Failory

The One-Hour Home

Nicolás Cerdeira — Thu, 10 Sep 2026 20:25:21 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Pallet, a startup building shelters for homeless people, shut down — learn why below
Why adding more AI agents makes your team slower
GPT 6 Astra and Fable 5.1 are out now and they are impressive — learn what people have been saying and building below

A big thanks to today’s sponsor…

Your best decisions start with listening ^AD

Every decision shapes your company's trajectory. SurveyMonkey ensures you're making them based on real insight, not assumptions.

Gather customer feedback through natural conversation. Get instant AI-powered analysis. Benchmark your metrics against peers in your space.

No expensive research team needed. Just real data, real fast. Integrate with ChatGPT and Slack so feedback flows directly into your workflow.

The companies learning fastest from their customers grow fastest.

→ Learn how today

This Week In Startups

🔗 Resources

Why adding more AI agents makes your team slower

Nobody cares about your new product (sorry)

How to turn your AI into a world-class designer

Will ChatGPT recommend your brand? Run a free 2-minute AI readiness scan on your site. *

📰 News

GPT has solved one of the Millenium Problems

Meta released a new personal AI agent called Muse

Apple announced a new foldable phone: the Iphone Duo

Nvidia confirms it will buy Hugging Face for $12.9 billion

💸 Fundraising

Gimlet Labs, an AI infrastructure startup building a multi-silicon inference cloud, raises ⁠$300M Series B funding

Aitan, a defense tech startup building autonomous robotic warfare systems, raises ⁠$41M funding round

Resect AI, an AI accountability platform for detecting hallucinations and controlling model behavior, raises ⁠$25M funding

Diameter Pay, a fintech infrastructure platform for global US dollar payments and virtual accounts, raises ⁠$10M Series A funding

* sponsored

Fail(St)ory

The Missing Village

Pallet built flat-pack sleeping cabins that cities could assemble in under an hour and group into staffed shelter villages.

On September 4, 2026, the company ceased operations after nine years. Pallet's own figures show annual village openings falling from 35 in 2021 to five in 2026.

What Was Pallet:

Pallet made private shelters for people experiencing homelessness or displacement. Its cabins had beds, windows, locking doors and optional heating or air conditioning. Kitchens and bathrooms were initially communal, which kept the units light enough to move and deploy quickly.

For residents, a lockable room offered privacy and security that a tent or congregate shelter could not always provide. For cities, it created shelter capacity faster than permanent housing and gave service providers a stable place to reach people.

Cities, counties, tribes and nonprofits bought the cabins, usually with public or philanthropic funds. A separate provider often ran the village, handling intake, meals, sanitation, security, case management, maintenance and housing placement. The cabin was one line in a much larger operating budget.

Pallet primarily generated project-based revenue from sales of shelters and village components. That made growth depend on large public purchases, with appropriations, procurement and site approvals sitting between interest and revenue.

Assembly could take less than an hour. However, opening a site could take two years once community opposition, zoning, utilities and operator readiness entered the process.

The gap showed up after customers had already spent money:

Nashville bought 108 cabins for $1.2 million, then left most in storage for years while officials worked through code and operating requirements.
Rochester's 19 cabins also sat unused as insurance, nonprofit capacity and funding stalled.

Pallet responded by refusing to ship without an essential-services plan and adding advisory work on funding, sites, policy and operators. Those moves were meant to reduce the risk of unused cabins, while pulling the company further into problems outside manufacturing.

During the pandemic, Pallet reported producing 1,222 shelters and opening 35 villages in 2021. It raised a $15 million Series A the following year and used outside capital for materials, factory space and hiring.

The later S2 cabin added safety, energy and comfort features. A furnished, delivered and installed unit cost roughly $16,000 to $17,000 in 2026.

Pallet still had buyers near the end:

Seattle opened 50 units in June.
Dallas selected Pallet for a proposed 75-unit site in August
and existing villages remained open after the shutdown.

Those late projects show that demand remained through the end. Pallet never disclosed its order book, leaving its shrinking deployment count as the clearest measure of the wider decline.

The Numbers:

🏗️ Founded: 2016
🏠 Shelters shipped: 5,780
📍 Reach: 150+ sites in 139 communities
📉 Annual villages: 35 in 2021 to five in 2026
💰 Series A: $15M in 2022
🛑 Operations ceased: September 4, 2026

Reasons for Failure:

Public demand arrived in waves: Pallet expanded factory capacity and hiring near the pandemic-era peak. Village deployments then declined for several years. Seattle's mayor said shifting public funding left Pallet without enough purchasing demand, and California placed zero orders with Pallet after selecting it for a 1,200-home program. The likely consequence was a factory carrying fixed costs against a thinner project pipeline, although Pallet never disclosed its utilization or margins.
Every sale depended on a second operating system: Land, utilities, code approval, insurance, services and permanent-housing exits determined whether purchased cabins became a working village. Delays left units idle and stretched the time from political commitment to repeatable revenue. Pallet's move into advisory services suggests these customer-side dependencies had become part of its sales problem.
Product and mission choices raised the entry price: Early cabins cost $3,500 to $7,500, while later quotes reached roughly $16,000 to $20,000 for a more capable installed unit. Pallet also used paid, supported labor rather than volunteers. Those choices may have strengthened durability and social impact, but cash-constrained buyers could count doors more easily than lifecycle value.

Why It Matters:

Match growth to the buyer's clock: Before adding factory space, inventory or headcount, prove implementation-ready orders arrive fast enough to support the cash cycle.
Treat external dependencies as part of the product: When customers need permits, infrastructure or partners before they can use it, solve those steps profitably or focus on buyers that already have them covered.

Trend

The AI Title Fight

Anthropic released Claude Fable 5.1 on September 1. OpenAI released GPT-6 Astra two days later. Since then, users have been conducting the industry's most orderly possible war: running the same prompt twice and arguing about the screenshots.

This is the closest thing frontier AI has to a main event. The models have roughly one-million-token context windows, the same 128,000-token maximum output, and identical list prices of $10 per million input tokens and $50 per million output tokens.

So, who is the winner?

Why it Matters

The two releases are close enough that a leaderboard cannot make the buying decision for you. Artificial Analysis gives both models an intelligence score of 53, while Epoch scores Astra at 167 and Fable at 164, with heavily overlapping confidence intervals.
Their identical token prices also produce very different bills. In Artificial Analysis's maximum-effort tests, Astra cost $3.26 per task and Fable cost $7.63. Fable generated about 78,000 output tokens per task, versus 27,000 for Astra.

What People Are Saying

One builder had Astra play through Portal using screenshots for 23 hours and 43 minutes, making 3,336 tool calls. The bespoke setup says little about an ordinary coding task, but it gives Astra unusually concrete evidence of persistence.

Simon Willison tested Fable by asking it to draw a pelican in SVG (a task AI usually struggles with), then increasing the model's reasoning effort. The quality rose sharply, but so did the bill. His max-effort result consumed 65,927 output tokens, took almost 14 minutes, and cost $3.30. Animating it cost another $1.37. It was, in fairness, a very nice pelican.
Matt Shumer pushed both mdoels through Blender, Three.js, Unreal Engine, subagents, reference gathering, and visual critique. His launch-week builds included a 3D Manhattan, a browser neighborhood filled with zombies, and an
agent civilization.

Bart Slodyczka ran the clearest direct comparison: five prompts, five unattended app builds, and both models working from the same assignments.

Tom Krcha used it to model a house. It seems that Astra is specially good at working with 3D spaces.

The Trend

So, what is the conclusion after a week? Which model seems to be better at which task?

Astra looks like the better default when the model needs to act: operating software, using a terminal, navigating a browser, or building inside CAD and 3D tools. It leads the relevant independent benchmarks, finishes many public builds faster, and reached the same overall intelligence score with far fewer tokens.
Fable looks stronger when the job depends on judgment: planning, reviewing, following an existing codebase, reasoning across long context, or adding visual polish. It may take longer and spend more, but builders often prefer what it notices and how the final result feels.

The evidence is less than ten days old and still too thin to turn these into permanent job descriptions. Codex and Claude Code shape the experience, public demos follow launch marketing, and some developers report the reverse.

Either way, Astra and Fable are currently the two models to beat. Artificial Analysis has them tied at the top, so the winner remains unclear. Everyone else, at least for now, is on the undercard.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Outspent and Outdelivered

Nicolás Cerdeira — Thu, 03 Sep 2026 20:05:15 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Satvacart, a grocery delivery startup, shut down — learn why below
Read this before deleting your AGENTS.md
There are two new insane video models: Atlas and Solaris — learn all about them below

A big thanks to today’s sponsor…

Can you afford your next hire? ^AD

Founders learn the answer to this the hard way. Hiring changes your expenses, burn, and the month your cash runs out.

Parallel connects to your accounting and gives you a live forecast built from your real numbers. Add a hire, see the fully loaded cost, and the impact on your runway & cash. So you can hire (and raise) knowing exactly where you stand.

Founders backed by Y Combinator, Khosla Ventures, and Techstars use it.

Start your 15-day free trial (no credit card)

This Week In Startups

🔗 Resources

Read this before deleting your AGENTS.md

The channel that pays you to use it

The Creator Economy boom in B2B GTM

Will ChatGPT recommend your brand? Run a free 2-minute AI readiness scan on your site. *

📰 News

OpenAI launches Astra

NVIDIA agrees to acquire Hugging Face for $12.93B

Anthropic launches Claude Fable 5.1 and Mythos 5.1

Meta is paying to peek at how you use their latest AI model

💸 Fundraising

Wonderful, an enterprise platform for deploying and coordinating AI agents and workflows, raises $550M Series C

HiddenLayer, a security platform protecting AI models, agents, and workflows, raises $100M Series B

Superluminal Medicines, a biotechnology startup using machine learning and structural biology to discover small-molecule drugs, raises $60M Series B

Light, a platform that lets businesses offer branded retail-electricity plans, raises $46M Series A

* sponsored

Fail(St)ory

The Second Grocery War

Satvacart began delivering groceries around Gurugram in 2014, during India’s first wave of online grocery startups. It survived India’s first online-grocery shakeout and learned how to make one local cluster economical.

On August 28, 2026, the company stopped operating and disbanded its team after twelve years.

By then, Blinkit, Zepto and Instamart had turned grocery delivery into a race for dense citywide networks. Satvacart’s discipline had produced a carefully tuned local operation. The new contest required far greater coverage and access to much larger pools of capital.

What Was Satvacart:

India’s first online-grocery startups were trying to move the weekly supermarket run onto a phone. Behind each app sat a physical operation that sourced thousands of products, packed every order and sent riders across a city.

Expansion was expensive. A new area needed a warehouse, staff, stock and marketing before enough nearby orders existed. One playbook used investor money to open quickly and absorb early losses. Satvacart, on the other hand, chose to make one small area economical first.

Satvacart launched in early 2015 with 4,000 products in Gurgaon. It planned to reach six cities by year-end. The rollout did not happen, and the company remained local.

Inspired by the online grocer FreshDirect, Satvacart built one warehouse serving homes within roughly five kilometers. It called this a micro-cluster: one fulfillment point, nearby customers and short delivery routes. Shoppers could choose half-hour delivery slots.

The economics improved when nearby households ordered regularly. The warehouse and riders stayed busy, while several deliveries could share a route. Prepaid subscriptions helped create that density. By 2016, Satvacart claimed they had cut delivery costs from ₹300 to ₹25 and brought orders to unit-level break-even.

This discipline appears to have helped Satvacart survive an e-grocery meltdown that removed many better-funded startups. By 2019, it reported a small positive EBITDA with support from a separate B2B marketing activity. The grocery business still occupied one micro-cluster.

Quick commerce then raised the cost of competing. Blinkit, Zepto and Swiggy’s Instamart promised groceries in minutes using many small warehouses near customers. Each “dark store” had to be rented, stocked and staffed before demand filled it.

Satvacart tried to adapt. In 2021, it claimed workers could pick 23 items in two minutes and deliver them within ten minutes inside a two-kilometer radius. It announced thirteen-city expansion and talks for a $50 million round.

The round never appeared, and Satvacart remained Gurgaon-focused. By Q3 FY2026, Instamart operated 1,136 dark stores across 131 cities and still reported a ₹9.08 billion adjusted EBITDA loss.

Satvacart had survived the first shakeout with one efficient cluster. The next phase required a network it could no longer finance.

The Numbers:

🏁 Founded: 2014; commercially launched in early 2015
🛒 Early assortment: About 4,000 products
💰 Lifetime funding: Roughly $2.2M-$2.3M
📍 Core radius: About five kilometers
📉 FY2023: ₹1.4 crore revenue and ₹93 lakh net loss

Reasons for Failure:

One cluster never became a rollout system: Satvacart learned how to concentrate orders around a single warehouse. Each new area still required rent, staff, inventory and customer acquisition before reaching useful density. Repeated city launches remained plans, leaving the company too localized for the scale investors and buyers wanted.
Profitability labels obscured the cash available for growth: Unit break-even, contribution margin, two cash-positive months and small EBITDA described different periods and cost layers. Filed accounts later showed a company-level loss. None of those earlier milestones proved that one mature cluster could finance the next.
The large round never arrived: Satvacart appears to have raised roughly $2.2 million to $2.3 million across its lifetime, mostly through angel and pre-Series A rounds. In his shutdown account, the founder said smaller tranches were insufficient, two significant investment discussions failed, and acquisition talks stalled because the company lacked scale.

Why It Matters:

The strategy that helps a startup survive one wave of competition may leave it exposed to the next. Revisit old assumptions when the market changes.
A profitability claim needs a clear scope. A profitable order or location can still sit inside a company that is losing money.
In network businesses, scale changes the customer experience. A strong local operation may become less attractive once rivals offer broader coverage and faster service.

Trend

AI Video Specializes

Runway and World Labs released two video-capable models within two days. Solaris turns clicks and drags into the next frame of an interface. Atlas takes camera geometry and spatial context, then generates controlled views or explicit 3D output.

They point to a wider split in AI video. The question is shifting from which model produces the best-looking clip to which model gives a product the right controls. Interactive software, navigable worlds and spatial reconstruction each ask the model to hold a different kind of state.

Why it Matters

Model choice now starts with the workload: live interaction, camera direction, reconstruction, 3D output or a fixed clip.
Value moves toward the context a model can maintain and the controls it can follow. That opens infrastructure budgets in software, games, robotics and simulation, along with recurring inference costs.
Aesthetics becomes one metric among several. Teams also need to test action fidelity, state drift, geometry, latency, physics and safety.

Solaris

Runway's Solaris generates an interface frame by frame. An LLM decides how the experience should respond, while a world model based on Gen-4.5 renders the next 720p frame after a click, drag or other action.

A conventional interface has coded screens, components and transitions. Solaris turns that fixed set of states into a running exchange: after every action, the LLM decides what should happen and the visual model draws it.

Runway's demos apply the system to virtual stores, rooms and courseware, where the experience can keep changing as the user moves through it.

That freedom is only useful if the generated interface still feels responsive. Runway tested this with 250 participants, who made nearly 7,500 comparisons across 30 examples. Each comparison put Solaris beside an interface coded by Claude Opus 5.

When judging instruction following, participants preferred Solaris 62% of the time, the coded interface 25%, and rated them equal 13%. For natural behavior, Solaris received 72%, compared with 21% for the coded version and 7% equal. These are Runway's own results, and the test measured how short interactions looked and felt. It did not test whether Solaris could preserve state, finish a task securely or remain coherent through a long session.

Runway says stable text, grounding, accessibility and integration also remain unresolved. Solaris could handle the exploratory, visual layer of an interface, with tested components managing payments, permissions and records behind it. A convincing button is still a picture until something reliable handles what happens when you click it.

Atlas

World Labs' Atlas accepts images or depth maps at explicit camera poses. It can generate up to one minute of 1440p video along a specified camera path, reconstruct a scene from a few images and return explicit 3D representations for editing or simulation.

A user tells Atlas where the camera is and where it should move, while the model tries to keep the scene coherent across the path. World Labs says two or three reference images can be enough for reconstruction, although Atlas may imagine areas the inputs never showed.

Those capabilities matter when a scene needs to live beyond one rendered clip:

Game and simulation teams could turn a few reference images into an editable 3D starting point, direct new camera moves through it and generate variations without rebuilding the environment by hand.
Robotics teams could use those reconstructed scenes in real-to-sim workflows for training and evaluation.

Together, those applications could reduce the hand-built work between capturing a place and using it inside another system.

The Trend

The AI video category keeps splitting into smaller ones. Some models make clips. Solaris generates interfaces you can interact with. Atlas generates views you can direct with a camera path and turn into 3D. Project Genie and Odyssey generate worlds that keep responding to actions.

Choosing a model now starts with what you want to control: the shot, the interface, the camera or the world. Benchmarks will have to measure interaction, consistency and control alongside visual quality.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Electricity Through Thin Air

Nicolás Cerdeira — Thu, 27 Aug 2026 19:43:21 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Wi-Charge, a startup using light beams to charge devices through the air, has shut down — learn more below.
ARR Doesn't Mean What It Used To
OpenAI just unveiled its first AI chip. It beat Nvidia. — learn more below

This Week In Startups

🔗 Resources

ARR Doesn't Mean What It Used To

How to Build a High-Growth Operating System

Execution Is The Moat

You are not a model. Don’t price per token.

📰 News

Chinese lab Z.ai says it’s behind the Ox Alpha model that wowed Silicon Valley

Apple debuts its ‘most powerful chip ever’

There’s a new Twitter in town

Nvidia closes in on Hugging Face acquisition

💸 Fundraising

Higgsfield, an AI image and video creation platform, raises a $400M Series B, Higgsfield raises $400M Series B

Groq, an AI inference cloud provider, raises a $350M Series A, Groq closes $350M Series A

Wispr, an AI voice dictation and interaction platform, raises a $280M Series B, Wispr’s Series B

Rillet, an AI-native accounting and ERP platform, raises a $100M Series C, Rillet raises $100M Series C

Fail(St)ory

Charging Through the Air

In 2015, a Samsung Galaxy S4 spun on a platform at Mobile World Congress while an invisible beam charged it from across the room.

That was Wi-Charge’s pitch in one neat demo: electricity could travel through the air and find a device without a cable trailing across the floor. In July 2026, the company filed for insolvency.

A co-founder later described the trap. Device makers wanted rooms full of transmitters before adding receivers to their products. Buyers wanted compatible devices before installing transmitters. Each side waited for the other, which is a difficult way to sell electricity.

What Was Wi-Charge:

Picture a hotel corridor full of smart locks. Someone has to notice the dying batteries, visit each door and replace them. Wi-Charge wanted one transmitter to keep several locks charged from across the hall.

Its AirCord system sent focused infrared light from a powered transmitter to a receiver built into the lock, display or charger. The receiver turned that light back into electricity. People could not see the beam, although walking through it stopped the power until the path cleared.

Devices carried a small rechargeable battery to survive those interruptions. The power also needed a direct view of the receiver. It could cross a room, but it could not turn a corner, pass through a pocket or negotiate with a badly placed cupboard.

The first working transmitter cost about $500,000 and required six PhDs to package, according to the co-founder. Years of engineering shrank later versions to roughly the size of a soft-drink can.

For stores, the appeal was movable digital signs without new wiring. Hotels and offices could avoid rounds of battery replacement for locks. Wi-Charge also demonstrated powered speakers, faucets, toothbrushes and a charging pad whose own power arrived wirelessly.

WIRED’s five-week bathroom test captured the experience nicely. A ceiling transmitter kept two electric toothbrushes charged from six feet away. Installing this wireless future required an electrician, an outlet in the loft and a hole in the ceiling.

It also used much more electricity than an ordinary toothbrush charger. Wi-Charge said the telemetry-heavy test unit was inefficient and a finished version would cut consumption sharply.

The company offered transmitters, receivers and integration kits to device makers, then built some finished products itself. These included store displays, an Alfred smart lock option and a $149 kit for Schlage Encode locks.

A co-founder claimed a footprint across 39 U.S. states, and the Schlage-compatible kit’s first batch reportedly sold out. Wi-Charge never disclosed that batch’s size, its revenue, margins or number of paid deployments.

The Numbers:

🧪 First transmitter: About $500K and six PhDs
📡 Published range: Up to 33 feet
💵 Consumer lock kit: $149
💰 Series C: $20M announced in May 2025
👥 Team: About 45 people in a 2024 estimate
🛑 Insolvency filing: July 14, 2026

Reasons for Failure:

Every room needed its own little network: Device makers told the co-founder they wanted a large installed base of transmitters before adding receivers. Property owners saw little value in transmitters while few compatible products existed. Wi-Charge had to persuade both sides to move together.
The annoyance rarely earned a new system: Batteries and cables already powered the same locks, signs and toothbrushes. Replacing them was tedious, but the co-founder said customers rarely treated it as urgent. An invisible laser was much cooler than changing batteries; cool alone did not make the installation essential.
Installation stayed physical: Each site needed powered transmitters, clear sightlines and receivers inside every target device. New product categories brought more integration work, from locks to bathroom fixtures and retail screens. The public record suggests growth kept depending on tailored deployments instead of one repeatable setup.
Commercialization consumed a decade: Wi-Charge expected smart-home products in 2016 and a phone kit in 2017. It was still describing deployments as early in 2023, and its first direct consumer product arrived months before insolvency. The company announced a $20 million Series C in 2025, although the amount funded and its finances at closure remain unknown.

Why It Matters:

Deep-tech demos compress years of engineering into a few delightful seconds. The buyer pays for the whole package around them: installation, compatible devices, maintenance and a changed workflow. Test that package before scaling the hardware.
Products that need infrastructure and compatible devices should measure useful density inside one site. Ten scattered pilots can look like traction while leaving the eleventh customer with the same empty-room problem.

Trend

OpenAI Goes Silicon

OpenAI says its need for compute is so large that it cannot imagine when it will have enough Jalapeño chips to sell one. For now, the chip is still being tested in its labs.

Jalapeño is OpenAI's first custom inference chip, built with Broadcom to serve ChatGPT, Codex, the API, and future agents. The strategy behind it is to use OpenAI's models and production workloads to improve the hardware, then use that hardware to make the products faster and cheaper to run.

At Hot Chips this week, the company published the chip's first measured performance results. Jalapeño ran three large open models and beat Nvidia's GB200 and GB300 on speed and energy efficiency.

Why it Matters

Every ChatGPT response carries a compute bill. Custom silicon gives OpenAI another way to reduce that bill and another source of capacity.
Latency is part of the product. A voice model that answers half a beat sooner feels better; a coding agent that can run efficiently for longer can take on larger jobs.
OpenAI can now design around its own bottlenecks and product roadmap. The company says it will continue buying Nvidia hardware at scale. Jalapeño gives it more control over how quickly its infrastructure follows new models.

Jalapeño

An LLM request is several hardware problems disguised as one answer. Reading the prompt requires heavy computation. Producing each new token requires repeatedly pulling model data through memory. Spreading the work across chips adds communication delays.

Jalapeño handles those phases on the same system. It keeps the model's working memory local, changes which compute, memory, and networking resources are active, and powers down unused units. OpenAI designed the system around traffic from its own products, which shows where time and electricity are being spent.

Each Jalapeño package can hold 216 GiB of model data and move 15.4 TB of it every second. A full pod connects 2,048 chips, so you can do the math (it’s a lot). Each chip is rated at 700 watts, though OpenAI says the benchmarked workloads never pushed it beyond 550.

OpenAI then put Jalapeño against Nvidia's GB200 and GB300 using three large open models. Its chip completed 1.5 to 1.9 times more work per watt and returned full answers 1.7 to 3.6 times faster. In the tests designed to measure highly interactive use, Jalapeño was 2.1 to 4.1 times faster. On OpenAI's scorecard, Nvidia lost every matchup.

SemiAnalysis ran its public InferenceX benchmark with OpenAI engineers in OpenAI's lab and concluded that Jalapeño beat every Nvidia, AMD, and Google chip it had tested across several major open models.

The Trend

Jalapeño is the chip nobody can buy. OpenAI says it has no plans to sell it, and its internal demand is too large to leave much spare capacity anyway. Any advantage will show up inside ChatGPT, Codex, and the API.

That changes the economics of the AI race. If Jalapeño delivers more work per watt at scale, OpenAI can use the gain to lower prices, increase capacity, run agents for longer, or keep more margin on the same products. Competitors without comparable private infrastructure would still be paying a supplier for that layer.

A rival such as Anthropic could build an equally capable model and face worse serving economics. Matching OpenAI would then require better supplier terms, a more efficient model, or private silicon of its own. Pricing and scale would depend on the model and the infrastructure hidden underneath it.

OpenAI has also joined the chip industry from the buyer's side. Jalapeño is intended for its own data centers. As deployment grows, each chip can shift inference demand away from Nvidia and other suppliers. Those companies now face a customer capable of becoming its own supplier for part of an enormous inference load.

The scale remains unknown. OpenAI plans only limited Jalapeño capacity in 2026, will keep buying Nvidia hardware, and has published no cost per token, manufacturing yield, or production volume. Its 2027 rollout will supply the missing evidence.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Buying Better Weather

Nicolás Cerdeira — Thu, 20 Aug 2026 22:16:12 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

ClimateAi, a startup using IA to predict weather, has shut down — learn more below.
This post will save you tokens
A new category of startups is building layers that let SaaS users create their own features — learn more below

This Week In Startups

🔗 Resources

How to make people care about your startup

This post will save you tokens

How the GTM playbook has changed in 2026

Dear VCs, Just Give Up

📰 News

OpenAI launches a safer ChatGPT for teens

Perplexity’s free AI offer left it with millions more users in India

SpaceX officially closes its Cursor acquisition

OpenAI introduces ‘Ultrafast,’ a new mode that makes GPT-5.6 Sol work at 14x the speed

💸 Fundraising

Neros Technologies, a defense startup building autonomous and interceptor drones, raises ⁠$250M Series C funding

Skan AI, an enterprise AI platform for mapping and automating business workflows, raises ⁠$63M funding

Axle, an AI-native insurance data clearinghouse, raises ⁠$17.5M Series A funding

Alloy Robotics, an AI platform for debugging and analyzing robot fleets, raises ⁠$8M to help engineers debug robot fleets

Fail(St)ory

AI Meteorologist

ClimateAi sold weather and climate forecasts to companies whose crops, supply chains, and commodity costs depended on what happened months or years ahead.

It shut down on two weeks ago, after raising about $38 million. The challenge was turning better forecasts into recurring software spend and real operating decisions.

What Was ClimateAI:

ClimateAi sold weather forecasts built for business decisions. Its main product, ClimateLens, combined weather data, satellite observations, climate models, and machine learning, with most of its work focused on food and agriculture.

A normal weather forecast tells you what might happen. ClimateAi pushed further into what that could mean for crop yields, supply, inventory, and prices over periods ranging from a few weeks to several years.

That could change when a company harvested, how much inventory it carried, where it sourced from, or when it bought a commodity.

Some of those decisions were worth a lot. ClimateAi said a wet-harvest warning helped one customer avoid losses worth hundreds of thousands to millions of dollars. Another customer used its crop outlook as part of a trading strategy that the company estimated produced $2 million to $6 million.

The product fit best when weather already sat close to an existing business decision. Procurement teams, for example, already managed supply, contracts, and inventory, so a forecast could feed directly into work they were doing anyway.

By late 2024, ClimateAi said it had worked with more than 56 partners across 40-plus crops and more than 60 countries.

Customers used it in different ways. Some paid for recurring access, while others hired ClimateAi for a specific analysis and came back when another planning problem came up. The company wanted the economics of software, with the same forecasting system serving many customers at low added cost.

The forecast still had to land somewhere useful inside the customer’s workflow. ClimateLens 2.0, released in early 2026, pulled in planting records, contracts, and delivery schedules so ClimateAi could send alerts tied to decisions the customer already had to make.

That was the basic bet behind the company: give businesses enough warning to change what they buy, plant, move, or protect before weather changes the outcome. ClimateAi eventually said it served roughly a quarter of the world’s top 200 food and beverage companies.

The Numbers:

💰 Raised: $38M
🌍 Countries covered: 60+ by late 2024
🌾 Crops: 40+ across 56+ partners
📈 ARR growth: 5x before April 2023
👥 Staff: ~52 in August 2025
🛑 Wound down: August 7, 2026

Reasons for Failure:

No clear budget owner: ClimateAi said customers often lacked a standing budget for climate adaptation. Procurement could pay when forecasts affected sourcing or inventory, but broader resilience work crossed several departments, which made sales and renewals harder to standardize.
The software required operating changes: A forecast sitting in a dashboard had limited value. Customers had to change contracts, inventory, planting dates, hedges, or capital plans early enough for the forecast to matter, often across teams with existing approval cycles.
ROI arrived unevenly: ClimateAi could point to cases where one forecast saved or generated millions. Those wins depended on specific weather events, crops, and decisions, so a customer might see a large payoff one year and little visible financial impact the next.
Repeatable scale was still unresolved: In late 2024, the company expected profitability in 2025 and discussed reaching roughly 200 partners before another funding round. It was also expanding across yield tools, APIs, demand forecasting, government work, and workflow products, suggesting it was still working out which buyer and use case could support the company at scale.

Why It Matters:

Budget ownership shapes adoption: products that affect several departments need a clear person who controls the money and the decision.
Risk products need value between big events: one avoided loss can justify years of fees, but customers still review the bill every year.
Forecasts become more useful when they sit inside the decision: linking weather data to contracts, schedules, and inventory reduces the work required to act on it.

Trend

Extensible SaaS

Lately, I’ve been noticing a new startup category with one pitch: add a generative layer to an existing SaaS product so every customer can build the features they are missing. Vendo calls them “user-generated features.”

Say your sales team uses a CRM but needs a page that groups renewals by month, shows the account owner and includes a button for sending a Slack reminder. You describe that page inside the CRM. The layer builds it using the records already there, then saves it to your account so the team can keep using it.

From the user’s point of view, the SaaS has learned a new feature for their workflow. The new page looks and behaves like part of the product, follows the same access rules and can be shared with colleagues. One product can now grow into a slightly different version for each customer.

Why it Matters

A feature request can become something the customer uses. Teams can build the niche dashboards, forms and workflows they need without waiting for them to reach a product roadmap.
The work stays connected to the software where the data already lives. Users can avoid exporting records to spreadsheets, buying another tool or asking a developer to rebuild the same context elsewhere.
Customer-made features create better evidence of demand. A vendor can see which additions people keep using, promote the best ones to the main product and let other customers reuse them.

The Signals

Vendo offers the clearest version of this experience. A user can ask for a dashboard, workflow, new action or persistent micro-app inside an existing SaaS product. The result uses the host product’s data and visual components, and it can connect to tools such as Slack or Salesforce when the vendor allows it. Sensitive actions still require approval.

OnClaw uses the line “Every user gets their own version of your app.” A user opens a command bar, describes a missing feature and receives a new component inside an area the vendor has made customizable. The feature stays with the product. Developers can later inspect and edit what was generated.

Gigacatalyst has the strongest disclosed production example through UpKeep Studio. Maintenance teams can ask its Nova assistant to build inspection forms, shift-handoff tools, compliance workflows and specialized dashboards from their existing UpKeep records. Users can review the result, revise it and publish it to colleagues. Gigacatalyst claims more than 800 generated features in six weeks and 70% day-30 retention, though those figures are unaudited.

Several other startups are selling variations of the same outcome. Legato lets business users create apps, workflows, automations and agents inside B2B platforms. Vezel offers generated dashboards, reports and internal tools, while Fork gives each customer a separate branch of the original application. Fork also lets vendors bring popular customer-made features back into the main product.

The Trend

People have customized software through spreadsheets, scripts and no-code tools for decades. These startups bring that behavior inside the SaaS product itself. The language model turns a plain-English request into a feature, and the layer gives that feature access to the right data, a place in the interface and a permanent home in the customer’s account.

I think this is becoming a real software category. The startups are selling the machinery that turns one fixed product into many customer-shaped versions, while the original vendor continues to control the data, permissions and core services.

The evidence is still early. UpKeep is the only named production deployment in this group. The category will matter if users keep returning to the features they generate and trust them for everyday work.

That is where this new category could matter. SaaS companies would keep building the core product, and customers would use these layers to create the smaller dashboards, workflows and tools that make it fit the way they actually work.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

80 Acres in One

Nicolás Cerdeira — Thu, 13 Aug 2026 19:42:30 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

80 Acres Farms, a startup building vertical farms, has shut down — learn more below.
The emptiest distribution channel with 800M users
Founders are using AI to quickly make tiny SaaS and sell them for very cheap — learn more below

This Week In Startups

🔗 Resources

The emptiest distribution channel with 800M users

10 ways Clay's GTM engineers use AI to accelerate sales

Meet your biggest competitor (👋 Claude)

Tech’s Search for Legitimacy

📰 News

OpenAI says it slowed Astra model development over security concerns

Cloudflare launches Kitesurf, a browser built for AI agents

Claude will now add invisible watermarks to text

Mark Zuckerberg wrote a controversial essay about the future of AI

💸 Fundraising

HappyRobot, an AI platform for automating enterprise operations, raises ⁠$150M Series C funding

Ambrook, an accounting and payments platform for independent businesses, raises ⁠$30M Series B funding

Buzz Solutions, an AI platform for electric grid inspections, raises ⁠$20M Series A funding

Valar Atomics, a nuclear startup developing mass-produced microreactors, raises ⁠$1B Series B funding

Fail(St)ory

80 Acres Farms

80 Acres Farms built indoor farms that could grow lettuce, herbs and tomatoes all year, close to the supermarkets selling them.

By August 2026, it had raised more than $350 million, supplied thousands of stores and claimed several farms had profitable unit economics. Then a planned sale fell apart, the company ran out of funding, and operations stopped almost immediately.

What Was 80 Acres Farms:

80 Acres grew produce inside warehouses using stacked growing beds, LED lights, sensors, automated handling and tightly controlled water, temperature and humidity. Crops could grow every week of the year without waiting for the right weather or season.

That gave supermarkets a fairly simple benefit: predictable local supply. A retailer in Ohio could buy greens grown nearby instead of bringing them across the country, with fewer weather shocks and less time between harvest and shelf.

80 Acres sold herbs, microgreens, tomatoes and salad blends, then added salad kits, dressings and prepared meals. Kroger became a major customer, and products eventually reached chains including Whole Foods, Meijer and The Fresh Market.

The farms also came with a sustainability pitch. 80 Acres said they used 95% less water per pound than conventional farming, ran on renewable electricity and grew without pesticides.

Behind the produce sat a large technical operation. Its GroLoop system tied together lighting, climate controls, crop data, forecasting, inventory and automated equipment. Running one of these farms meant managing biology, software, machinery and a food production operation under the same roof.

The farms were expensive to build and run. Indoor growing meant paying for lighting, climate control, automation, maintenance and labor every day, so each facility needed high, steady output to make the economics work.

80 Acres said it had reached that point at several farms. By 2024, it reported selling about 2,000 metric tons of produce and claimed profitable unit economics at the farm level.

The company then grew far beyond those individual farms. It added new facilities, acquired other businesses and merged with Soli Organic, expanding to more than 17,000 retail locations.

That growth brought a much larger cost base around the farms, including corporate staff, R&D, technology and a national operating network.

The Numbers:

🏗️ Founded: 2015
💰 Funding: $350M+ raised
🥬 Produce sold: ~2,000 metric tons in 2024
🏪 Retail footprint: 17,000+ locations after the Soli merger
👥 Jobs affected: 800+

Reasons for Failure:

Farm profits could not carry the company: 80 Acres said several farms had profitable unit economics, but the company still had to fund R&D, software, sales, headquarters and other central costs. By March 2026, management acknowledged that the overall business remained unprofitable.
Expansion kept adding fixed costs: New farms meant more staff, equipment, maintenance and working capital, whether each site was running near capacity or not. Expanding into Georgia, Texas and Colorado increased the amount of cash the company needed to keep the network operating.
The business became too broad: 80 Acres added salad kits, dressings, prepared meals, plant genetics and farm technology, then merged with Soli Organic and added roughly 1,000 employees. That left profitable farms supporting a much larger and more complex operation.
The company ran out of financial room: By summer 2026, 80 Acres expected a sale of the company to provide the cash needed to keep operating. The buyer withdrew on August 2, and WARN filings say there was no other funding available, forcing the shutdown the next day.

Why It Matters:

In capital-heavy businesses, “profitable units” can be misleading if each new unit still depends on a growing layer of central R&D, software, sales and management that never gets cheaper.
Buying distressed assets can look capital-efficient, but cheap capacity is only useful when you already have enough local demand to keep it full.

Trend

Micro-SaaS Flipping

AI has made it possible to build a tiny software product in days. A growing number of founders are taking the next step: get a few users, some revenue or a burst of traffic, then sell the whole thing for a few thousand dollars and start again.

Why it Matters

A $5,000 exit means more when the product took a week to build. Founders can run many more shots, recover money from the ones that show some demand, and move on before turning every experiment into a company.
Buyers increasingly pay for the head start around the code. Users, search traffic, a working payment setup, customer feedback and a clear niche can save months of uncertain work. AI keeps reducing the time required to rebuild software, which makes those other assets more important.
This creates room for a new type of founder. Someone can specialize in finding small opportunities and getting products to their first signs of demand, then hand them to operators who are better at sales, support and long-term growth.

The Signals

The example that pulled me into this was Livestamp.

Its founder says he had no app-building experience. He saw a photo effect going viral, used Claude to figure out how it worked, ChatGPT for the visual assets and Codex to build it. Three days later he had a working app.

Then his demo went viral, the app passed 10,000 downloads and roughly 1,000 daily users, and a Vietnamese CEO offered to buy it. Six weeks after starting, he sold for $8,000 with zero MRR.

Once I started looking for similar cases, they were surprisingly easy to find.

In June, an AI blog generator doing $100 MRR sold for $2,000 in 12 days on TrustMRR. A few weeks later, an AI virtual try-on app doing $300 MRR sold for $6,000. Around the same time, somebody sold a simple Stripe tool for recovering failed payments for $1,300 despite having zero revenue.

These are tiny exits from tiny products. That is exactly the behavior I’m interested in. Founders are getting products to a very early proof point and finding someone willing to take over from there.

This is Happening a Lot

The marketplace data is moving in the same direction. Flippa says SaaS transactions grew 73.5% in 2025, while 37% of its buyers made multiple acquisitions.

Buyer demand is still growing. In H1 2026, Flippa had 123,022 active buyers, up 18% from a year earlier.

And a new generation of marketplaces is forming around much smaller deals. TrustMRR only launched in October 2025 and now says it facilitates more than 20 acquisitions a month. It already gets around 200,000 monthly visitors, with many products on the marketplace making only a few hundred dollars in MRR.

That is the part I find interesting. There are now enough tiny software products for sale, and enough people willing to buy them, that a marketplace dedicated to micro-SaaS can clear several deals every week.

The Trend

I’ve started calling this micro-SaaS flipping.

The model is pretty simple: build, or occasionally buy, a tiny software product, get enough evidence that people want it, sell it to another operator, then move on.

People have been buying side projects and tiny SaaS businesses for years. SideProjectors was doing this in 2013. MicroAcquire made small startup acquisitions much easier starting in 2020. Founders were even running “build and sell a SaaS in 30 days” challenges before vibe coding existed.

AI is accelerating the cycle. A founder can now get from idea to working software in days, run more experiments and reach the point where there is something transferable with much less money upfront. On the buyer side, AI also makes an unfamiliar codebase cheaper to understand and maintain.

I think that combination is creating a proper second-hand market for software.

The useful mental model is that AI is turning software into inventory. A founder can manufacture a small product, prove some demand and sell the asset to someone who wants to handle the next stage.

Most of these exits will stay small. That is part of what makes the model possible in the first place.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Tap. Declined.

Nicolás Cerdeira — Fri, 07 Aug 2026 00:38:24 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Kulipa, a startup making infrastructure for crypto cards, has shut down — learn more below.
What nobody tells you about writing agent skills
AI has solved a lot of math conjectures in the last 2 weeks — learn why this matters below

This Week In Startups

🔗 Resources

Code Review (As We Know It) Must Die

Adding ambient AI: How to put AI to work where nobody is looking for it

The Endgame Of Vertical Integration

What nobody tells you about writing agent skills

📰 News

ChatGPT brings unlimited text chats to free users

Meta launches Muse Code, an AI agent for large code bases

Apple finally fixed Siri

OpenAI’s new AI smart speaker will reportedly sell for between $300 and $400

💸 Fundraising

Array Labs, a satellite radar startup, raises ⁠$21M strategic funding

ProphetX, a sports prediction market, raises ⁠$35M funding round

Mirae, an AI platform for continuous care of autoimmune diseases, raises ⁠$5.4M funding

Biota, an environmental diagnostics startup, raises ⁠$3M seed round

Fail(St)ory

Crypto Mastercard

Most crypto payments still hit the same wall: you can’t buy groceries, pay for dinner, or book a flight with a wallet full of stablecoins.

Kulipa was trying to solve that by building the infrastructure to let crypto wallets launch their own payment cards.

It worked fine until the cards stopped working.

What Was Kulipa:

Imagine you’re a crypto wallet. Your users already keep USDC in your app. They don’t want to cash out every time they buy coffee, and they don’t care about blockchain rails when they’re standing at a checkout. They just want to tap a card and move on.

That’s exactly what Kulipa made possible.

Instead of asking merchants to accept crypto, Kulipa went the other way. It connected stablecoins to the payment networks people already use. The customer paid with a normal Mastercard, the store received a normal card payment, and Kulipa handled everything happening in the middle.

For wallet companies, that removed an enormous amount of work. Launching a card program usually means finding licensed banking partners, passing compliance checks, integrating fraud systems, connecting to card processors, and figuring out how to move money between crypto and traditional finance. Kulipa wrapped all of that into a single API.

Its biggest selling point was that users didn’t have to preload a card account. Their USDC stayed in their own wallet until they actually bought something. At the moment of payment, Kulipa moved just enough funds to settle the transaction. The wallet stayed self-custodial, but spending felt like using any other debit card.

That was a pretty compelling pitch for crypto wallets. They could offer a feature people immediately understood without becoming payments companies themselves.

By April 2026, Kulipa said it had signed around 20 customers, issued more than 120,000 cards, and was seeing 70% month-over-month transaction growth. Partners ranged from crypto wallets to fintechs serving people who preferred saving in stablecoins instead of local currencies.

Then one of the licensed issuers behind the cards ran into trouble, and the whole setup unraveled fast. Within weeks, users lost access and partners were scrambling for a replacement.

The Numbers:

💰 Funding: ~$9.2M raised
💳 Cards issued: 120,000+
🤝 Customers: ~20 wallets and fintechs
📈 Reported growth: 70% month-over-month transaction volume
💵 Peak monthly volume: ~$9M (January 2026)
📉 Card services stopped: July 29, 2026

Reasons for Failure:

One issuer became a single point of failure: Kulipa depended on regulated issuers to connect its software to Mastercard. After one key issuer reportedly ran into regulatory trouble, Kulipa had to move programs to another partner with more limited coverage. Within weeks, cards started disappearing in some markets before the entire service stopped.
The product relied on too many moving parts: Kulipa looked like a software company, but every payment depended on banks, card networks, compliance providers, processors, and regulators working together. Each layer added another dependency outside the company’s control. That makes the business much harder to operate than a typical API startup.
Customers outsourced a critical feature: Wallets loved Kulipa because it let them launch cards without becoming payments companies. The tradeoff was that one infrastructure provider now sat between them and their users. When Kulipa went down, products from around 20 different companies went down with it.
Growth came before resilience: Kulipa moved quickly, reaching more than 120,000 issued cards and around 20 customers in a short time. That’s impressive traction, but payment infrastructure earns trust through years of reliability, not months of growth. Once partners started looking for replacement providers, that trust was difficult to recover.

Why It Matters:

Map every regulated dependency before scaling; one weak partner can shut down the whole product.
When customers build on your infrastructure, reliability matters more than feature velocity.

Trend

AI Math

AI has had a ridiculous two weeks in mathematics.

Anthropic helped disprove an 87-year-old conjecture. OpenAI answered by publishing ten new results on open problems in mathematics and theoretical computer science. These weren’t benchmark questions or Olympiad exercises. They were problems that mathematicians genuinely didn’t know how to solve.

This isn’t the first time AI has contributed to original math. The difference is the tempo. One breakthrough felt exceptional. A string of them, from multiple labs, starts to look like something else.

Why it Matters

Frontier AI is getting measured on original research. Benchmark scores don’t move the conversation like they used to. New discoveries do.
The cost of exploring ideas is falling fast. AI can search through thousands of possible directions. Humans spend more time deciding which ones are actually right.
Mathematics is probably just the first proving ground. It’s a field where new results are public, verifiable, and hard to fake.

What happened?

The biggest story came from Anthropic.

On July 20, mathematician Levent Alpöge casually posted on X that, while watching the World Cup final, Claude had helped find a counterexample to the Jacobian conjecture, an 87-year-old open problem. The entire result fit inside a tweet. Within hours, mathematicians started checking it. It held up.

A few days later, OpenAI published ten new results across mathematics and theoretical computer science, generated by an internal model called Astra. Instead of presenting one headline result, it released a collection spanning graph theory, geometry, coding theory, complexity, and more.

Neither company was working in isolation. AI had already been making steady progress on open math problems throughout 2026. What’s different is that the announcements are arriving much closer together, and they’re increasingly being used as public demonstrations of what frontier models can do.

It’s hard not to see some competition here. Once one lab shows its model can produce original mathematics, everyone else suddenly has an incentive to show they can too.

The Trend

Math is becoming the new benchmark.

Coding benchmarks are crowded. Leaderboards move by fractions of a point. Original research is much harder to dismiss. Either a result is new or it isn’t. Either other mathematicians can verify it or they can’t.

That’s why these announcements matter beyond mathematics.

If that capability keeps improving, mathematics won’t stay the main story for long. The same search process can be applied anywhere progress comes from exploring a huge space of possible ideas: chemistry, materials science, biology, economics, or algorithm design.

We’re still early. Human experts are doing most of the verification, and these results need time to be absorbed by the research community. But the last few weeks made one thing much clearer: frontier AI companies have started competing on scientific discovery, and mathematics is where they’re choosing to keep score.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

The Decentralized Google

Nicolás Cerdeira — Thu, 30 Jul 2026 23:50:48 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Presearch, a startup building a decentralized search engine, has shut down — learn more below.
How much can you delegate to agents?
Kimi K3 is out and everyone is debating about open-weight models — learn more below

This Week In Startups

🔗 Resources

Steal this data from Anthropic's 1 million user interviews

Why every marketer needs a GitHub

The next AI moat isn’t a better model

How much can you delegate to agents?

📰 News

Friend, the lonely AI wearable, returns with a new voice and a much bigger price tag

Google’s AI search is rapidly becoming the default

LinkedIn adds a button to report AI-generated ‘slop’

DoorDash is building its own drone delivery business

💸 Fundraising

Sila, a silicon-carbon anode materials startup, raises a $300M private equity round

Neo, an enterprise platform for securing and governing AI agents, raises $100M

Twenty, an offensive cyber warfare startup, raises a $30M follow-on investment

Natural, a payments infrastructure platform for AI agents, raises a $30M Series A

Fail(St)ory

Who Owns Search?

Presearch tried to build a private, community-run search engine and pay people to use and operate it with crypto tokens.

It shut down on this week, after nine years, caught between the cost of running search infrastructure and an audience too small to fund it.

What Was Presearch:

Presearch wanted to build a search engine that no single company controlled. Users would get more privacy, community members would help run the infrastructure, and the PRE token would reward both groups.

Its first product was a search dashboard. You entered one query and chose whether to send it to Google, DuckDuckGo, Wikipedia, LinkedIn, or dozens of other services. Most users kept using the default, so Presearch moved toward a normal search experience with one query and one results page.

Presearch did not crawl most of the web itself. It gathered results from other search engines and data providers, combined them, and displayed them under its own interface.

To process those searches, Presearch used computers run by members of its community. These computers, called nodes, received anonymized queries, fetched results from outside services, and sent them back to Presearch.

The goal was to spread the work across many independent operators instead of relying only on company-owned servers. This could lower infrastructure costs, make searches harder to link to individual users, and reduce the amount of control held by one company.

Anyone could run a node by installing Presearch’s software on an online computer with a public IP address. Operators staked PRE tokens and earned rewards based on the number of searches they handled, plus their speed and reliability.

Users also earned small amounts of PRE for searching, giving them a reason to switch from Google. Advertisers were supposed to fund the system through keyword ads, display ads, paid ad-free search, an AI subscription, and an API.

Presearch later started building its own index, called Indee, focused on forums, small publishers, and creator sites. The full plan was a private search engine with its own index, infrastructure supplied by its community, and tokens paying the people who used and operated it.

The Numbers:

🚀 Founded: 2017
👥 Monthly active users: Around 115,000
🔎 Daily searches before shutdown: Nearly 500,000
🖥️ Active community nodes: More than 38,000
💵 2024 revenue: $208,547
📉 2024 net loss: $267,414
🏦 Cash at the end of 2024: $23,735

Reasons for Failure:

Rewarded searches were hard to monetize: PRE tokens gave people a reason to search, but they also attracted bots and users chasing rewards. Presearch spent a lot of money checking traffic that advertisers might not value.
Search needed much greater scale: Presearch was close to 500,000 daily searches and still could not cover its costs. Management believed it needed one million legitimate searches per day, with no clear path to reach that level fast enough.
The system stayed expensive to run: Community nodes handled some work, while Presearch still managed routing, fraud checks, external data sources, ads, accounts, and the core search product. Its own index added another large technical cost.
Funding ran out first: Presearch ended 2024 with $23,735 in cash and raised only $183,328 through its 2025 crowdfunding campaign. New products and enterprise deals were still under development when the company needed immediate revenue.

Why It Matters:

Pay users only for behavior that creates clear economic value, because cheap actions like searches are easy to fake and costly to verify.
Decentralized infrastructure can increase costs when the company still handles routing, quality control, fraud, support, and monetization.
Model the exact usage level needed for break-even before scaling incentives, then prove revenue per user at a smaller size first.

Trend

Open-Weight Week

This week, everyone has been talking about open-weight models.

The long-awaited Kimi K3 finally released its downloadable weights, Nvidia backed a major industry letter, Anthropic published its objections, and Mark Zuckerberg joined the debate with an op-ed of his own.

Why it Matters

Open weights shift control from model providers to startups. You can run models yourself, keep data in-house, and adapt them—but you also take on infrastructure, security, and maintenance.
Most companies won’t run K3 directly. At 2.8 trillion parameters, it’s too large for most teams, which means value moves to hosting, inference, and model compression layers.
Every major player has a different incentive. Nvidia wants more compute demand, startups want flexibility, and frontier labs worry about losing control once weights are released.

Kimi K3

Moonshot AI launched Kimi K3 on July 16 and released the weights on July 27. It is the largest open-weight model released so far: 2.8 trillion total parameters, with 104 billion active at a time, native vision and a one-million-token context window.

K3 can handle long coding jobs, agent tasks, research, reasoning and work across text and images. Moonshot’s tests place it among frontier models, though still behind the strongest closed systems. The release drew so much attention because nobody had made a model this large available to download before.

Its mixture-of-experts design only activates a small part of the model for each token. That makes inference far lighter than running a dense 2.8-trillion-parameter model, though serving K3 still requires serious infrastructure.

The Controversy

While everyone was discussing Kimi, many of the biggest AI companies started debating the role open-weight models should have in the future of AI.

On July 24, Nvidia, Microsoft, Meta, Hugging Face, Mistral and more than 20 other companies signed a letter supporting open weights. OpenAI, Google, Amazon and others joined soon after. They argued that downloadable models give companies more choice, let them run AI on their own infrastructure and reduce dependence on a small group of model providers.

Nvidia then launched the Open Secure AI Alliance, which focuses on using open models for cybersecurity. The company pointed to a recent Hugging Face investigation where proprietary models refused some forensic tasks, while an open model could run locally and review more than 17,000 suspicious actions.

Anthropic published the most prominent counterargument. It supports open releases when the risks remain limited, but wants stronger testing as models become more capable in areas such as cybersecurity and biology. Once weights are released, safeguards can be removed, usage becomes hard to track and the model cannot be recalled.

The debate now comes down to where companies draw that line. Nvidia, Meta and Microsoft see broad access as useful for competition, security and customer control. Anthropic wants more evidence before the most capable models can be downloaded by anyone. Kimi arrived at the exact moment the industry was being forced to take a side.

The Trend

This week was a reminder that a large part of the AI race is happening among companies most people in the West barely follow. Moonshot turned Kimi K3 into one of the week’s biggest releases, then watched American tech companies spend days arguing over what models like it mean for competition, security and control.

These releases keep widening the field. Every capable open-weight model gives developers another serious option and puts more pressure on the major labs to explain what they will release, what they will keep closed and where they draw the line.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Another Agent Down

Nicolás Cerdeira — Thu, 23 Jul 2026 22:41:39 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Relay.app, an AI automation startup, has shut down — learn more below.
How to control what AI says about you
Twitter has gone from loving “AI Loops” to boasting about “AI Graphs” — learn what these actually are below

This Week In Startups

🔗 Resources

How to control what AI says about you

Pricing used to last 18 months. Now it's down to 6.

The 7 hires a hardware startup needs to get right.

2030-shaped software

📰 News

Anthropic updates Claude voice mode with more capable models

OpenAI makes ChatGPT Health available to all US users

OpenAI says Hugging Face was breached by its pre-release models

Moonshot AI Launches World’s Largest Open-Source Model

💸 Fundraising

Thira, an AI platform for automating enterprise back-office workflows, raises $21M

BRINC, a public-safety drone developer, raises $125M financing

Speakeasy, an operations platform for live-event businesses, raises $8.8M round

Fora, a platform for independent travel advisors, raises $60M Series D

Fail(St)ory

The Agent Pileup

Another week, another AI automation startup telling customers to export their workflows.

Relay.app helped teams stitch together the apps they already used, with AI doing the dull bits and humans checking the parts that could go wrong.

Last week, the company said it was shutting down. Customers got a short migration window. Everyone else got one more reminder that “AI teammate” is a cleaner pitch than it is a business.

What Was Relay.app:

Relay was an AI workflow builder. You’ve seen many of these, so you know the deal: connect your apps, pick a trigger, add a few steps, let AI do the boring work, and keep a human around for the parts where a bad answer could annoy a customer or break an internal process.

The use cases were the usual company plumbing: onboarding, lead routing, renewals, invoices, meeting briefs. The buyer was usually someone in ops, sales, marketing, HR, or customer success who knew the process because they were trapped inside it.

When the process was repetitive enough, Relay could save real time. The company said one customer cut the time from contract signing to technical setup by 80%. Another brought onboarding down from about a month to two or three days.

Relay’s cleaner idea was packaging these workflows as “AI teammates” with specific jobs. Instead of thinking in terms of random automations, a customer could build something that looked more like a support agent, onboarding agent, or research agent. Same plumbing, clearer mental model.

That price was quite competitive (starting from $19 per month) and came with a lot of product: hundreds of app connections, AI model options, forms, internal tables, reusable workflow blocks, MCP tools, and every weird customer workflow hiding behind a clean product page.

This was nice for customers but annoying for the nine-person team maintaining permissions, broken fields, API changes, rate limits, and whatever happened when someone’s “simple onboarding flow” turned into a 47-step monster.

By 2026, the whole category was flooded. Every automation tool was becoming an AI automation tool. Every agent startup was promising to run company workflows. Relay had useful features and real customers, but it was still selling into a market where the pitch had become very familiar and the maintenance work stayed very real.

The Numbers:

💰 Raised: $8.2M
👥 Team: 9 employees in March 2026
💳 Customers: About 1,000 paying customers
🧾 Spend: Around $100–$200/month for a typical business customer
🔌 Integrations: 200+ apps, with some pages claiming 350+
📅 Shutdown: July 16, 2026

Reasons for Failure:

A small team carried a big surface: Relay had hundreds of app connections, AI model options, workflow logic, forms, tables, approvals, and reusable blocks to maintain. Every API change, permission issue, broken field, or weird customer setup could turn into support work for a nine-person team.
The product was cheap for how much it had to do: A typical business customer spent around $100 to $200 per month. That is a tough price point when the product has to run important workflows, support many integrations, pay for AI usage, and keep improving in a crowded market.
The category got crowded fast: By 2026, every automation tool had AI features, and every agent startup claimed it could run company workflows. Relay had real customers and a useful product, but its pitch was getting harder to separate from the noise while the maintenance burden kept growing.

Why It Matters:

Its another reminder of how tough the AI automation category is nowadays.
It shows that cheap pricing gets adoption, but it leaves little room for support-heavy workflows.

Trend

While Loops, Rebranded

AI Twitter discovered “loops” in June. Developers were told to stop prompting agents and start “loop engineering”: give an agent a goal, let it work, check the result, and send it around again.

By mid-July, loops had already died. “Graph engineering” arrived with roadmaps, diagrams, courses, and the usual claims that everyone else was working at the wrong layer. One of the posts that started the graph wave was a joke about how fast the vocabulary changes. The timeline treated it as a product launch.

These constant re-brandings of not-so-new engineering ideas are a bit cringe in my opinion. Still, I think there is value in being able to understand what they are all about and how to apply them.

Why it Matters

The builders picked it up. Claude Code creator Boris Cherny and OpenClaw founder Peter Steinberger popularized loops. Anthropic had already published guides on agent loops and long-running harnesses.
It reached the conference circuit within weeks. LangChain published The Art of Loop Engineering, and AI Engineer World’s Fair put the topic on a their conference.
The renaming has serious reach. The post declaring loops dead and introducing graphs drew roughly 677,000 views, followed days later by LangChain’s 3 Years of Graph Engineering.

What are “AI Loops”?

In June, OpenClaw founder Peter Steinberger told developers to stop prompting coding agents and start designing loops that prompt them. The timeline soon filled with “Loop Engineering” guides, courses, diagrams, and people explaining that everyone else had been using agents wrong.

You could reasonably assume that a subject with this much educational material must be hard.

The truth is that an AI loop is very straightforward: the agent receives a goal, takes an action, checks what happened, and decides whether to continue.

The label now covers several setups. Some loops run inside the agent as it calls tools and reads their output. Others use an external script to restart the agent, a schedule to launch it every hour, or a verifier that rejects the work and sends it back. “Loop engineering” has become a broad name for making an agent do something repeatedly with less human input.

In every case, you give the agent a task, run it, check the result, and feed the latest state back into the next attempt. A basic script can keep doing that until a test passes, the task is complete, or a retry limit stops it.

What Are “AI Graphs” Then?

Ok, that is “Loop Engineering” clarified. So what the hell are graphs then?

An AI graph is a workflow drawn as a set of connected steps. Each node does something, such as calling a model, running a tool, checking a result, or asking a human for approval. The connections decide what happens next.

A simple example would be a support system that reads a ticket, sends billing questions to one agent, technical issues to another, and refunds above a certain amount to a human. Each route can have its own tools, rules, and retry loop.

You can build the same thing with normal code. The graph makes the routes, state, and failure points easier to see once the workflow starts branching or several tasks need to run at the same time.

Loops also live inside graphs. A failed test can send one node back for another attempt, while the rest of the workflow stays finished. That is the main reason graphs become useful: one part can branch, pause, retry, or wait for approval without restarting everything.

The Trend

AI Twitter keeps turning familiar software patterns into new engineering disciplines, then replacing them before the definitions settle.

Nevertheless, each term does point to a real design choice that is often worth paying attention to. That is, without paying $100 usd for a course that explains a while loop.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

The Maggot Robot

Nicolás Cerdeira — Thu, 16 Jul 2026 17:39:11 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Goterra a maggot-powered food waste startup, has shut down — learn more below.
The new MVP (Meme Viable Problem), and a 3-step playbook to using it
OpenAI had a busy week — learn about all the new products below

This Week In Startups

🔗 Resources

The new MVP (Meme Viable Problem), and a 3-step playbook to using it

How to measure the impact of AI search the right way

LLMs are picking winners. Here’s how to become one

The word that kills startups is "Interesting"

📰 News

Apple sues OpenAI over alleged trade secret theft

OpenAI is shutting down Atlas

OpenAI releases a $230 keyboard for Codex

Meta’s new AI chips will begin production in September

💸 Fundraising

Vendelux, an AI platform for measuring B2B event performance, raises ⁠$50M Series B funding

Quaise Energy, a developer of deep geothermal drilling technology, raises ⁠$134M Series B funding

Venus Aerospace, a developer of rotating-detonation rocket engines, raises ⁠$91M Series B funding

Ollama, an open-source platform for running AI models, raises ⁠$65M Series B funding⁠

Fail(St)ory

Waste to Protein

Goterra made big, shipping-container machines that used maggots to turn food waste into useful stuff like protein and fertilizer.

It shut down in 2026 after running out of money. The core problem was pretty simple: there was interest in what they were doing, but every bit of growth meant building more expensive hardware first.

What Was Goterra:

Goterra sold organic-waste processing infrastructure. Its core system, called MIBS, was often described as a “maggot robot”: a modular unit that received food waste, processed it, fed it to larvae, and harvested useful outputs.

The product made most sense for customers with steady organic waste: supermarkets, airports, hotels, city programs, commercial kitchens, and food processors. These buyers were paying to avoid landfill, reduce waste transport, and improve their emissions story.

The unit was designed to sit near the waste source. This was important because wet food waste is heavy, smelly, and expensive to move. A decentralized plant could, in theory, make waste treatment closer to a service network than a large remote facility.

Inside the container, automation handled feeding, material movement, environmental conditions, and parts of harvesting. Larvae consumed the waste over roughly 12 days, turning it into insect biomass and frass.

Goterra had two revenue pools:

Customers or waste partners paid service fees to divert waste.
The company could also sell larvae as protein ingredients and frass as fertilizer or soil conditioner.

That second pool made the model attractive on paper. Goterra was paid to receive feedstock, while many insect-protein companies had to buy inputs. If the waste fees covered a big part of processing cost, protein sales could improve the economics.

The operating reality was heavier. Each contract required enough units, site work, maintenance, permits, collection logistics, uptime, and working capital. A customer signing a multi-year deal did not instantly become revenue unless Goterra could build and deploy capacity.

The company said it had contracted demand on both sides: waste-management services and protein offtake. The bottleneck became manufacturing infrastructure quickly enough to activate that demand.

By the end, the business looked like an infrastructure company with startup funding. Revenue was growing, but the gross margin was deeply negative because the system was still being developed, operated, and scaled at the same time.

The Numbers:

🗑️ Processed: 35,000+ tonnes
📈 Revenue: A$365k → A$1.43M
💸 Gross margin: -198%
🧾 Net loss: A$11.74M
🧱 Spend: ~A$25M
🔚 Liquidation: July 8, 2026

Reasons for Failure:

Capacity had to be built before revenue could catch up: This wasn’t SaaS, you couldn’t just flip a switch and scale. Every new customer needed real machines, real installs, and real upkeep. Goterra had contracts lined up, but without enough units in the field, that demand stayed stuck on paper. When funding slowed, so did everything else.
The unit economics were still moving in the wrong direction: Revenue was growing, but costs were sprinting ahead. In 11 months, Goterra made about A$1.43M while spending A$4.26M to do it. The hope was scale would fix it, but the runway ran out first.
Liquidity depended on tax offsets and bridge funding: A big chunk of the balance sheet came from expected R&D tax credits. That’s fine until you need cash now. Goterra borrowed against those future refunds, which only works if new funding arrives on time. It didn’t.
The final financing had too many conditions at once: Goterra needed about A$25M to keep scaling, but the deal never quite came together. A key investor didn’t sign in time, which also killed the bridge funding. Plenty of buyers looked, but no one wanted to take on both the tech and the heavy capital bill.

Why It Matters:

Demand only matters when you can serve it. Contracts stayed on paper without enough deployed units.
Growth can make the cash problem worse. Every new customer required more hardware and working capital upfront.

Trend

OpenAI’s New Stack

OpenAI had a busy week filling in the stack around agents. There’s a new model (actually three models), a whole new app, and a pretty impressive new voice mode.

Here’s everything about it.

Why it Matters

Model choice gets simpler and cheaper. Three tiers map cleanly to task difficulty, so you can match cost to need without overpaying.
Long tasks and voice both level up. Work handles end-to-end assignments, while Live makes conversations fluid and interruptible.

GPT-5.6 now comes in three sizes

GPT 5.6 is out and the biggest change is the model lineup.

Sol is the expensive flagship for hard reasoning, coding, science, cybersecurity, and long tasks. Benchmarks put it at or near the top, alongside Fable.
Terra is the middle option, with performance OpenAI compares to GPT-5.5 at roughly half the token price.
Luna is the fast, cheap model for high-volume work.

OpenAI also launched Programmatic Tool Calling. GPT-5.6 can generate a small JavaScript program that runs inside OpenAI’s sandbox. That program can orchestrate multiple tool calls, loop over results, filter and deduplicate data, and transform it before returning anything to the model.

In practice, the model offloads structured, repetitive work to code. The script pulls exactly what it needs, cleans it up, and compresses it into a smaller, more relevant payload. The model then reasons over that reduced dataset.

One more thing: OpenAI also announced GPT-Red, an internal model trained to attack other models and agents. It tries prompt injections hidden in emails, webpages, files, and tool outputs, then feeds the successful attacks back into training. GPT-5.6 was one of the models trained against it.

ChatGPT Work

GPT Work is essentially OpenAI’s answer to Claude Cowork: Codex for non-technical work.

You give it a larger assignment, it breaks the work into steps, pulls information from connected apps and files, uses the browser or desktop, and keeps going until it has something finished.

The actual new product is the app around it. OpenAI merged Chat, Work, and Codex into one desktop app for macOS and Windows. Existing Codex users get it through an update, while the previous ChatGPT desktop app is now called ChatGPT Classic.

You can chat, hand off a longer Work task, or jump into Codex from the same app. It is a much cleaner way to understand how OpenAI now sees the product: three modes for three different kinds of work, sitting next to each other.

GPT-Live finally feels live

GPT-Live is pretty cool.

The old voice experience still had the rhythm of a walkie-talkie. You spoke, waited, then the model answered. GPT-Live listens and speaks at the same time, so you can interrupt it, pause mid-sentence, change direction, or keep talking while it figures out when to respond.

It can also send harder work to another model in the background. You might talk through customer notes, ask it to check previous emails, keep discussing the account, and get a structured brief once the deeper work is done.

That combination makes voice much more useful. Before, it seemed more like a toy and I would always avoid it for actual work. This week, I am not ashamed to confess that I have spent more time speaking with it than typing to it.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Send Pics

Nicolás Cerdeira — Thu, 09 Jul 2026 18:08:21 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

PicSee, an app that use face recognition to find your photos on your friends’ phones, shut down — learn why below
How tech workers are feeling in 2026: a workforce splitting in two
Cloudflare announced Monetization Gateway, a new business model for the web — learn why this matters below

This Week In Startups

🔗 Resources

How tech workers are feeling in 2026: a workforce splitting in two

How we booked 614 meetings with one inbound agent

The AI cost crisis is entirely self-inflicted

GTM Strategy 101: ICPs, Positioning, Messaging & More Explained

📰 News

Anthropic, OpenAI, and SpaceX are bigger than the last 25 years of tech exits

SpaceXAI releases Grok 4.5

OpenAI releases new voice models for more natural live conversations

Meta just launched a new AI generator

💸 Fundraising

Together AI, an AI cloud platform for training and running open-source models, raises $800M Series C funding⁠

TwelveLabs, a video intelligence platform for search and analysis, raises $100M Series B funding⁠

OXMIQ Labs, a licensable AI chip architecture startup, raises $35M Series A funding⁠

Queue, an autonomous robotic pharmacy startup, raises $12.6M seed funding⁠

Fail(St)ory

Where’s My Photo?

PicSee tried to fix a very annoying social problem: your best photos are often sitting on someone else’s phone.

The app launched publicly in October 2025 and shut down in July 2026, after failing to turn that irritation into a large, self-propelling network.

What Was PicSee:

PicSee wanted to help you get every photo your friends have of you, without begging them to send it.

The app looked at a behavior everyone already understands. Friends take pictures of each other at trips, weddings, parties, college events, dinners, and family gatherings. The photos have value, but the act of sharing them is annoying enough that people delay it, forget it, or dump everything into a chaotic group chat.

PicSee used facial-recognition AI to scan a user’s gallery and identify people in their photos. It could then help invite those people to connect. Once both sides approved the relationship, the app could automatically exchange photos they had taken of each other.

The core loop was “give to get.” A user received photos of themselves by sharing the photos they had of others.

The product tried to remove the small chores that kill sharing. No album creation. No link. No manual upload. No scrolling through a camera roll while deciding which photos are safe to send. The app could keep finding new photos later, so the exchange was meant to become recurring once the relationship was approved.

The awkward part was that one user could not make the product useful alone. If you joined PicSee after a trip and your friends ignored the invite, the magic stopped there. The app needed clusters of real-world relationships to move together.

That made the business harder than the product demo suggested. PicSee was free at launch and expected to monetize later through subscriptions or premium features. Before any of that mattered, it had to prove that users would pull their actual friends and families into a new sharing habit without heavy paid marketing.

The Numbers:

🏢 Founded: 2024
🚀 Public launch: October 2025
🛑 Shutdown: July 2026
📸 Photos shared: 150K+ by October 2025
📱 Downloads: 10K+ on Google Play
💰 Funding: Around $4.25M raised
👥 Team: 12 people
🔁 Capital returned: 60–65%

Reasons for Failure:

PicSee needed friend groups to arrive together: A normal consumer app can often win one person at a time. PicSee had a tougher path because one download only became useful when the right people joined too. Founder Mayank Bidawatka described the challenge well: the company was trying to “acquire networks,” and “there isn’t a distribution channel in the world that lets you do that.” The app needed your friends, your family, your travel group, or your wedding guests, not just another anonymous install.

The old behavior was messy but deeply embedded: People complain about WhatsApp photo dumps, but they still use WhatsApp because everyone is already there. The same applies to iMessage, Instagram, Snapchat, Google Photos, and shared albums. PicSee had to move a low-frequency habit into a new app and then get the user’s social circle to follow. The pain was obvious after an event, but it faded quickly once the moment passed.

Paid growth created misleading signals: Much of PicSee’s user growth came from “junk signups” acquired through Meta ads that did not convert into revenue. That is especially dangerous for a social product because installs can look like traction while adding almost no useful density. A user in one city does not help someone recover photos from their cousins, classmates, or travel friends somewhere else. PicSee needed connected groups, and ads were better at producing scattered users.

Why It Matters:

A relatable problem is not automatically a habit. Missing photos is annoying, but the pain shows up around specific events, not every day.
Social products should measure density, not just signups. PicSee needed connected groups joining together, not scattered users from ads.

Trend

A New Internet Economy

Last week, Cloudflare announced Monetization Gateway, a way for websites to charge AI agents when they access pages, APIs, datasets, and tools.

On its own, it sounds like a niche infrastructure product. Put it next to AWS adding AI traffic monetization, x402 moving into the Linux Foundation, and publishers preparing for lower search traffic, and the signal gets clearer.

The internet economy is starting to move away from charging for human attention and toward charging for machine consumption.

Why it Matters

The request becomes the new pageview. If agents consume the web without sending humans back to websites, founders need to think about pricing access directly, not only capturing attention after the click.
Infrastructure companies are becoming the toll booths. Cloudflare and AWS are putting payment, identity, and access rules at the edge, which means machine monetization may be configured more like security rules than checkout flows.
Open content is getting harder to justify. Pages, APIs, datasets, docs, and community archives can all become raw material for AI systems, so startups need a clearer answer for what stays free, what gets blocked, and what gets priced.

The Problem

The old web bargain was simple enough: let search engines crawl your site, get traffic back, then monetize that traffic through ads, subscriptions, affiliates, or leads.

AI agents weaken that bargain because they can take the useful part without creating a visit. A user asks for an answer, comparison, summary, recommendation, or action. The agent fetches the page, extracts what it needs, and returns something useful inside another interface.

Some traffic still comes back, but the ratio is ugly. Cloudflare has shown AI crawlers fetching far more pages than they refer visitors back to, which leaves the original site with the expensive side of the relationship: producing content, maintaining data, serving pages, and handling bot traffic, while the monetizable moment moves somewhere else.

Cloudflare’s Solution

Cloudflare’s Monetization Gateway wants to turn access control into a payment flow.

A site owner can put a price on specific resources behind Cloudflare: a page, an API route, a dataset, or an MCP tool call. When an AI agent requests that resource, Cloudflare can respond with HTTP 402, the old “Payment Required” status code, plus the price and payment instructions.

The agent pays (initially through stablecoins), sends proof of payment, and gets the resource. No account, no checkout page, no sales call, no custom billing setup.

The Trend

Cloudflare is not moving alone. There have been several signals these past months that all support the same trend: the internet is moving from charging attention to charging access.

AWS had already added AI traffic monetization to WAF: Publishers and API owners can return an HTTP 402 response to bots and agents at the edge, then serve the content after payment is verified.
x402 moved into the Linux Foundation: x402 is the payment protocol behind much of this: it lets a server ask for payment inside a normal HTTP request. Coinbase started it, but the Linux Foundation gives it a more neutral home, with early support from Cloudflare, Stripe, AWS, Visa, Mastercard, Google, Microsoft, and Shopify.
Crawler controls are becoming more granular: Cloudflare is separating search, agent, and training crawlers, which lets site owners treat each type differently instead of choosing between blocking all bots or giving everything away.
AI answers are already reducing click-through: A 2026 academic study on Google AI Overviews and Wikipedia found that exposure to AI Overviews reduced daily traffic to affected English Wikipedia articles by about 15%. If answer engines can satisfy the user before the click, content owners need a way to monetize consumption earlier in the chain.
Large content owners are signing licensing deals: News Corp and Reddit, and others have shown there is real money in AI access to content. Request-level pricing is the version that could work for smaller sites without a custom enterprise deal.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Paging Dr. WhatsApp

Nicolás Cerdeira — Thu, 02 Jul 2026 17:13:49 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

JiviAI, a startup building an AI Whatsapp doctor, has shut down — learn why below
How to steal a customer base
Anthropic just released Claude Science, a new AI workbench for scientists — learn why this matters below

This Week In Startups

🔗 Resources

How to steal a customer base

AI's Optimisation Era

Frontier no more? The five forces challenging Frontier Models

If your product hasn't taken off

📰 News

Google introduces a faster, cheaper image generator with Nano Banana 2 Lite

Anthropic launches Claude Sonnet 5 as a cheaper way to run agents

OpenAI releases GPT 5.6 only to trusted organizations

Anthropic restores Claude Fable 5 worldwide as US lifts export controls

💸 Fundraising

Patronus AI raises $50 million to develop digital environments for testing AI agents.⁠

⁠Assort Health scores $120M series C to scale voice AI agent platform for healthcare

Attention Raises $30 Million Series B To Build AI Platform For Revenue Teams

Coval raises $28M as enterprises push voice agents into production

Fail(St)ory

The Cost of Being Everyone’s AI Doctor

JiviAI tried to put a doctor-like assistant inside a phone, a WhatsApp chat, and a clinical AI platform. They shut down in June 2026 after less than three years in market.

The tension was clear: healthcare AI can attract millions of curious users very quickly, while running medical-grade models, agents, voice systems, and multilingual support can burn cash faster than those users create a business.

What Was JiviAI:

JiviAI sold the idea of instant health guidance. Its consumer product was positioned as a personal AI doctor that could answer symptom questions, explain reports, track health signals, and help users make everyday decisions around stress, sleep, diet, and care.

The obvious use case was the moment before a doctor visit. A user has a fever, a confusing blood report, chest discomfort, or anxiety about a symptom, and wants a plain-language answer now. Jivi gave them a low-friction first step.

The app promised symptom checking based on medical cases and clinical data. Users could describe what they felt and get possible explanations, guidance on when to rest, and advice on when to see a doctor.

Report explanation was another strong wedge. Medical reports are full of ranges, acronyms, and scary-looking flags. Jivi translated that into simpler language, which is useful even for people who already have access to doctors.

The product also moved into mental health and lifestyle coaching. It offered stress, sleep, and mood self-assessments, plus AI chat using CBT and DBT-style methods.

Jivi’s reach depended on accessibility. Public app copy claimed 100+ languages, and the company pushed WhatsApp as an entry point.

That wide surface area made Jivi more ambitious than a simple symptom checker. It was trying to be a consumer health companion, a multilingual access layer, and eventually a clinical AI platform. The user experience looked lightweight, but the company behind it had to carry medical AI infrastructure, language coverage, and trust expectations on every interaction.

The Numbers:

🏥 Founded: 2023
📍 HQ: Gurugram, India
💸 Funding: $2.99M across one seed round
👥 Team: 11–50 employees publicly listed
📱 Users: 1.5M+ claimed in app-store copy
🌍 Reach: 3.5M users across 200 countries claimed later
🛑 Shutdown: Reported on June 26, 2026

Reasons for Failure:

AI health had expensive usage before it had durable revenue: Jivi claimed millions of users across many countries, but health queries are long, sensitive, multilingual, and costly to serve. Free or cheap guidance only works if it converts into paid care, provider workflows, insurance deals, or enterprise contracts.
Proprietary medical AI put Jivi in a costly race: Jivi was promoting MedX, medical LLMs, voice AI, multimodal systems, agents, and open-source model work. That is a wide technical surface for a seed-stage company. Larger model labs keep improving the baseline and lowering prices, which makes proprietary differentiation expensive to defend.
The product straddled consumer and clinical markets: Consumer health apps need trust, habit, and low acquisition cost. Clinical platforms need procurement, compliance comfort, workflow integration, and clear accountability when advice affects care. Jivi’s public materials spoke to both sides, which made the vision bigger and the execution burden heavier. A user may like instant report explanation, while a clinic needs reliability, auditability, escalation rules, and a clear operational owner.
The financing path closed before the model matured: A planned funding round did not close, and acquisition talks failed to materialize. That left Jivi trying to fund a heavy medical AI platform before usage had become a repeatable revenue engine. Once a company has committed to proprietary models and broad platform claims, it has limited room to shrink into a small, cheap experiment.

Why It Matters:

Usage in AI healthcare can be misleading when sessions are easy to start but hard to monetize.
Building proprietary models changes the funding requirement from app-company money to infrastructure-company money.
Medical AI gets cleaner when one buyer has the budget, workflow pain, and accountability to pay early.

Trend

AI Scientists

This week, Anthropic added a new Claude to its growing family: Claude Science.

Its announcement did not generate the same splash as Claude Code or Claude Cowork, but this product is worth watching.

Claude Science is Anthropic’s attempt to move Claude into the actual workflow of scientists: papers, datasets, code, figures, compute jobs, citations, and all the annoying glue work that sits between them.

Why it Matters

Accuracy becomes inspectable: Claude Science attaches the code, environment, explanation, and history behind an output. That pattern should spread to any AI product where users need to trust the result before acting on it.
Chat is turning into a workbench: The product sits inside the workflow: reading, coding, running jobs, making figures, revising them, and keeping the path back. The interface matters less than the job history.
Domain context becomes the moat: Claude Science is useful because it knows the tools, databases, file types, and compute setup scientists already use. In vertical AI, the generic model is only the starting point.

Claude Science

Claude Science puts Claude inside the scientific workflow: reading papers, working with datasets, running code, making figures, checking citations, and helping prepare research outputs.

The product combines tools scientists usually use separately. Instead of jumping between PubMed, Jupyter, R, terminals, scientific databases, file viewers, and custom scripts, a researcher can coordinate more of that work from one place.

Claude Science is preconfigured for fields like genomics, single-cell biology, proteomics, structural biology, and cheminformatics.

It can connect to scientific databases, journals, preprints, and domain-specific models, so researchers do not have to manually pull context from a dozen separate places.

It can produce research artifacts, not just text. Claude Science can run analyses, create figures, render 3D protein structures, show genome browser tracks, and work with chemical structures.

Claude Science is also trying to make the work easier to trust. When it creates a figure or artifact, it keeps the code, environment, explanation, and message history behind it. A scientist can inspect the steps, rerun the work, catch mistakes, and challenge the result instead of treating the output like a black box.

Other Signals

OpenAI GPT-Rosalind: OpenAI’s life-sciences model is built to help researchers reason through real research tasks, not just answer biology questions. It can pull evidence, analyze biological data, help design experiments, troubleshoot lab protocols, and use tools to produce reviewable outputs. It was launched in April.

Google DeepMind Co-Scientist: DeepMind published Co-Scientist in Nature in May. It is a multi-agent system that generates, debates, ranks, and refines scientific hypotheses. The “tournament of ideas” setup is the interesting bit. It treats hypothesis generation as something agents can compete over, improve, and review before a human decides what deserves attention.

Google Gemini for Science: Google also packaged Co-Scientist into a broader Gemini for Science push. The bundle includes Hypothesis Generation, Computational Discovery, and Literature Insights. That gives Google coverage across three common jobs: coming up with ideas, testing many computational approaches, and turning literature into something researchers can actually use.

FutureHouse Robin: Robin was published in Nature in May and is focused on biological discovery. It combines literature search, hypothesis generation, experimental planning, data analysis, and follow-up insight generation.

Amazon Bio Discovery: AWS launched Amazon Bio Discovery in April. It gives scientists access to 40+ biology models, helps them choose models and inputs, and connects computational candidates to lab testing. In one MSK collaboration, AWS said the platform generated nearly 300,000 antibody molecules and narrowed them to 100,000 candidates for testing.

The Trend

AI-for-science is turning into its own product category.

Claude Science is the cleanest example this week because it feels like a real workspace, not a demo pretending to be a lab. The bigger point is that OpenAI, Google, Microsoft, AWS, FutureHouse, and Anthropic are all circling the same opportunity.

Science has huge workflows full of expensive people, fragmented tools, private data, slow review loops, and repetitive analysis. AI products that can live inside those workflows, leave a trail, and earn trust will be much more valuable than another smart sidebar.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

The Empty Factory

Nicolás Cerdeira — Thu, 25 Jun 2026 20:03:44 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

SWAP Food, a startup making plant-based chicken, shut down — learn more below
Using context-engineering to 5x conversion and 2x activation
Founders are using AI assets to market non-exist products and validate ideas — learn all about this strategy below

A huge thanks to today’s sponsor, HubSpot. Download their free AI guide to supercharge your workday.

100 ChatGPT Prompts to Revolutionize Your Day ^AD

Supercharge your productivity with HubSpot's comprehensive guide. This free resource is your fast track to AI mastery:

Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies

Plus, in-depth sections on email composition, content creation, customer support, and data analysis.

Find Your Guide Here

This Week In Startups

🔗 Resources

How to pivot well

What AI-native GTM looks like at public company scale

Using context-engineering to 5x conversion and 2x activation

Your unfair advantage in the age of AI

📰 News

OpenAI unveils its first custom chip

Anthropic released Claude Tag

New data suggest engineering jobs are the most resilient

Anthropic’s Claude is winning over paid consumers, a market owned by ChatGPT

💸 Fundraising

Arcade.dev, a secure action layer for production AI agents, raises $60M⁠

Convey, an AI teammate platform for enterprise operations, raises $38M ⁠

Verse, an energy infrastructure platform for AI data centers, raises $54M ⁠

HyperLight, a photonics chip startup for AI infrastructure, raises $80M

Fail(St)ory

Plant Chicken

SWAP Food raised over €100M to make plant-based chicken look and cook like a real chicken breast.

By 2026, the Paris-based company was in judicial liquidation. Its factory in Alsace had been built for industrial scale, but the installed line ran at under 10% capacity.

This is a story of what happens when you bet on industrial scale before the market is ready to absorb it.

What Was SWAP Food:

Most plant-based meat companies started with the easy shapes: burgers, nuggets, sausages, mince. SWAP went after the harder format.

It wanted to make a plant-based chicken breast that looked like a real filet, cooked like one, and could sit in the middle of a plate without looking like a compromise.

SWAP’s technology was called Umisation. In simple terms, it was a way to turn plant protein into long, meat-like fibers without the usual heavy extrusion process. The company wanted whole-cut texture from soy and pea protein, with a shorter ingredient list than many alt-meat products.

Its main product, SWAP Chicken, was a thick plant-based filet with 19g of protein per 90g serving. It used eight ingredients and avoided GMOs, added gluten, cholesterol, artificial flavors, colorants, and texturizers.

The pitch to restaurants was simple: use it like chicken. Put it in a sandwich, bowl, salad, or plated dish without rebuilding the menu. For chefs, that was the appeal. They could add a plant-based option without making it feel like a separate vegan menu item.

The 2024 rebrand from Umiami to SWAP made the proposition clearer. The company was selling a swap from animal protein to plant protein. The target was the mainstream diner who still ate meat but wanted to cut back, not the tiny group already committed to vegan food.

That made SWAP different from many alt-meat startups. It was trying to solve the two complaints that kept hurting the category: weak texture and products that felt too processed. A chicken-style filet with a clean label gave foodservice buyers a better story to tell.

So, what do you do when you’re still proving a new food technology and the market has not fully shown up yet? You build a giant factory, apparently.

SWAP made a huge bet on a production site in Alsace. The plant was built for 7,500 tons of annual capacity, with plans to reach 20,000 tons. By the end, the installed line was reportedly running at under 10% capacity.

The Numbers:

💰 Funding: €100M+
🏭 Factory: 7,500 tons of annual capacity at launch, with plans to reach 20,000 tons.
📉 Utilization: Under 10% capacity.
🧾 Debt: Around €16.5M
🔥 Burn: Around €1M per month after cost cuts.
👥 Team: 66–70 employees

Reasons for Failure:

The factory made small demand look ridiculous: A pilot line can live with messy early sales. SWAP had already built the expensive version: rent, debt, salaries, utilities, maintenance, and a lot of stainless steel waiting for orders. When utilization sits under 10%, every slow week burns money in the most physical way possible.

The product was too premium for a cheap-protein fight: SWAP Chicken sold at around €20/kg. That can work in a nice restaurant where the menu needs one strong plant-based option. It is much harder when the benchmark is ordinary chicken, one of the most efficient proteins on the planet.

The balance sheet left no patience for a slow category: SWAP was carrying debt, heavy losses, and an underused factory while the plant-based meat market grew slower than expected. Management cut fixed costs by 50%, but losses were still around €1M per month. The company needed fresh financing to keep going. Investors had become much less willing to fund companies that needed industrial-scale capital before proving industrial-scale demand.

Why It Matters:

Industrial scale turns slow adoption into an expensive fixed-cost problem very quickly.
“Drop-in replacement” still has to win on price, format, storage, prep, distribution, and menu economics.
A category can be growing and still be too small for one company’s factory, debt, and burn.

Trend

Synthetic Validation

This week I saw a tweet that caught my attention because it was a perfect example of a marketing strategy I am seeing more and more of. The tweet was about BYLT, a Kickstarter campaign for miniature construction kits, and how an AI-generated promo video helped turn a stalled campaign into real pledge demand.

People have been calling this strategy a bunch of things: AI smoke tests, synthetic MVPs, testing the fantasy first. I’ll use Synthetic Validation, mostly because it gets closer to what is actually happening: founders are using AI-generated assets to test demand before the real product is fully built.

Why it Matters

The launch asset is becoming the first product test. Founders can now test whether people want the story, use case, aesthetic, or emotional payoff before spending months on production.
Realism changes the trust equation. A rough mockup tells users they are looking at a concept. A polished AI video can make an unfinished product feel like a finished one, which pulls more attention and also raises the bar for disclosure.
The asset can backfire once people realize it is AI. BYLT’s video helped drive attention and pledges, but also triggered backlash over what was real, generated, or actually finished.

The BYLT Story

BYLT: Real Construction, Tiny Scale was a Kickstarter campaign for miniature home-building kits. The promise was oddly specific and very internet-friendly: build miniature homes using real construction methods, including blueprints, a concrete foundation, framing, electrical, siding, and finishing.

After launch, the campaign became stagnant for around 15 days. Then Ethan, the founder, posted a sleek AI-generated launch video. It reportedly crossed 5M views and helped bring in more than $30K in pledges for a product that was still basically non-existent.

Of course, this was not all good. The video also got a lot of backlash once people realized it was AI-generated. Some called the campaign fake, others questioned what was real, what was rendered, and how much of the product had actually been built. Kickstarter later suspended the campaign.

But the key thing is that we now know there is potential in Ethan’s idea. Before the video, miniature construction kits could have sounded like an insane niche product. After the video, it looked like something that might actually have PMF.

And Ethan tested that with basically zero production budget. That is the power of synthetic validation: you can test the fantasy before building the product. Because the video looked like a finished product launch, rather than the classic Kickstarter video of a founder talking into a camera and pitching a concept, people got more excited and were more willing to pledge.

Other Signals

Wonders of the First, a collectible card game, raised more than $1.4M on Kickstarter while using a lot of AI art. The campaign showed the same dynamic at a bigger scale: AI assets can make a project look polished fast, but they also make backers inspect the whole thing more aggressively.
Maket used AI-generated influencers to test creative concepts and marketing hooks before spending more on campaigns. That is synthetic validation applied to messaging: cheap fake people, real customer reactions.
Pieter Levels had already described the indie playbook. In 2023, he suggested validating ideas with a mockup, daily TikToks, Linktree, Stripe preorders, and repetition across many ideas. AI makes the mockup, video, product shots, copy, and variants cheaper to produce.
A 2025 Alibaba-linked paper described a “sell it before you make it” system where merchants generate photorealistic fashion items from text, show them to customers, and produce after enough orders come in. The paper reported over 13% relative improvements in CTR and conversion versus human-designed counterparts.
Kickstarter is now requiring creators to disclose when they use AI-generated images, text, or other outputs, and explain what was generated versus original. That is a signal by itself. Platforms do not write policies for edge cases unless enough projects are already doing it, and synthetic campaign assets are now common enough to need their own trust rules.

The Trend

Synthetic validation is becoming a normal part of startup and creator workflows. The founder no longer has to start with the product.

They can start with the promise, package it into a believable asset, distribute it, and measure whether anyone reacts strongly enough to pay, pledge, or preorder.

That makes the next question less about whether you can make something look real, and more about how clearly you tell people what they are actually buying.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

The $500 AI Accountant

Nicolás Cerdeira — Thu, 18 Jun 2026 20:25:42 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

FINNX, a startup building an AI agent for finance teams, has shut down — learn why below
Your vendor is now your competitor
OpenRouter Fusion is the latest tool in a new way of using AI models — learn why it matters below

This Week In Startups

🔗 Resources

Your vendor is now your competitor

How to build your AI GTM system

Why we're bullish on loops

AI customer support for startups. Up to 90% off your first year *

📰 News

The US government shut down Anthropic’s latest Claude AI model

SpaceX valuation balloons to $2.6T, briefly passes Amazon

Snap finally debuts its long-awaited AR glasses

ChatGPT’s market share slips below 50% for first time

💸 Fundraising

Standard Bots, an AI-native industrial robotics manufacturer, raises $200M Series C funding⁠
Jedify, a context-graph platform for enterprise AI agents, raises $24M Series A funding⁠
Turnout, an AI-powered consumer advocacy service, raises $35M Series A funding⁠
Reset, an earned-wage access platform for credit unions and community banks, raises $6M seed funding⁠

* sponsored

Fail(St)ory

Would you hire Edith?

FINNX built Edith, an AI finance employee for accounting teams buried in manual work. It shut down in June 2026, after nearly two years of building.

Another vertical AI startup that found a painful workflow and learned how far a convincing demo is from software trusted to run it.

What Was FINNX:

FINNX sold Edith as “your first finance team member.” The product was built for the daily accounting work that falls between software, spreadsheets, inboxes, and someone following up when the process breaks.

Edith was meant to pull invoices from email, post them into the right systems, match them to purchase orders, flag unusual items, chase vendors, and send messy cases to a human. Month-end reconciliation was part of the pitch too.

The buyer was a finance team that already had accounting software, AP tools, and approval workflows, yet still spent time cleaning up missing information and mismatched records.

Their thesis was that existing finance software handled discrete tasks but left the exception-heavy “last mile” behind: missing documents, PO mismatches, timing differences, and judgment calls.

FINNX wanted Edith to sit in that gap. The company’s pitch was that finance tools automate individual tasks, while Edith would carry a process through to completion.

Edith was built around deterministic, classical AI rather than LLM-first agents, with an emphasis on rules, traceability, and auditability. Finance teams need to understand why an invoice was matched, why an exception was flagged, and how a number reached the ledger.

This meant the team had to build a lot before Edith could operate safely. The product needed access to accounting systems, bank feeds, ERP tools, AP workflows, approval rules, and each customer’s own tolerances.

A vendor match that works for one company can be wrong for another because of different chart-of-account structures, approval policies, tax treatment, or purchasing habits.

FINNX advertised Edith at $500 per month and compared it with a $70,000-a-year finance hire.

More than 700 companies joined the waitlist before the shutdown. That showed showed that finance teams wanted the outcome; FINNX still had to prove Edith could deliver it inside live accounting workflows.

The Numbers:

🏢 Founded: 2024
📍 Headquarters: Singapore
👥 Team size: Around 5 people
🧾 Waitlist: 700+ companies
💸 Capital sought: $1.5M in March 2026

Reasons for Failure:

One bad entry mattered: A wrong invoice match, tax treatment, or journal entry could create cleanup work and audit risk for the customer. The founder said the team “ran out of runway before she was ready.” Until Edith could handle those mistakes safely, finance teams still had to check the work themselves.
The technical scope was large for a five-person startup: Edith needed to connect with accounting systems, bank feeds, ERP tools, approval rules, and each customer’s own finance logic. Building that safely required more integration and infrastructure work than a tiny team could finish quickly.
Buyers needed category education: FINNX had to explain why its deterministic approach mattered and why a generic AI agent was not enough for accounting. That added work to every sales conversation before the company could even get to implementation or pricing. The founder later said explaining the difference between FINNX and generic AI agents had become a full-time job.
The waitlist did not provide enough proof for the next round: The waitlist was not enough to change the financing outcome. More than 700 companies had signed up to hear about Edith, but the founder said the company’s final VC diligence process made it clear that raising the capital needed to finish the model was unrealistic.

Why It Matters:

Vertical AI is easy to pitch and hard to trust. The closer the product gets to owning a core workflow, the higher the bar for reliability.
Waitlists do not reduce implementation risk. Interest is cheap; live customers still need to connect real systems and let the product make real decisions.

Trend

Answer Fusion

OpenRouter shipped a very cool feature last week that deserves more attention than it got.

Fusion sends the same question to several AI models, compares what they say, spots where they disagree, and produces one final answer.

It sounds like a minor upgrade to a model API, but it really points to a different way of building AI products: treat a single model response as a draft, then give the system a way to challenge it before the user sees it.

Why it Matters

The best AI answer may come from a process, not a model. A strong model can still miss an edge case, follow a bad assumption, or give a convincing answer with thin reasoning. Fusion adds a second layer that looks for overlap, disagreement, and missing coverage before producing the final response.
Startups can buy more confidence only when it matters. Running several models costs more and takes longer, so this will not replace everyday chat.
Model choice becomes less visible to the user. Customers rarely care whether an answer came from GPT, Claude, Gemini, or a cheaper open model. They care whether the answer feels complete, catches obvious issues, and gives them a reason to trust it.

OpenRouter Fusion

Most AI products still follow a simple pattern: ask one model a question and return its answer.

Fusion changes the sequence. Several models first work on the same prompt independently. A judge then reviews their answers and identifies where they agree, where they conflict, what each one added, and what the group missed. A final model uses that review to write the response.

Think of it like giving the same research task to three analysts, asking a fourth person to compare their notes, then having an editor write the final memo. The value comes from the gaps that appear when several answers sit next to each other.

OpenRouter tested the idea on 100 deep-research tasks from Perplexity’s DRACO benchmark. Its best panel, Fable 5 and GPT-5.5 with Claude Opus 4.8 synthesizing the results, scored 69.0%. Fable 5 alone scored 65.3%. GPT-5.5 alone scored 60.0%.

Its cheaper panel is even more relevant for startups. Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro reached 64.7%, close to Fable 5’s score at roughly half the cost. That creates room for products that reserve expensive frontier models for judging and final synthesis while letting cheaper models do the first pass.

A Pattern That Keeps Reappearing

Together AI’s Mixture-of-Agents came early. In 2024, Together showed a layered system where multiple open-source models generate responses, and later models build on earlier outputs.
Karpathy’s LLM Council made it a power-user workflow. His open-source project sends a query to several models, asks them to review and rank each other’s answers anonymously, then has a “Chairman” model compile the final response. It looks like the manual version of what products are now turning into features.
Perplexity brought the idea into consumer AI. Model Council runs multiple frontier models on the same query, compares their outputs, and uses a separate model to synthesize a higher-confidence answer. That is the same customer promise in a different wrapper: better answers through cross-model comparison.
CollectivIQ is taking the enterprise angle. It launched in March 2026 as an AI consensus platform, pulling from ChatGPT, Gemini, Claude, Grok, and other models to give teams more trusted answers. The buyer here is less interested in model fandom and more interested in reducing hallucinations, bias, and vendor dependence.
Even n8n has workflow templates for it. Operators can send one question to several models, have them evaluate the answers, and return a final verdict inside an automation.

The Trend

A new layer is starting to emerge above AI models.

The early products all use different names: Model Council, Mixture-of-Agents, consensus AI.

I prefer the name Answer Fusion. Across these products, the shared pattern is broader: multiple models contribute intermediate work, and a system synthesizes that work into one final answer.

That makes sense because model performance varies by task, cost, latency, context window, tool use, and reasoning style.

The first products in this category will likely win in places where people already double-check AI manually. They copy a prompt into Claude, GPT, Gemini, and Perplexity, compare the outputs, and try to work out what they missed. Answer Fusion just turns that manual comparison into a product.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Fintech on Fumes

Nicolás Cerdeira — Thu, 11 Jun 2026 15:01:59 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Chimoney, a startup building a global payout API, shut down — learn why below
LLMs are picking winners. Here’s how to become one
There’s a new Claude model: Fable 5 — learn why you should try it below

A big thanks to today’s sponsor, SparrowDesk. Let their AI agents handle 60%+ of your support tickets automatically.

What if 60% of your support tickets never reached you? ^AD

As a startup, most of what lands in your inbox is repetitive. The same questions, the same responses, resolved manually by someone who should be closing deals or building your product.

SparrowDesk fixes this. Our AI agents auto-resolve 60%+ of queries before it reaches your team. One unified inbox for chat, email, and more, so nothing goes missed, and every customer gets a reply.

Early-stage teams get 90% off the first year through our startup program.

→ Try SparrowDesk for free

This Week In Startups

🔗 Resources

How AI changes the software value chain

Why you should focus on the Return on Tokens (ROT)

LLMs are picking winners. Here’s how to become one

Your startup didn't work out. Now what? *

📰 News

Decart’s new world model can simulate hours of photorealistic driving

Google just fired a warning shot in the AI subscription price wars

OpenAI files confidentially for IPO, following Anthropic

Apple’s long-awaited AI Siri overhaul is finally here

💸 Fundraising

Arpio, an AI-native cloud recovery and resilience platform, raises $15M Series A

Forage, a financial infrastructure platform for government benefits payments, raises $40M Series B

Terra AI, an AI platform for mineral and energy exploration, raises $20M Series A

ZeroDrift, an AI compliance firewall for enterprise communications, raises $10M seed round

* sponsored

Fail(St)ory

Money Out, Not Money In

Chimoney tried to make messy global payouts feel simple.

It wanted to build a payout layer that could sit between companies, recipients, banks, wallets, local rails, and regulators across multiple markets.

It raised less than $1M to do it, which is roughly “nice seed round for a calendar app” money, not “move money across continents without breaking anything” money.

What Was Chimoney:

Chimoney was a payout orchestration layer for businesses sending money into fragmented markets.

It lived after the moment a company already owed someone money:

A marketplace needed to pay sellers.
A remote-work company needed to pay contractors.
A creator or affiliate platform needed to pay users across countries where the preferred payout method changed by market.

That was the job Chimoney tried to handle. Integrate once, fund the payout, and let recipients receive money through the method that actually worked where they lived.

The product supported 41 currencies across North America, Africa, and Latin America. Customers could send bank transfers, mobile money, airtime, gift cards, and even stablecoins.

It was not trying to be a new Stripe. It did not help businesses accept card payments, run checkout, manage subscriptions, or process merchant payments. It focused on the other side of the flow: getting money out to recipients who might not have the same banking access, cash-out options, or local rails.

The messy part sat underneath that simple integration. If one recipient wanted mobile money, another needed a bank transfer, and another could only use airtime or a gift card, Chimoney had to make those options work behind the scenes.

That meant routing payouts, managing local partners, handling failed transactions, reconciling balances, processing refunds, answering support tickets, and keeping enough liquidity in the right places. The customer saw one API. Chimoney carried the market-by-market mess underneath it.

Near the end, Chimoney tried to reposition part of this infrastructure around AI wallets and programmable payments. It was a cleaner story for the future, but the company still needed the payout business to work in the present.

By the time Chimoney stopped accepting new transactions in April 2026, the shape of the failure was clear: the product solved a real last-mile payout problem, but the company had to carry too much infrastructure before enough volume arrived.

The Numbers:

💸 Funding: Under $1M raised over the company’s life
🌍 Coverage: 41 currencies supported
🏢 Customers: Hundreds of businesses served
🛑 Shutdown: Stopped accepting new transactions on April 30, 2026

Reasons for Failure:

Distribution never became a real engine. The founder’s own explanation was direct: Chimoney had a technically solid, licensed platform, but no customer acquisition engine. In payout infrastructure, buyers need proof that money lands, failures get handled, compliance is covered, and switching risk is worth taking.
The last mile was expensive to operate. Chimoney simplified payouts for customers by absorbing the operational mess itself. Every local rail, mobile money option, gift card route, off-ramp, refund path, and partner relationship added work behind the API.
Fintech costs arrived before fintech scale. Under $1M is thin for a company moving money across multiple jurisdictions. Licenses, audits, banking partners, compliance, support, and liquidity management all cost money before transaction volume is large enough to carry them.
Volume density never caught up. Payout infrastructure gets healthier when repeat transactions flow through the same rails often enough. Without strong distribution, Chimoney still had to maintain the platform, partners, support, refunds, and compliance systems on too small a revenue base.

Why It Matters:

Payout startups need dense, repeat transaction volume before the operational load becomes manageable.
Developer experience helps, but payout buyers mostly care about whether money arrives and who fixes it when it does not.
Future-market repositioning is hard when the current infrastructure business is already running out of room.

Trend

Claude Fable 5

There’s a new Claude model in town, so naturally, everyone is talking about it.

Anthropic says it’s SOTA, many developers claim it’s the new coding model to use, and a lot of people seem angry that it won’t help them make a bomb.

Here’s everything you should know about it: the good, the bad, and the biohazard disclaimers.

Why it Matters

The best model is no longer always the model you get. Fable 5 is public. Mythos 5 is for vetted partners. Same family, different permissions, which means AI access is becoming a status layer.
The coding hype seems earned. Fable 5 looks built for the annoying work developers actually hate: messy repos, migrations, broken tests, weird dependencies, and tasks that take more than one clean prompt.
The guardrails are part of the product now. Sometimes Fable answers. Sometimes it refuses. Sometimes it quietly sends you to Opus 4.8.

The Coding Hype Has Substance

The strongest case for Fable 5 is coding.

Anthropic says it is better at long-running agentic work, large migrations, complex implementations, test writing, design fidelity, and reasoning across big codebases.

Early developer reactions mostly point in the same direction: it stays useful for longer, handles more context, and needs less babysitting once the task gets messy.

The specs help explain why. It reportedly has a 1 million token context window and 128,000 max output tokens. That gives it more room to read, plan, edit, test, and produce larger changes in one run.

Anthropic’s biggest example is Stripe using it to cut a 50-million-line Ruby migration from months to a day. My favorite example so far has to be this one:

The problem: it is very very expensive. Fable 5 costs twice as much as Opus 4.8, so letting it loose on a repo has the energy of giving a very smart intern your credit card.

The Guardrails Are Everywhere

The weirdest part of Fable 5 is that it is not just a model. It is a model with a bouncer.

Anthropic says Fable 5 uses the same underlying model as Mythos 5, but with stronger safeguards for public use. Mythos 5 is the more restricted version, available only to vetted partners, because it can help in areas that could be dangerous (mostly cybersecurity and biology).

So Fable 5 sometimes does something strange. If a cyber or bio prompt gets flagged, it may route the request to Opus 4.8 instead.

The biology filters seem especially jumpy. Reports say Fable 5 refused or rerouted questions about mitochondria, mRNA vaccines, hay fever, asthma medication and antibiotic resistance. It's like if a high school worksheet got stopped at airport security.

Mythos Already Changed the Security Workload

Anthropic’s caution makes more sense when you look at what Mythos has already done.

In Project Glasswing, around 50 partners used Mythos Preview for defensive cybersecurity. Anthropic says they found more than 10,000 critical severity vulnerabilities.

Then Mythos scanned more than 1,000 open-source projects, initially estimated 23,019 vulnerabilities, and produced thousands of critical severity findings. Anthropic says independently reviewed findings had a 90.6% true-positive rate.

That’s the fear: the same model that can find vulnerabilities for defenders can also help attackers find them first. So Anthropic is keeping the public version on a much shorter leash.

That is the tradeoff with Fable 5. The model looks genuinely useful, especially for coding, but the launch is also a reminder that the best AI tools may increasingly come with routing, permissions, and a bouncer at the door.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

The Pizza Robot Graveyard

Nicolás Cerdeira — Thu, 04 Jun 2026 17:33:41 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Picnic, a startup building robots that made pizza, has shut down — learn why below
24 tips for giving S-tier demos
Microsoft Build 2026 was this week — all important announcements below

A huge thanks to today’s sponsor, HubSpot. Download their free AI guide to grow your business today.

Demystify AI for Your Business in Just 4 Steps ^AD

Overwhelmed by AI? HubSpot's free guide cuts through the noise. Get the ultimate crash course for non-technical entrepreneurs who want to harness AI's power—without getting lost in the jargon.

"AI for Business Builders" delivers:

A 4-part roadmap to AI mastery
Jargon-free explanations of large language models
Practical prompt engineering tips you can use today
Real-world examples of AI boosting businesses like yours

Arm yourself with the knowledge to make informed AI investments and skyrocket your startup's growth.

Download Your Free Guide Now

This Week In Startups

🔗 Resources

Startups become what they hunt

24 tips for giving S-tier demos

Small Businesses are the Next Frontier for AI

Your startup didn't work out. Now what? *

📰 News

OpenAI launches new Codex tools for white-collar work

Anthropic files to go public

Claude Opus 4.8 is out and is SOTA

Meta rolls out a new AI creator assistant on Facebook

💸 Fundraising

OpenRouter, an AI model exchange and inference routing platform, raises $113M Series B funding

Reactor, a real-time generative video platform for developers, raises $59M Series A funding

Catena Labs, an AI-native banking infrastructure startup, raises $30M Series A funding

Saris, an AI workflow automation platform for banks and credit unions, raises $28.8M Series A funding

* sponsored

Fail(St)ory

Robot Aquarium

Picnic built a robot that could sauce, cheese, and toppings at assembly-line speed.

It could sit inside a commercial kitchen and help one worker turn out 100+ pizzas an hour. That made it easy to like and easy to demo.

The problem was turning a pizza-making robot into something enough operators needed every month.

What Was Picnic:

Picnic built the Picnic Pizza Station, a robotic pizza assembly system for commercial kitchens. It applied sauce, cheese, pepperoni, and other toppings onto dough, then handed the pizza off to the rest of the kitchen workflow.

Pizza was a smart first target. The work is repetitive, the ingredients are standardized, and rush-hour demand can be painful to staff.

The company claimed one employee could produce roughly 100+ customized 12-inch pizzas per hour with the system, depending on setup and size.

That was the pitch: fewer hands on the line, more predictable output, less waste from inconsistent topping.

Picnic sold the machine through a leasing model. When it reached the market in 2021, pricing started around $3,500 per month.

That made the product feel closer to an operating expense than a large capital purchase, but the customer still had to believe the machine would earn its place every month.

The strongest use cases were dense-volume environments: stadiums, campuses, military sites, large venues, and high-throughput foodservice operators. Those are the kinds of places where pizza behaves more like production.

A normal restaurant is messier: some hours are slow, menu changes matter, and kitchen space is tight. Staff still need to load ingredients, monitor the machine, clean it, recover from jams, and trust it during the exact moments when failure is most expensive.

Picnic had real customer interest. It partnered with Ethan Stowell Restaurants, tested with Domino’s Pizza Enterprises in Berlin, listed customers like Aramark, Chartwells, and Compass Group, and pushed into venues and institutional foodservice.

It also carried the cost structure of a hardware company. Machines had to be manufactured, shipped, installed, maintained, monitored, and supported. In 2024, Picnic raised $5M to scale production and meet demand across North America.

By May 2026, the company had become insolvent. Its assets and IP were sold to an undisclosed buyer through liquidation.

One customer described being left with a $250,000 “robot aquarium,” which says a lot about the risk operators take when their automation vendor disappears.

The Numbers:

🍕 Founded: 2016 in Seattle
⚙️ Throughput claim: roughly 100+ pizzas per hour
💸 Lease pricing: started around $3,500/month in 2021
📈 Funding: roughly $50M-$53M raised
👥 Peak team: about 100 employees by 2023
🏟️ Customers and pilots: MOTO Pizza, Domino’s Pizza Enterprises, Aramark, Chartwells, Compass Group, military and venue installs

Reasons for Failure:

The ROI depended on dense, predictable pizza volume: Picnic made the most sense when pizza production was concentrated: stadium rushes, campuses, bases, large venues, and very busy shops. In lower-volume restaurants, a $3,500/month lease had to compete with slow hours, uneven staffing needs, and limited kitchen space.
The robot automated only one visible part of the job: Topping is important, but it is only one step in a pizza operation. People still had to prep ingredients, manage dough, bake, cut, box, serve, clean, and fix problems when the workflow broke. For operators, the buying decision was really about total kitchen reliability, not demo throughput. A robot that needs monitoring and support can still save labor, but the savings need to survive real shifts.
Hardware burn met a colder funding market. Former CEO Clayton Wood said Picnic was “caught in the squeeze” between the free-money period and the harsher post-2022 market. That matters for a company selling physical systems. Picnic needed capital for manufacturing, deployment, maintenance, and sales cycles that move slower than software.
Food robotics had become a scarred category. Picnic was selling into a market where Zume, Basil Street, Piestro, and other restaurant automation startups had already failed. That gave buyers a fair reason to ask what would happen if the vendor disappeared and the robot became expensive furniture.

Why It Matters:

Sell the customer the full outcome, because automating the most visible task may still leave them with the annoying parts.
Your best use case is only a market if there are enough buyers with the same pain, budget, and workflow.
A buyer’s fear of being stuck with dead equipment can matter as much as your ROI math.

Trend

Microsoft Build 2026

Microsoft used Build 2026 to show how much of the enterprise agent market it can absorb by default.

It already controls the work context, identity layer, data platform, dev workflow, runtime, and endpoint. Now it is wiring those pieces together so agents can use company files, calendars, permissions, business data, and approval rules without customers stitching the stack themselves.

For founders, generic workplace agents just got harder to sell.

Why it Matters

Enterprise context becomes distribution. Microsoft IQ gives agents company context from files, people, workflows, data, and the web. Startups have to rebuild that through integrations; Microsoft already sits where the work happens.
Generic coordination agents get squeezed. Scout handles meetings, prep, stalled work, calendars, and deliverables across Microsoft 365. Startups need a narrower workflow, unique data, or a buyer Microsoft will not serve well.
Microsoft wants model control. Copilot leaned heavily on OpenAI. The new MAI models give Microsoft more control over cost, latency, routing, and product behavior.

Microsoft IQ makes company context the product

Microsoft IQ is Microsoft’s name for the context layer underneath its agents.

Think of it as the system that tells an agent what it needs to know before it acts: who works on what, which files matter, how the business defines key terms, what happened in meetings, and what outside information is current.

It has four pieces:

Work IQ: context from Microsoft 365: files, chats, meetings, people, calendars, and workflows.
Foundry IQ: internal company knowledge for agents built in Microsoft Foundry.
Fabric IQ: business meaning for company data, like how customers, products, invoices, and metrics connect.
Web IQ: fresh external information from web pages, news, images, and video.

The idea is that companies define this context once, then reuse it across Copilot, Foundry, Copilot Studio, and Microsoft 365. That makes agents less dependent on long prompts and one-off integrations.

Scout is Microsoft’s always-on work agent

Scout is Microsoft’s first Autopilot: an agent that can run in the background instead of waiting for a prompt every time.

It connects to Microsoft 365 tools like Teams, Outlook, OneDrive, SharePoint, calendar, email, contacts, and browser context. The job is coordination work: scheduling meetings, preparing users, flagging stalled decisions, blocking focus time, and noticing upcoming deliverables.

Under the hood, Scout is built on OpenClaw, Microsoft’s open-source agent runtime for autonomous workflows.

The important part is control: Scout gets its own governed Entra identity, so its actions are visible and attributable. Companies can limit what it can access, require approval for sensitive actions, and apply existing Purview policies when it handles data.

GitHub Copilot is getting a new app

The new GitHub Copilot app is a desktop app for running coding agents as separate work sessions.

Instead of asking Copilot for code in the editor, a developer can spin up agent sessions from GitHub Issues, Pull Requests, or previous sessions. Each session gets its own branch, files, conversation history, and task state through git worktrees.

The practical change is parallel work. One agent can investigate a bug, another can draft a feature, and another can respond to review comments, without mixing files or branches.

The developer still decides what gets merged. Copilot is taking on more of the execution loop, while the human keeps the review, testing, and approval role.

Microsoft is reducing its OpenAI dependency

Microsoft also announced seven in-house MAI models for reasoning, coding, image generation, transcription, and voice.

The main one is MAI-Thinking-1, a 35B-active-parameter reasoning model for multi-step tasks, long-context work, and code generation. It is still limited to early partners.

This is important because Copilot and Microsoft’s AI products have relied heavily on OpenAI models. The MAI models give Microsoft more control over cost, latency, routing, and product packaging. Simple tasks can run on cheaper Microsoft models, while harder tasks can still use frontier models when needed.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Turned Into A Feature

Nicolás Cerdeira — Thu, 28 May 2026 18:54:45 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Huxe, an AI audio app that turned your inbox into a personal daily briefing, has shut down — learn why below
How to Avoid AI Code Slop
Tokenmaxxing is the new obsession with measuring AI productivity — learn why it matters below

A huge thanks to today’s sponsor, SimpleClosure. When startups hit the wall, close cleanly and move forward with expert-guided support.

When It’s Time to Close, Do It Properly ^AD

Across the startup landscape, founders are facing tougher markets, crowded categories, and harder paths to sustainable growth.

When shutdown becomes the right next step, the work is not over.

SimpleClosure helps founders bring structure to the process, from state filings and investor communications to final distributions, compliance, and selling remaining assets. So you can close cleanly, responsibly, and with a clearer path through.

👉 Simplify Your Shutdown

This Week In Startups

🔗 Resources

How to Avoid AI Code Slop

In B2B customer support, AI is a Copilot, Not a Replacement

The Founder's Black Box

Anthropic and OpenAI have finally found PMF

📰 News

Meta launches Instagram, Facebook, and WhatsApp subscriptions

YouTube adds new podcast features, including an AI recommendation tool

Pope Leo uses first major papal text to warn about dangers of AI

Robinhood now lets your AI agents trade stocks

💸 Fundraising

AI coding startup Cognition raises $1B

Mercury, a fintech banking platform for startups, raises $200M Series D funding

Moment, an AI operating system for investment management, raises $78M Series C funding

Modal, a serverless cloud platform for AI infrastructure, raises $355M Series C funding

Fail(St)ory

Personalized Podcasts

Huxe was an AI audio app that turned your inbox, calendar, and interests into a personal daily briefing.

You opened it in the morning, pressed play, and got something closer to a podcast about your day.

It came from the same world as NotebookLM’s viral Audio Overviews. Ex-Googlers built it. Good investors backed it.

Then, on May 21, 2026, Huxe announced it was winding down, in the same month Spotify and Amazon pushed deeper into AI-generated personal audio.

What Was Huxe:

Huxe was a consumer AI startup in the Bay Area built around one bet: people do not want to read everything.

They want the important stuff spoken to them while they drive, walk, cook, work out, or avoid another screen.

The team came from Google’s NotebookLM project, where they had already seen AI-generated audio take off.

At first, Huxe looked more like a work tool. The early idea was to connect systems like Salesforce and let teams chat with company data.

Then the team shifted toward consumers. The bigger opportunity seemed to be audio.

The core product was a personalized daily briefing. Users connected email, calendar, and interests, then got a morning audio summary covering meetings, important emails, news, and topics they cared about.

The app also had personalized audio feeds, AI-generated DeepCasts built from prompts, live AI radio-style stations around topics like tech or sports, and interactive hosts users could interrupt to ask questions or change direction mid-conversation.

It was pretty ambitious. Huxe touched productivity, news, learning, search, and voice assistants all at once.

However, at the same time, Google, Spotify, and Amazon were pushing into the same category with products tied directly into platforms people already used every day.

Huxe raised $4.6 million in September 2025. Eight months later, it shut down.

The Numbers:

🗓️ Founded: late 2024, after the team left Google
💰 Funding: $4.6M
📱 Public launch: June 5, 2025
🛑 Shutdown announced: May 21, 2026
📉 Public launch to shutdown: about 11.5 months

Reasons for Failure:

The product was turning into a platform feature: A personalized AI briefing fits naturally inside Google, Spotify, Amazon, Apple, or Microsoft. Google had NotebookLM and Daily Listen. Spotify announced Personal Podcasts the same week Huxe wound down. Amazon had Alexa Podcasts. Huxe had to convince users to download a new app, while the platforms could place similar features inside products people already used every day.
The trust barrier was high: Huxe worked best when users connected email, calendar, and personal context. That is sensitive territory for a young startup. Many users hesitate before giving inbox access to a product they discovered a few weeks ago. Big platforms already had the accounts, the data, and years of user familiarity behind them
The product needed near-perfect reliability: The whole pitch was saving users time. That breaks fast if people still feel the need to double-check their inbox afterward. Some users liked the summaries but still worried important emails could get missed. Once that doubt appears, the product shifts from essential tool to optional layer.
Audio has very little margin for error: Users tolerate mediocre text more than mediocre audio. You can skim weak writing in seconds. Audio forces you to sit through every awkward pause, robotic sentence, and bad transition. Small quality issues feel much bigger in a product people are supposed to use every morning.

Why It Matters:

In consumer AI, the biggest risk is often becoming a feature inside a larger platform.
Audio products have a brutal quality bar because users feel every second of bad output.

Trend

Burning Tokens

This week, Uber’s COO said AI token spending is getting harder to justify. A few days earlier, OpenAI offered YC startups up to $2 million in API credits and encouraged founders to go “tokenmaxxing.”

That’s where the market is right now. One side is subsidizing massive AI consumption. The other is asking whether all this compute actually produces better products.

The interesting thing is how quickly tokenmaxxing escaped startup Twitter and became corporate behavior. Internal leaderboards. Usage quotas. Employees running agents for no reason just to look AI-active.

A productivity tool quietly turned into a status game.

Why it Matters

AI spend is becoming a hiring decision. Founders increasingly compare token costs against salaries, not software budgets. Uber even said it’s hiring fewer people while trying to balance AI spend against headcount.
AI adoption and AI productivity aren’t the same thing. Companies like Meta and Amazon reportedly tracked token usage aggressively. Others, like Indeed, focused on outcomes instead.
Tokens are becoming a real cost center. AI usage scales unpredictably through agents and automation loops. Companies are starting to treat tokens more like cloud infrastructure than SaaS seats.

Tokenmaxxing

Tokenmaxxing is what happens when companies start treating AI usage as proof of productivity.

Executives want employees using AI more aggressively. But AI adoption is hard to measure. Most workers say they use ChatGPT or Claude occasionally. Few actually change how they work around it.

So companies started looking for a measurable signal.

Tokens became the obvious choice. Every prompt, response, workflow, or agent loop consumes them. More tokens usually means more AI usage. That made token consumption an easy internal metric.

Then the dashboards appeared.

Meta built an internal leaderboard called “Claudeonomics” tracking AI usage across roughly 85,000 employees. Microsoft created internal token leaderboards as well. Amazon managers monitored AI adoption rates and usage dashboards internally.

YC encouraged founders to think about “tokenmaxxing over headcountmaxxing.” OpenAI reinforced the trend by offering YC startups up to $2 million in API credits. Small startups started viewing tokens as a substitute for hiring.

Nectir reportedly pushed engineers to spend hundreds or even thousands of dollars per month on AI tools because leadership believed aggressive usage would force employees to discover better workflows. One founder described AI agents as becoming “an army of coders.”

Why it Broke

The problem with tokenmaxxing is simple: activity is easy to measure. Value isn’t.

Once token usage became visible inside companies, employees started optimizing for the metric itself.

According to reports, some Meta employees left AI agents running for hours just to inflate usage numbers. One user consumed 281 billion tokens in a month.

Amazon employees used internal AI tools for unnecessary tasks because they believed managers were paying attention to adoption dashboards

Then the financial reality showed up.

Uber burned through its annual Claude Code budget just four months into the year. Leadership started asking the question: does more token consumption actually create proportional business value?

Some people still believe tokenmaxxing is a useful signal, at least before employees start gaming it. Their argument: strong engineers using AI seriously naturally burn more tokens because they’re testing more ideas, running more workflows, and shipping faster.

Others think the metric corrupts itself the moment companies start tracking it publicly. Employees stop optimizing for output and start optimizing for visibility instead.

The real lesson is probably that token usage works as an early adoption push, but breaks once consumption becomes the goal rather than the byproduct of useful work.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

23M Users. Only 75 Paid.

Nicolás Cerdeira — Thu, 21 May 2026 18:36:49 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

StreamElements, a startup building creator tools, is about to shut down — learn why below
The Dark Marketplace: why Vertical AI is coming for commerce
Google’s I/O 2026 event just finished — learn every important announcement below

A huge thanks to today’s sponsor, SimpleClosure. When startups like StreamElements face an uncertain future, close cleanly and move forward with structured, expert-guided support.

Closing Well in Uncertain Times ^AD

StreamElements’ potential shutdown is a familiar startup story: a company can build for years, raise significant funding, support millions of users, and still reach a point where the next step is uncertain.

Across the creator economy and AI tooling space, founders are running into similar pressure: crowded markets, changing customer behavior, and harder paths to sustainable growth.

When shutdown becomes the right next step, the work is not over.

SimpleClosure helps founders bring structure to the process, from state filings and investor communications to final distributions, compliance, and asset sales. So you can close cleanly, responsibly, and with a clearer path through.

👉 Learn more

This Week In Startups

🔗 Resources

YC’s Request for Startups - Summer 2026

The Dark Marketplace: why Vertical AI is coming for commerce

After Automation: AI progress creates more work for humans, not less

Stuck between US LLC and EU? Honest comparison of 12 countries, real tax data *

📰 News

OpenAI co-founder Andrej Karpathy joins Anthropic’s pre-training team

Elon Musk has lost his lawsuit against Sam Altman and OpenAI

OpenAI launches ChatGPT for personal finance, will let you connect bank accounts

Figma adds an AI assistant to its collaborative canvas

💸 Fundraising

Hark raises $700M Series A at a $6B valuation

Mercury raises $200M Series D funding

GridCARE raises $64M to squeeze more power out of grids

Commure, an AI platform for Healtchare, raises $70M in funding

* sponsored

Fail(St)ory

The Cost of Free

StreamElements was a creator-tools company built for livestreamers.

It had millions of creators, deep Twitch and YouTube roots, $111 million in funding, and a product people actually used.

Then, in May 2026, the company entered shutdown mode while trying to find a buyer, with creators reportedly told to save their assets during a 30-day window.

What Was StreamElements:

Launched in 2017 the company started with a simple pitch: give streamers the tools they need to run a professional live channel, for free.

That meant overlays, alerts, chatbots, tipping pages, loyalty tools, merch stores, analytics, moderation, and sponsorship infrastructure. The plumbing behind the show.

A streamer does not just press “go live.” They need alerts when someone subscribes, commands for chat, widgets for donations, sponsor graphics, and a setup that does not collapse the second the audience shows up.

StreamElements became one of the default options in that stack.

By 2019, more than 200,000 Twitch and YouTube channels were using it. Then the pandemic hit and livestreaming exploded. By 2021 more than 1.1 million creators were using the tool.

That same year, SoftBank led a $100 million round, bringing total funding to roughly $111 million.

So, they had the users, but how were they making money?

StreamElements kept the tools free, then made money through brand sponsorships. Brands paid to run campaigns with creators, creators got paid for the work, and StreamElements took a cut in the middle.

In theory, it was a good idea. Creators loved free tools. Brands wanted creator distribution. StreamElements sat in between.

However, free software still has bills.

By 2026, StreamElements said it was spending $300,000 per month just to keep its tools running. That did not include the heavier stuff around payroll, sales, support, and product.

So in January 2026, it launched Keep It Live, a crowdfunding-style campaign asking creators and viewers to help keep StreamElements free.

The campaign accidentally told the real story.

StreamElements said 23 million-plus creators had used its tools over the past decade, and 2 million-plus had gone live that month. The campaign had raised $1,700 from 75 supporters in the research snapshot.

Millions used it. Almost nobody paid for it.

By May 2026, GamesBeat reported that StreamElements was preparing to close its website and sunset its creator tools. Creators were reportedly told they had 30 days to log in and save their assets.

The company later said it was in “positive discussions with potential acquirers.”

The Numbers:

💰 Funding: $111M
👥 Users: 23M+ creators served.
🧾 Costs: $300K monthly tool costs
📉 3 major layoff rounds between June 2022 and August 2023.
🧑‍💻 200+ employees at peak, down to about 72 listed on LinkedIn by May 2026.

Reasons for Failure:

The free-tools bargain broke. StreamElements built enormous reach, but most of that reach sat outside the paywall because there was no real paywall. That only works when sponsorship revenue is growing fast enough to subsidize the product.
Sponsorships were too cyclical for the cost base. The company depended heavily on brand campaigns, and those budgets were the first thing to wobble when the ad market slowed. StreamElements said as much during its 2023 layoffs, pointing to weaker advertising demand and reduced budgets from existing clients. That is a bad place to be when your product costs are fixed and your revenue depends on marketers staying brave.
The 2021 round sized the company for a boom that did not last. StreamElements raised $100 million when livestreaming, gaming, and creator monetization all looked like they would keep climbing forever. Then growth cooled, ad spend tightened, and the company had to cut staff three times in 14 months
The creator product and the buyer product were not the same thing. Creators came for alerts, overlays, bots, and tipping tools. Brands came for measurable sponsorship performance. Those are two different products, two different sales motions, and two different definitions of success. StreamElements had creator adoption, but the business depended on advertiser conversion.

Why It Matters:

Creator love does not matter if the buyer is the advertiser.
Boom-era funding can lock a startup into boom-era assumptions.

Trend

Google I/O 2026

Google just wrapped its I/O 2026 event with a ton of AI announcements.

Most of them point in the same direction: Google wants AI embedded across everything it already owns. Search, Gmail, Chrome, YouTube, Shopping, Android.

Gemini 3.5 Flash became the foundation for most of the launches. Spark handles long-running tasks in the background. Omni pushes AI deeper into editing workflows. Search gets more agentic. Shopping gets automated.

Here’s what matters and where this is going.

Why it Matters

Google launched the operational model first: Gemini 3.5 Flash shipped before Pro because agents care more about reliability and speed across repeated actions than perfect reasoning on one big query.
Google is quietly turning Search into a monitoring product: The shift is from “search this” to “keep watching this for me.” That changes how people discover products, content, and information online.
Omni is a bigger threat to creative software than Veo was: Generating clips is one thing. Editing existing assets through conversation is much closer to actual creative work.

Gemini 3.5 Flash

The biggest launch from I/O was Gemini 3.5 Flash.

Google says it outperforms Gemini 3.1 Pro across several coding and agent benchmarks while running much faster. More importantly, Google immediately made it the default model behind AI Mode in Search.

Google clearly optimized Flash for constant execution across tools and workflows. Search queries, coding agents, background tasks, enterprise agents, shopping flows.

These systems constantly make API calls, retrieve information, compare outputs, and retry actions. Speed starts mattering a lot once that stack grows.

Spark

Spark feels much closer to OpenClaw-style agents than to a traditional assistant. Google is basically building a persistent cloud worker connected to Gmail, Calendar, Search, and eventually payments.

The examples were practical: reviewing inboxes, tracking internship openings, turning emails into task lists, scheduling calendar blocks, preparing recurring workflows, the kind of thing you would expect OpenClaw to be able to handle.

Spark runs on dedicated virtual machines inside Google Cloud, so it keeps running even when your devices are offline.

Google also said Spark will support third-party tools through MCP integrations later on. The rollout starts with trusted testers and AI Ultra subscribers in the U.S.

Daily Brief

Daily Brief organizes your morning inside Gemini: inbox updates, schedule context, tasks, reminders, priorities.

It’s nothing groundbreaking technically, but the packaging is probably the important part. Google is wrapping agent behavior into an existing habit instead of asking users to build workflows manually. Open app in the morning, get context, move on.

That approach showed up across a lot of the I/O announcements. Less “build your own AI system,” more lightweight automation built directly into products people already use.

Omni

Google also introduced Gemini Omni, a multimodal model focused on video generation and editing.

The important part is that Omni works with existing assets, not just prompts. Users can combine text, images, audio, and video inputs, then edit outputs conversationally inside tools like Flow and YouTube Shorts.

Search and Commerce

Google also upgraded AI Mode in Search with Gemini 3.5 Flash and introduced information agents that continuously monitor topics for users.

The shift here is pretty straightforward: instead of repeatedly searching for the same things manually, users hand monitoring tasks to the system.

Google connected this directly into shopping too.

Universal Cart and the new Agent Payments Protocol let users define spending limits, product constraints, and purchase conditions while agents handle transactions automatically once those conditions are met.

Overall, most of the I/O announcements pointed in the same direction: AI systems that stay connected to ongoing workflows instead of isolated prompts.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Sudden Bankruptcy

Nicolás Cerdeira — Thu, 14 May 2026 19:28:24 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

Parker, a startup making corporate cards for ecommerce, shut down abruptly — learn why below
Is Software Losing Its Head?
OpenAI’s new Voice Models and Thinking Machines new “Interaction Models” — learn everything about them below

A huge thanks to today’s sponsor, Warmy. Fix your email deliverability so your outreach reaches inboxes and generates leads.

Sending Emails but Getting Zero Replies? ^AD

It might not be your copy.

Many startups unknowingly have broken email deliverability, causing messages to land straight in spam folders.

That means lost leads and wasted outreach. Warmy.io automatically warms up your email accounts, protects your sender reputation, and keeps your emails landing in real inboxes.

If email is part of your growth strategy, this matters.

Fix your deliverability before your next campaign

This Week In Startups

🔗 Resources

How to Build Services-as-Software Business

The state of B2B monetization in 2026

Is Software Losing Its Head?

Your startup didn't work out. Now what? *

📰 News

Notion just turned its workspace into a hub for AI agents

Amazon launches an AI shopping assistant for the search bar

Anthropic now has more business customers than OpenAI

Google unveils Googlebook, a new line of AI-native laptops

💸 Fundraising

Vapi raises $50M Series B powering the next generation of enterprise voice AI

AI-powered wholesale broker Novella raises $21M

Rivian spinoff Mind Robotics raises another $400M

Exaforce Raises $125M Series B to Combat AI-Powered Attacks

* sponsored

Fail(St)ory

Parker

Parker built corporate cards and financing tools for ecommerce brands that needed cash before revenue showed up.

Then, almost overnight, it disappeared.

Three days after shutting down, the company filed for Chapter 7 bankruptcy. Reports suggest Parker had been trying to sell itself shortly before the collapse, and whatever was supposed to happen behind the scenes clearly did not happen.

What Was Parker:

Parker launched in 2019 and went through YC’s Winter 2019 batch.

Its focus was entirely on ecommerce brands, especially businesses doing a few million to tens of millions in annual sales.

Most startup card companies are built around SaaS startups. Those companies have predictable margins, recurring revenue, and relatively clean cash flow.

Ecommerce businesses are different. They constantly spend money before revenue shows up. Inventory gets paid upfront. Ads get paid upfront. Freight, packaging, suppliers, warehouses, agencies, manufacturers, all upfront. The money comes back later, if everything goes well.

Parker built its product around that gap.

The company’s pitch was that traditional banks and corporate cards underwrite ecommerce brands badly because they look at outdated financials and generic risk models. Parker wanted to underwrite using live business data instead.

Customers connected systems like Shopify, Amazon, and QuickBooks. Parker used that data to decide credit limits and repayment terms.

That let the company make a much more aggressive offer than normal business cards. Parker claimed it could offer limits 10x to 20x larger than traditional cards, with some businesses getting access to up to $10 million in credit.

The more interesting part was the repayment structure. Normal business cards use monthly billing cycles. Parker thought that made no sense for ecommerce operators managing inventory cycles and ad spend.

So instead, every transaction got its own repayment window. A purchase made on March 1 could be repaid on May 1. A purchase made two days later would have its own separate due date.

If you are buying inventory before Q4 or ramping ad spend before a launch, timing matters more than rewards points. A normal statement cycle can create random cash pressure depending on when purchases land. Parker tried to make repayment timing predictable.

Parker had real traction: they claim more than $1 billion in payments volume across hundreds of brands and over 65 million in revenue.

But then everything collapsed very quickly. Parker abruptly stopped operating on May 4. The company filed for Chapter 7 bankruptcy three days later.

The Numbers:

🗓️ Founded: 2019
🚀 YC Batch: Winter 2019
💰 Funding: $200M+
📈 Claimed payments volume: $1B+
💳 Claimed card limits: up to 10x-20x higher than traditional cards
📉 Bankruptcy filing: Chapter 7 liquidation

Reasons for Failure:

The business was riskier than it looked: Parker looked like software, but it was really an ecommerce credit business built around larger limits and longer repayment windows tied to volatile cash cycles.
Ecommerce is a brutal market to finance: Parker picked ecommerce because the pain was real. The problem is that ecommerce volatility cuts both ways. Ad costs spike, inventory gets stuck, suppliers miss deadlines, consumer demand swings, marketplaces delay payouts. A lender sitting inside those cash cycles inherits that chaos very quickly.
The funding headlines probably overstated how safe the company was: Parker often talked about raising more than $200 million. But a large part of that was asset-backed lending capacity supporting the card program. That money helps finance customer spending. It is not the same thing as having hundreds of millions sitting in the bank to absorb losses or survive operational problems.
The company seems to have run out of options very suddenly: The bankruptcy filing reportedly mentioned acquisition talks, mergers, and out-of-court wind-down options before the Chapter 7 filing. Reports also suggest a potential acquisition collapsed shortly before shutdown. If that is accurate, Parker may have been depending on a deal to survive and simply ran out of time when it disappeared.

Why It Matters:

Parker proved that ecommerce brands still need financial products built around cash flow, not generic startup banking.
Huge funding totals and payments volume can hide how fragile a fintech company actually is.

Trend

AI That Interrupts You

For years, AI voice demos have all had the same problem.

You ask something. The model waits politely. Then it answers in a slightly awkward voice while pretending interruptions don’t exist. If you correct yourself halfway through, the whole thing starts wobbling immediately.

Last week this started to change. Two major labs pointed at the same problem: voice AI still handles conversation like software, not like people. The bottleneck is the interaction layer.

OpenAI launched GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. A few days later, Thinking Machines introduced what it calls “interaction models,” a research preview built around continuous audio, video, and text interaction.

Why it Matters

The friction is moving from intelligence to interaction: The models are already smart enough for plenty of workflows. The annoying part now is the conversation itself. Delays, awkward pauses, broken interruptions, losing context after tiny corrections.
Voice is becoming a workflow layer: You can see it in OpenAI’s examples: scheduling, support, meetings, translation, real estate. The AI handles the software while you keep talking. Less clicking around dashboards.

OpenAI’s Realtime Push

OpenAI launched three new realtime audio models:

GPT-Realtime-2 for live voice conversations
GPT-Realtime-Translate for live translation
GPT-Realtime-Whisper for streaming transcription

GPT-Realtime-2 can:

reason while talking,
handle interruptions more cleanly,
call tools in parallel,
and keep long conversations alive with a 128K context window.

Developers can also tune reasoning effort depending on the task and add spoken status updates like “checking that now” while the model works in the background.

The translation and transcription models are also built for live interaction instead of post-processing. Translation happens while both people keep talking naturally. Whisper streams text continuously instead of waiting for complete speech segments.

Thinking Machines’ Interactive Models

Thinking Machines is aiming at a bigger shift.

Their argument is basically that current AI interfaces still behave like software pretending to be conversational. The model waits for you to finish talking. Then it responds. Then it waits again.

Real conversations are messier than that.

People interrupt each other constantly. They hesitate. They change direction halfway through sentences. Timing carries meaning. Silence carries meaning too.

Thinking Machines built its interaction model around those messy moments.

The system processes conversations in tiny 200 millisecond chunks called “micro-turns.” It keeps listening while speaking. It reacts while watching video. It can jump in naturally instead of waiting for a perfectly clean stopping point.

Some of the demos feel surprisingly human in small ways. The model notices hesitation. It reacts to visual changes on-screen. It interrupts when someone says something incorrect.

Most current voice systems fake this behavior with external turn-detection systems sitting outside the model. Thinking Machines is trying to build interactivity directly into the model itself.

You can already see the direction this pushes things toward. The useful AI products over the next few years probably won’t feel like chatbots with better answers. They’ll feel like systems that stay in sync with you while work is happening.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Too Early To AI

Nicolás Cerdeira — Thu, 07 May 2026 17:23:55 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

NeuroPixel.AI, a startup building AI tools for fashion, shut down — learn why below
The stuff nobody tells you about startup marketing
An overview of the Agent-Orchestration industry — learn about the key players below

A huge thanks to today’s sponsor, SimpleClosure. When startups like NeuroPixel hit the wall, close cleanly and move forward with structured, expert-guided support.

When the market shifts, we'll help you handle closure right ^AD

NeuroPixel.AI’s shutdown is a familiar startup story: tougher distribution, stronger competition from major tech players, lost revenue, and a market that moved faster than the company could.

Across the space, founders are running into similar pressure: shrinking differentiation, harder-to-sustain product-market fit, and scale becoming more difficult to reach.

When shutdown becomes the right next step, the work is not over.

SimpleClosure helps founders bring structure to the process, from state filings and investor communications to final distributions, compliance, and asset decisions. So you can close cleanly, responsibly, and with a clearer path through.

👉 Learn more

This Week In Startups

🔗 Resources

The best AI-native GTM plays you're not running

The "AI Job Apocalypse" is a complete fantasy

The stuff nobody tells you about startup marketing

Your team isn’t using AI. Here’s why that’s your fault.

📰 News

OpenAI launched GPT 5.5 Instant, the new default in ChatGPT

Image AI models now drive app growth, beating chatbot upgrades

Amazon opens up its global logistics network to all businesses

X announces a rebuilt ad platform powered by AI

💸 Fundraising

China’s Moonshot AI raises $2B at $20B valuation

Corgi raises $160M at $1.3B valuation to expand AI-native insurance platform

CodeWords has raised a $9m seed round to make you a workflow automation genius

UK stablecoin startup OpenTrade raises $17M

Fail(St)ory

Eaten by Giants

NeuroPixel.AI shut down in April after five years building AI tools for fashion ecommerce. The company raised a lot of money, worked with brands like Myntra and Decathlon, and got into generative AI before most people even knew what “GenAI startup” meant.

Then the market shifted underneath them.

What’s interesting is that they were not wrong about the trend, they just saw it early. They built real tech for a real problem. And then larger image models got good enough, fast enough, to crush the advantage they spent years building.

What Was NeuroPixel:

NeuroPixel started in Bengaluru in 2020 with one problem in mind: Fashion catalogs are operational nightmares.

Every product needs photos. Then more photos for different marketplaces. Different body types. Different campaigns. Different countries. Different seasons. A simple shirt turns into hundreds of image assets before it ever reaches a customer.

NeuroPixel’s pitch was simple enough that buyers immediately understood it. Instead of doing endless photoshoots, brands could shoot apparel on mannequins and let AI render the clothes onto synthetic models.

This sounds very normal now. Any AI model could do it. But in 2020 things were very different: generative AI was very unreliable, image models struggled with hands, consistency, fabric textures, proportions.

NeuroPixel was one of the companies trying to solve those ugly edge cases for a specific vertical instead of building generic AI tools for everyone.

The company claimed its system could reduce catalog production costs by 30% and cut processing time by 90%. Later, it pushed further into synthetic humans, virtual try-ons, AI-generated marketing images, and model customization.

In 2021, the startup raised an $825K seed round. A year later, Flipkart Ventures put in another roughly $299K. That second round valued the company at around $5.8M.

The company reportedly worked with brands including Myntra, Fabindia, Van Heusen, and Decathlon.

According to CEO Arvind Venugopal Nair, NeuroPixel spent nearly four years building deep IP assuming the competition would mostly come from other startups.

Then frontier image models exploded.

In his shutdown post, Nair said the company got “massively outgunned overnight sometime in late 2025.” He specifically mentioned Google’s NanoBanana Pro as the moment things changed.

NeuroPixel did not miss the GenAI wave. It caught it early. Then the wave got too big.

The Numbers:

🧠 Founded: 2020
💵 Funding: around $1.2M total
⚡ Claimed it could reduce catalog production timelines by 90%
📸 Claimed image production costs could drop by up to 70%
🚪 Shut down service operations in April 2026

Reasons for Failure:

General-purpose AI models caught up fast: NeuroPixel built around the assumption that fashion-specific image generation would stay technically difficult for a while. That was true for a brief period. Then larger models improved much faster than expected.
By late 2025, companies like Google were shipping image models powerful enough to erase most visible differences between specialized tools and frontier systems. The CEO openly admitted NeuroPixel’s product “didn’t hold up” against models like NanoBanana Pro.
Distribution became more important than product quality: Nair said the company assumed competitors would be other startups. Instead, the market consolidated around larger ecosystems with APIs, enterprise relationships, massive compute access, and existing user bases. Even if NeuroPixel’s output quality stayed competitive, customers naturally gravitated toward tools already integrated into broader workflows.
One unpaid client damaged the runway at the worst possible moment: The CEO said NeuroPixel’s largest customer collapsed and failed to pay for more than six months of work. For a startup already under pressure, that can end the company by itself.

Why It Matters:

AI moats disappear fast when your edge is just model quality.
In AI, distribution is starting to matter more than product quality.
Being early helps until Big Tech finally shows up.

Trend

Who Watches The Agents?

The AI market spent the last two years building agents.

Over the last year, a second layer started forming around them: the software that manages the agents.

At first, it looked scattered. A few orchestration tools here, some observability startups there, random “AI employee” demos on Product Hunt. Now the category is mature enough to map properly.

Why it Matters

The market is shifting from “can AI do the work?” to “how do you operate fleets of agents safely?” Different product. Different moat.
The winning products increasingly look like infrastructure software, not chatbots. Dashboards, logs, policies, orchestration layers, runtime environments.
The strategic fight is moving upward. The agent itself may commoditize quickly. The management layer probably won’t.

Agent Control Planes

This is the core category.

The interface here looks less like a chatbot and more like an admin console. Policies, retries, permissions, logs, audit trails, supervision dashboards, orchestration layers.

The pitch becomes: run agents like infrastructure, not experiments.

OpenHands Agent Control Plane: Launched May 6 as an “operational layer” for managing agent fleets. Includes orchestration, security, observability, scheduling, retries, spend tracking, and audit logs.
BAND: Raised a $17M seed to build communication infrastructure for the “Internet of Agents.” Focuses on agent discovery, delegation, context exchange, and collaboration across frameworks.
Wayfound: Supervision dashboard for business users. Tracks agent behavior, evaluates performance, analyzes interactions, and suggests improvements.
TectoAI: Governance layer for agentic systems. Monitors behavior, policy alignment, misuse, unauthorized actions, and compliance risks. Basically HR plus security for AI workers.

Company Runtimes

This is the founder-friendly wrapper around the same trend.

Instead of presenting orchestration as infrastructure, these startups present it as a company. The interface looks like an org chart, workspace, or operating system for a business.

The pitch is simpler: manage the company, let the agents handle the work.

Cofounder 2: Launched this week with the pitch “Run an entire company with agents.” Coordinates agents across engineering, sales, marketing, ops, and design.
Naïve: Calls itself “autonomous company infrastructure.” Focuses on planning, memory, isolation, and coordination between agent teams.
Buda: “Recruit agents to run your company as a synchronous team.” Includes marketplaces, team coordination, and live views of agents working inside browsers and terminals.
WUPHF: Shared memory, persistent knowledge bases, agents collaborating in public channels. Basically Slack for AI employees.

Underneath the branding, these companies are still building orchestration systems. Tasks, memory, roles, coordination, visibility. The org chart is just the UI layer.

Agent Observability and Debugging

Agents fail in weird ways. They loop. Misread pages. Call the wrong tools. Spend too much money. Hallucinate state. Trigger workflows nobody expected.

Companies need to see what happened after the failure.

Laminar: Raised a $3M seed to build observability for long-running agents. Captures LLM calls, browser actions, workflow traces, tool usage, replay systems, and anomaly detection for agent behavior.
AgentOps: Agent-first debugging platform. Tracks sessions, tool calls, costs, and behavior across popular agent frameworks.
LangSmith: Observability and evals for agent teams already building with LangChain or LangGraph. Useful for tracing failures and turning messy runs into test cases.

Big Tech

The category is now important enough that Big Tech is moving into the same layer.

Microsoft Agent 365: Microsoft is positioning it as a control plane for enterprise agents with governance, observability, and security controls.
OpenAI Agents SDK: OpenAI is building orchestration primitives directly into the developer stack: handoffs, agent routing, manager-style workflows.
Anthropic Managed Agents: Anthropic is building infrastructure around harnesses, sandboxes, and long-running managed agent systems.

That creates the obvious tension.

Can startups own agent management? Or does this collapse into the platforms that already control the models, cloud infrastructure, enterprise identity, and distribution?

Because the next fight is probably not over who builds the best agent. It’s over who owns the system around the agents.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

One step away from Nasdaq

Nicolás Cerdeira — Thu, 30 Apr 2026 18:50:58 +0000

Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

EO Charging, a startup building charging infrastructure for commercial EV fleets, shut down — learn more below
How to use Claude for GTM and pricing analysis
Claude Design and Claude for Creative Work are doing for design what Claude Code did for Software Engineering — learn how below

A huge thanks to today’s sponsor, Galactic Fed. Fix inefficient growth and scale smarter with a free expert marketing audit.

Overpaying for Growth? ^AD

Get a Free Marketing Audit from the team behind 600+ growth campaigns

If revenue increases only when budget does, you’re scaling cost, not efficiency. We'll help you fix the math before scaling it further.

As a Failory reader, you've been selected for one of our limited 20-minute executive review spots.

You’ll meet with senior growth specialists who will:

Assess your strategy
Identify smarter growth levers
Deliver a custom audit + action plan you can execute immediately

10 Spots Only: Confirm Yours

This Week In Startups

🔗 Resources

How to use Claude for GTM and pricing analysis

How to Compete with Legacy Incumbents

When the Dashboard Says ‘Go’ and Your Gut Says ‘No

Finally build that app this weekend (no-code hackathon) *

📰 News

X announces a rebuilt ad platform powered by AI

Jack Dorsey-backed Vine reboot Divine launches to the public

OpenAI releases GPT-5.5, bringing company one step closer to an AI ‘super app’

Lovable launches its vibe-coding app on iOS and Android

💸 Fundraising

AI customer service startup Netomi raises $110 million

Parallel Web raises $100M to build a parallel web for AI agents

German robotics startup Sereact raises $110M

Israeli AI startup Aidoc raises $150m for medical tools

* sponsored

Fail(St)ory

IPO to Zero

EO Charging entered administration in April 2026 after a failed sale process and another liquidity crunch.

Five years earlier, it was preparing to go public at a $675M valuation.

This is what happens when you try to build a global infrastructure company before the infrastructure market is ready to behave like one.

What Was EO Charging:

EO Charging started in the UK in 2014 with a clear idea: fleets were going electric, but the infrastructure required to support them didn’t exist yet in a usable form. Logistics operators didn’t just need chargers. They needed depots that worked reliably every morning when vehicles rolled out.

EO built itself around solving that operational gap.

Instead of selling standalone chargers, the company handled the full rollout inside fleet depots. That included site planning, grid upgrades, hardware installation, charging management software, and ongoing maintenance support.

If a company wanted to electrify hundreds or thousands of vehicles, EO could make the transition actually work on the ground.

By the early 2020s the company had deployed roughly 50,000 chargers across more than 35 countries. Its customers included Amazon, DHL, Tesco, Uber, and Go-Ahead.

At the center of the system sat EO Hub, the company’s charge point management platform. It coordinated charging schedules, integrated with fleet telematics tools, and helped operators manage power usage across entire depots.

In August 2021 EO announced a planned Nasdaq listing through a merger with First Reserve Sustainable Growth Corp. The deal implied a $675M enterprise value and more than $150M in expected cash. At the time, EO presented itself as a European fleet-charging leader preparing to scale globally.

The listing never happened. The SPAC deal was terminated in March 2022 as market conditions shifted and growth infrastructure companies suddenly became harder to finance.

EO responded the way many climate infrastructure startups did during that period: It kept expanding.

In 2023 the company raised around $80M and pushed further into international markets including the US, Australia, New Zealand, and Italy. The strategy aimed to turn EO into a global fleet electrification platform before the category consolidated around a few dominant providers.

That expansion came with operational weight. Each new geography required local grid coordination, installation partners, permitting workflows, and support coverage across active depots.

By late 2025 the company started narrowing its focus again. EO exited the US market, sold its domestic charger manufacturing business, and shifted attention toward its cloud charging platform. The plan was to simplify the business and move closer to a platform-led model built around software and fleet charging management.

A few months later the company began a sale process. No transaction followed.

The Numbers:

📅 Founded: 2014
🌍 Chargers deployed: ~50,000 across 35+ countries
🏢 Major customers: Amazon, DHL, Tesco, Uber, Go-Ahead
📈 SPAC valuation attempt (2021): $675M
💵 Equity round (2023): ~$80M
👥 Employees at administration: 93
❌ Immediate layoffs: 69

Reasons for Failure:

The business model looked scalable on slides, not in depots: EO sold a clean promise. One partner handles everything. Hardware, installation, grid coordination, software, uptime support. Fleet operators loved that. Startups usually don’t. Every depot became its own infrastructure project. Custom timelines. Custom engineering constraints. Custom integrations. That slows revenue recognition and eats working capital fast.
International expansion arrived before operational leverage: PwC administrators explicitly connected losses to overseas expansion across the US, Australia, New Zealand, and Italy. Expanding a hardware-plus-installation model internationally is not like opening a sales office. It means new permitting systems. New grid relationships. New subcontractors. New reliability expectations.
The SPAC collapse removed the fuel mid-flight: In 2021 EO expected more than $150M from a Nasdaq listing. That capital would have funded deployment capacity, hiring, and inventory. Infrastructure companies scale through balance sheets as much as through revenue. When the deal disappeared in 2022, the plan didn’t disappear with it. The company kept moving as if the market would reopen. It didn’t.

Why It Matters:

Enterprise infrastructure customers scale depot by depot, not quarter by quarter. Their rollout speed quietly sets your growth ceiling.
One-stop-shop positioning turns you into the system integrator holding everyone else’s risk. Grid delays, installs, and timelines all land on your balance sheet.

Trend

Claude Goes After Designers

Claude Code already changed how a lot of people build software.

Now Anthropic seems to be aiming that same playbook at design.

In the last two weeks, it launched Claude Design and Claude for Creative Work, connecting Claude to tools like Adobe Creative Cloud, Blender, Autodesk Fusion, Ableton, SketchUp, Splice, Affinity, and Resolume.

AI design is not new. But a Claude Code-style workflow for design would be.

Why it Matters

Claude is moving into tools AI usually could only talk about: Blender, Fusion, Ableton, Adobe, and SketchUp are not simple text boxes. If Claude can help operate those tools through natural language, complex software becomes easier to use without making the software itself dumber.
The workflow is starting to connect instead of resetting every time: Most AI creative tools are impressive until you need to keep working. Claude Design plus the connectors suggests a different direction: generate the rough asset, refine it, export it, move it into Canva, Adobe, Blender, or Claude Code, and keep going.

Claude Design

Claude Design launched on April 17 through Anthropic Labs as a research preview for Claude Pro, Max, Team, and Enterprise users.

The basic idea is simple. You describe what you need, and Claude creates a visual draft. That could be a prototype, a pitch deck, a one-pager, a slide, a dashboard, a landing page, a marketing asset, or some HTML-style output.

Then you keep working on it.

You can refine it through conversation, leave inline comments, make direct edits, use custom sliders, and export the result to formats people actually use: Canva, PDF, PPTX, standalone HTML, or Claude Code.

That matters because AI design tools often break down right after the first output. The first draft looks impressive, but then the real work starts somewhere else. Rebuilding it. Editing it. Making it on-brand. Turning it into something a team can actually use.

The Claude for Creative Work Connectors

Then came Claude for Creative Work.

To me, this was the more revealing announcement, even if it got less attention than Claude Design.

Anthropic announced a new set of connectors with tools across design, 3D, music, video, and live visuals. The list includes Adobe Creative Cloud, Blender, Autodesk Fusion, Ableton, Splice, SketchUp, Affinity by Canva, and Resolume.

This is what makes the Claude Design launch feel less isolated.

Claude Design gives Claude a visual workspace. The connectors give it a path into the professional tools where the work already happens.

Adobe brings Claude into the Creative Cloud world, including tools like Photoshop, Premiere, and Express. Affinity by Canva focuses on production tasks like batch image edits, layer renaming, exports, and custom feature generation inside the app.

Blender gives Claude access to its Python API and documentation. Autodesk Fusion lets people create and modify 3D models through conversation. SketchUp can turn a prompt into the starting point for a room, furniture setup, site concept, or 3D scene.

The common thread is simple: creative work is full of tool friction.

The tools are powerful, but they are deep, messy, and fragmented. A lot of the work is not some grand act of creativity. It is changing formats, finding the right command, applying the same edit twenty times, cleaning up files, moving assets around, or turning a rough idea into something editable.

The connectors allow Claude to handle all of these. So, it seems to me that design may be getting its Claude Code moment.

As Claude Design keeps improving, and more tools get connected to it, Claude has a real shot at doing for design what it already did for software engineering: make expert work faster, more accessible, and much easier to iterate on.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico

Failory

The One-Hour Home

Your best decisions start with listening AD

This Week In Startups

🔗 Resources

📰 News

💸 Fundraising

Fail(St)ory

The Missing Village

Trend

The AI Title Fight

Why it Matters

What People Are Saying

The Trend

Help Me Improve Failory

Outspent and Outdelivered

Can you afford your next hire? AD

This Week In Startups

🔗 Resources

📰 News

💸 Fundraising

Fail(St)ory

The Second Grocery War

Trend

AI Video Specializes

Why it Matters

Solaris

Atlas

The Trend

Help Me Improve Failory

Electricity Through Thin Air

This Week In Startups

🔗 Resources

📰 News

💸 Fundraising

Fail(St)ory

Charging Through the Air

Trend

OpenAI Goes Silicon

Why it Matters

Jalapeño

The Trend

Help Me Improve Failory

Buying Better Weather

This Week In Startups

🔗 Resources

📰 News

💸 Fundraising

Fail(St)ory

AI Meteorologist

Trend

Extensible SaaS

Why it Matters

The Signals

The Trend

Help Me Improve Failory

80 Acres in One

This Week In Startups

🔗 Resources

📰 News

💸 Fundraising

Fail(St)ory

80 Acres Farms

Trend

Micro-SaaS Flipping

Why it Matters

The Signals

This is Happening a Lot

The Trend

Help Me Improve Failory

Tap. Declined.

This Week In Startups

🔗 Resources

📰 News

💸 Fundraising

Fail(St)ory

Crypto Mastercard

Trend

AI Math

Why it Matters

Your best decisions start with listening ^AD

Can you afford your next hire? ^AD