Physical AI is having a busy fall

The next few weeks will take Spatial Capital from a private founder roundtable to a live podcast, TechCrunch Disrupt, and our first SF TechWeek event.

Here’s a closer look at what we’re up to.

Our largest event yet

On October 7, Spatial Capital and Capgemini will host a Physical AI Demo Day at Capgemini’s San Francisco office.

This is our first SF TechWeek event and our largest in-person gathering to date. We expect approximately 150 investors, founders, and industry leaders.

Several Spatial portfolio companies will give short presentations and demonstrate their products in person. Each team will also have its own table, giving attendees time to meet the founders, see the technology up close, and dig into the businesses.

It’s rare to bring this many of our portfolio companies together in one room. For investors interested in Physical AI, it’s an efficient way to see where the market is heading and meet some of the founders pushing it forward.

Request an invitation →

ConeLabs heads to TechCrunch Disrupt

ConeLabs has been selected as a finalist for TechCrunch Startup Battlefield.

The company uses AI and computer vision to automate infrastructure inspection and condition assessment. Its software analyzes imagery of roads, bridges, and other assets to identify, measure, and track defects.

That helps engineering firms and infrastructure owners inspect assets faster, produce more consistent assessments, and monitor how defects change over time.

Congratulations to CEO Albert Mansour and the entire ConeLabs team. We’re excited to see them take the stage at Disrupt.

Did someone forward this to you? Want more? 

A live conversation about Physical AI

On September 10th, Doug will be joining a live podcast to discuss what it takes to build, finance, and scale Physical AI companies.

He’ll get into where founders are finding commercial traction, what investors often miss about the category, and the realities of bringing AI into industries shaped by hardware, field operations, and long sales cycles.

The team spends much of our time helping technically ambitious companies turn complex products into durable businesses, and are excited to share some of this publicly.

Inside Spatial: Vivid Machines

On September 17th, Spatial Capital LPs will join Vivid Machines founders Jenny Lemieux and Jonathan Binas for a private virtual roundtable.

Vivid is building a computer vision and edge AI platform for commercial fruit agriculture. Its camera systems travel through orchards collecting detailed information about fruit count, size, and variability, giving growers earlier and more accurate information about their crops.

The founders will discuss the product, their commercial progress, and what comes next.

These private sessions give our LPs direct access to the people we back. They can hear the story firsthand, ask detailed questions, and follow the company’s progress over time.

More to come

Building a company that brings AI into the physical world? Apply here

Interested in learning more about investing with Spatial? Contact us.

Best, Spatial Capital

Few pieces of infrastructure have shaped modern life more than the electrical grid. Its history helps explain why a new generation of infrastructure companies is emerging today.

The First Grid

On the morning of September 4, 1882, the streets of lower Manhattan looked much as they had the day before. Beneath them, however, something entirely new had been assembled.  For months, workers had dug trenches through the city, laying nearly 100,000 feet of copper wire beneath the cobbled streets. Those cables led back to a coal-fired generating station on Pearl Street, where Thomas Edison was preparing to switch on a new kind of utility. When he did, electricity flowed to 80 customers, powering about 400 light bulbs.

Thomas Edison, outside the Pearl Street Power Station

At the time, it barely made the news. The New York Times mentioned the opening in its Miscellaneous section. A year later, the system still served less than 500 customers. Few could have imagined that this experiment would become one of the largest infrastructure projects in human history.

Electricity spread slowly at first. Early generating stations could supply only a few city blocks, so every neighborhood needed its own source of power. That changed as engineers learned to transmit electricity over longer distances using alternating current. Power no longer had to be generated where it was consumed.

AC distribution opened up the opportunity for the Niagara Falls hydroelectric project. Completed in the 1890s, it carried electricity more than 20 miles to Buffalo, proving that a single generating station could power an entire city.

Construction of the turbines and penstocks at the Niagara Falls Power Company in the 1890s.

By the turn of the twentieth century, electric lights illuminated city streets, factories, and department stores across America. But only a few miles beyond those growing cities, kerosene lamps and hand pumps remained a way of life.

Bringing Electricity to America

For millions of Americans, an electric future still felt very far away. While cities added skyscrapers, factories, and electric streetlights, millions of rural families were being left behind. In 1935, only about one in ten farms had electricity. Water came from hand pumps, barns were lit by lanterns, and washing clothes could consume an entire day. Running power lines across sparsely populated countryside was expensive, and private utilities had little financial incentive to do it. 

Franklin Roosevelt wanted to change that. As part of the New Deal, he created the Rural Electrification Administration, a federal program that provided low-interest loans to rural electric cooperatives to build power lines where private utilities would not.

What followed became one of the largest infrastructure projects in American history. Crews crossed fields with augers and shovels, setting pole after pole by hand. In just five years, the program financed nearly 300,000 miles of new distribution lines, enough to circle the Earth more than a dozen times, bringing electricity to roughly 1.5 million farms. Many of those distribution lines, and the rights-of-way they established, remain part of the electrical grid today.

REA crews extending power lines across rural Georgia in the 1940s. Many of the routes established during the rural electrification era still form part of today's electrical distribution network.

Did someone forward this to you? Want more? 

Building the Modern Grid

When World War II ended, America started building again. New suburbs spread outward from every major city. Factories expanded. Shopping centers appeared where farmland had once stood. Air conditioners, televisions, refrigerators, washing machines, and countless other electric appliances quickly became part of everyday life.

Utilities found themselves in a race they couldn't afford to lose.

Electricity demand climbed by roughly 7% every year throughout the 1950s and 1960s, doubling about once a decade. Every forecast seemed to underestimate how much power Americans would need next.

Levittown became the model for postwar suburban growth. As millions of families moved into new homes filled with electric appliances, utilities raced to keep up with soaring demand.

Across the country, crews returned to work. Surveyors marked new rights-of-way, linemen erected transmission towers and strung new conductors, and substations appeared at the edge of growing cities.

Between 1950 and 1970, the United States added hundreds of thousands of miles of transmission lines, connecting power plants, substations, factories, and cities into one of the largest machines ever built.

Keeping the Lights On

As the grid expanded, reliability became the next challenge.

On November 9, 1965, a protective relay at an Ontario generating station faltered. Within minutes, a cascading failure swept across the interconnected grid. More than 30 million people from Ontario to New York lost power, creating one of the largest blackouts in North American history.

Passengers walk through a New York City subway tunnel during the 1965 Northeast Blackout.

For years afterward, engineers studied what had happened. By the time the next great blackout arrived nearly four decades later, electricity was far more critical to our daily lives. 

On a sweltering August afternoon in 2003, another fault triggered a cascading failure across the North American power grid. Within minutes, more than 50 million people from Cleveland and Detroit to Toronto and New York City were without electricity.

As air conditioners failed, office towers emptied into streets already choked with traffic. Nearly 600 subway and commuter trains stopped between stations, trapping thousands of passengers in tunnels where temperatures approached 100°F. Emergency crews spent hours evacuating riders, hundreds of whom walked through dark subway tunnels to reach the surface.

Elevators froze in place, trapping hundreds more, including visitors at the top of the Empire State Building. ATMs went dark. Neighborhoods that depended on electric pumps lost water. Grocery stores and restaurants watched refrigerated food spoil in the summer heat.

The blackout prompted another wave of investment in grid reliability, setting the stage for the next chapter in the grid's evolution.

Grid Modernization

The blackout of 2003 became another turning point.

Over the next several years, engineers studied what had happened and began rethinking how the grid should operate. When Congress passed the American Recovery and Reinvestment Act in 2009, nearly $10 billion was set aside to modernize the nation's electric infrastructure.

Across the country, digital meters replaced analog ones. New communications networks connected substations that had long operated in relative isolation. Sensors began streaming data back to utility control rooms, giving operators a view of the grid that would have been unimaginable only a decade earlier.

The ISO New England electrical control room, circa 2010

The Cost of Failure Has Changed

By the early 2010s, the electrical grid was becoming connected. Communications networks gave utilities a far better view of the network. But when something went wrong, pinpointing the fault and understanding its cause could still take valuable time, or worse, cause disaster.

On the morning of November 8, 2018, a steel C-hook, installed years before Franklin Roosevelt launched the Rural Electrification Administration, finally gave way. The energized conductor fell into the dry brush below.

Driven by powerful winds, the Camp Fire swept forward with devastating speed. Eighty-five people lost their lives. More than 18,000 homes and buildings were destroyed. Investigators ultimately traced the ignition back to that single worn piece of hardware.

A washer and dryer still standing after the Camp Fire in Paradise, CA

The Next Generation

The Camp Fire wasn't the result of new technology failing. It was the result of aging infrastructure that had quietly remained in service for decades.

That illustrates a challenge facing utilities across North America. Much of today's electrical grid was built by previous generations. Rather than rebuilding it from scratch, utilities are maintaining, upgrading, and gradually making it more intelligent.

Much of the grid was built over generations, and utilities don't always know exactly what equipment remains in the field. A new generation of infrastructure companies, including Looq AI, one of our portfolio companies, is helping utilities build digital twins of their infrastructure, creating accurate assets of power poles, conductors, and other equipment that, in many cases, haven't been comprehensively documented for decades.

Watch a 30-second flythrough of Looq's digital reconstruction of utility infrastructure. 

But knowing what you have is only part of the challenge. Utilities also need to understand what's happening across the network in real time. That's what led us to invest in GridIQ.

GridIQ continuously monitors the electrical network, detecting, classifying, and pinpointing faults with meter-level precision. Instead of searching miles of power line after something goes wrong, utilities know where to look and can respond far more quickly.

For nearly 150 years, engineers have made the electrical grid larger, stronger, and more reliable. Every generation inherited what the last had built, then found new ways to improve it.

Today, that challenge is no longer simply expanding the grid. It is understanding it. As sensors, computer vision, and AI bring greater visibility into aging infrastructure, we believe the next chapter will be making the grid smarter.

Summer usually slows things down. Apparently Physical AI didn't get the memo.

The past few weeks brought a handful of announcements that caught our attention, plus a nice product release from one of our portfolio companies.

In this issue:

• 1X takes another leap in robotic dexterity

• Autodesk acquires MaintainX for $3.6B

• Odyssey reaches a $1.45B valuation

• EyePop broadens access to computer vision

1X NEO robotic hands picking a grape

Market

🤖 Robotic Hands - The Next Frontier?

One of the biggest challenges in robotics is object manipulation. While walking gets most of the attention, picking up a grape without crushing it is surprisingly difficult.

Elon Musk recently made the point that "you must solve the hands problem" before you can build a general-purpose humanoid robot. He went on to estimate that roughly 25% of the technology in the robot is in the hand itself.

A robotic hand needs dozens of degrees of freedom, tactile sensing, precise force control, and enough durability to repeat the same task thousands of times a day.

1X recently unveiled its latest robotic hands. Compared to earlier designs, they have more joints, more degrees of freedom, and much more natural finger motion. More joints alone aren't enough. They have to move together while continuously adjusting grip force as objects slip, bend, or change weight. Watching the videos, it's easy to see how quickly manipulation is improving.

🗞️ Robotic hands are a fascinating topic. There are several competing approaches to dexterity, force control, and tactile sensing. I'll dig into those in a future post.

💰 Autodesk acquires MaintainX for $3.6B

Autodesk announced its acquisition of MaintainX for $3.6B (link).

MaintainX reportedly reached about $135M ARR before the acquisition. The company helps manufacturers, utilities, facilities, and other industrial businesses manage maintenance, inspections, work orders, and frontline operations.

Autodesk's acquisition is another large bet on software that manages and understands physical assets. This is adjacent to what our portfolio company, Vizonare, is building, and provides further validation that this category is attracting strategic interest.

MaintainX is focused on helping people manage physical assets. Another area we're watching closely is AI's ability to understand those assets in the first place.

🦄 Odyssey Joins the Unicorn Club

Odyssey announced a new financing at a reported $1.45B valuation (link). The startup builds world models, AI systems that generate synthetic versions of the physical world. Their goal is to let AI understand, predict, and eventually interact with real environments.

We met the founders back in 2024 when the company was raising at roughly a $25M valuation. We didn't end up investing, and like every VC, we’ve added this one to our growing collection of "remember when..." stories.

It's great to see them doing well.

Portfolio

🎥 EyePop makes computer vision easier to build

Computer vision is finding its way into far more than autonomous vehicles and factory automation. We're seeing new applications across transportation, construction, manufacturing, utilities, agriculture, and dozens of other industries.

Building those applications still requires developers to stitch together models, workflows, and infrastructure before solving the customer's actual problem. Most projects require a dedicated AI team.

EyePop just launched Workflow Designer to dramatically reduce this effort. Developers build complete vision workflows visually, then export the underlying Python code. Object detection, OCR, tracking, counting, and classification become reusable building blocks instead of custom engineering projects.

EyePop recently demonstrated Workflow Designer using a vehicle-counting application. A large North American ferry operator is already using it to determine, in real time, how many vehicles can still be loaded before departure.

Swap in different vision models and the same platform can inspect manufactured parts on a production line, monitor construction progress, read utility meters, or identify safety issues. Whether it's a loading dock, a ferry terminal, or a factory floor, the opportunity is the same: turn visual information into something a business can act on.

👉 Learn more about EyePop

It's been a fun month to watch this space and we can’t wait to see what comes next. 

If you have questions, feedback, or a company we should meet, just hit reply. We read every email.

Best,
Doug & the Spatial Capital team

Last month, one of our portfolio companies, mbue, was acquired by Bluebeam. The announcement got me thinking about where value is actually accumulating in this market.

For years, the software industry rewarded people who knew how to build software. That made sense. Software was hard to build. Today it's getting cheaper fast. A founder can spin up a prototype in a weekend. Small teams ship products that would've needed dozens of engineers a decade ago. Coding assistants write a meaningful portion of production code.

Knowing how to build software still matters. Knowing what to build matters more.

Jean Pierre Trou (CEO of Mbue), Doug Griffin (General Partner of Spatial Capital), Babak Rasolzadeh (Advisor Spatial Capital)

Jean-Pierre Trou, mbue's founder, spent 20 years as a practicing architect before starting the company. Late nights at his dining room table in Austin, red pen in hand, reviewing 250-page drawing sets for a single office building. Checking for the kinds of errors that cost six figures to fix once construction starts: uncoordinated curtain wall details, missing penetrations, conflicting mechanical routes. A full week of work, for one project.

He didn't start with a model and go looking for a problem. He started with a problem he'd lived with for two decades.

Architects and engineers spend 30-40% of their time on quality control, and they're not particularly good at it. Human eyes miss things. Those mistakes compound into 5-10% of total construction costs. The U.S. construction industry wastes over $100 billion a year on errors, changes, and omissions.

mbue was built around that. Years of construction work generated not just software, but data: blueprints, specifications, markups, submittals, approvals. The artifacts created by real projects. That combination is hard to replicate.

The same pattern runs through our portfolio. Albert Mansour studied civil engineering and spent years selling measurement technology before founding Conelabs. Levi Farrand was a wildlife guide, obtained master degrees in environmental conservation and physical geography, all before starting Deep Forestry. Jenny Lemieux grew up on a farm, earned a master's in AI, and built Vivid Machines around challenges she understood firsthand.

None of them started by asking what a model could do.

Enterprise software still gets sold by people. Relationships matter. Industry reputation matters. Knowing how decisions get made inside a firm matters. A founder who's spent years inside an industry starts with credibility that outsiders spend years trying to earn.

Bluebeam could've hired engineers. They could've licensed models. Instead, they acquired a company that had spent years learning how construction documents actually move through the real world.

As AI lowers the cost of building software, the question shifts. The winners won't be the teams that generate the most code. They'll be the ones who understand their customers well enough to know which problems are worth solving.

As AI keeps lowering the cost of building software, the scarce thing isn't the technology. It's the 20 years at the table.

We backed JP because he'd never stopped doing the work. Turns out that's exactly what Bluebeam was buying.

Best,
Doug Griffin, Spatial Capital

The weather's heating up, and so is the news from Spatial. Here's the latest from the fund and portfolio.

In this issue:

• Vivid Machines counts grapes at Gallo's vineyards

• Looq AI wins "Power Player" at DTECH, the grid's biggest conference

• Maket's let’s you chat with your floorplan

• Conelabs heads to the International Bridge Conference

• Doug speaks on AI, jobs, and whether college still makes sense

Firm News

On May 29th, Doug joined a panel at Piedmont High School covering what’s next in AI - namely Vibe Coding, Agentic AI, and Physical AI. The room included Jonah Sachs, named by Fast Company as one of the 50 most influential social innovators and Jaime Peters, President of Countable - a platform for social impact. 50 people packed in on a Friday night during a holiday weekend, and the conversation quickly evolved from the technology itself to what AI means for jobs and whether college is still the right call. No easy answers, but good to see the questions being asked in the right room.

Lee Bellon, EIR at Spatial Capital, is joining Vizonare as Fractional CMO. He'd been helping with go-to-market informally for a few months. Then Vizonare added a serious upgrade to its AI-for-property-management platform, and suddenly there was a real GTM story to tell. Lee will help tell it.

Lee, Doug, and Przemek (CEO Vizonaire) celebrate in Sausalito

Portfolio News

Vivid Machines hosted a Physical AI Field Day at a Gallo vineyard outside Healdsburg. (Gallo produces about 30% of all US wine. Not a small sandbox.) We met with members of their research team, one of the largest in ag-tech, to watch the platform run live: counting every cluster, on every vine, in real time. We followed it up with some wine tasting of our own, at a happy hour with local LPs.

Jonathan & Jenny, co-founders of Vivid Machines, join Doug for a Spatial Happy Hour

Maket added an agentic layer to its AI floor plan editor. You describe what you want, the design updates in real time. No weeks waiting on a drafter, multi-revision email chains, or pricey hourly charges for a bathroom addition. Maket’s AI architectural assistant creates code-compliant drawings in minutes, something an LLM cannot do, and is getting major improvements on a regular cadence.

Did someone forward this to you? Want more? 

Looq AI won the "Power Player" next-gen vision award at DTECH, the largest electrical transmission and distribution conference, with 20,000 energy professionals, 700 exhibitors, and utilities from 40 different countries. These people literally keep the lights on.

The US electrical grid is aging, undersized for the EV and data center boom, and overdue for a serious upgrade. It starts with knowing what you have and where it's broken. Looq's computer vision and AI platform does exactly that. See more in the video below.

Conelabs is exhibiting at the International Bridge Conference next week. (Yes, we had the same reaction. Apparently bridges have their own conference.) Their platform converts drone footage into inspection-grade 3D models, helping surveyors spot cracks and defects before they become headlines. Our next question, “is there a conference for potholed roads?”

Who’s Hiring

EyePop is hiring an Account Engineer. Helping their growing client list implement vision language models into real products.

Drafter is hiring a Full Stack Engineer. Drafter helps engineers translate design intent into manufacturing-ready drawings - the unglamorous work that determines if something actually gets built.

We have more exciting news that we can’t wait to share. Stay tuned!

Best,

Doug Griffin, Spatial Capital

The first wave of Physical AI is reaching commercial scale. We published a longer reflection on this recently. Today’s newsletter highlights several proof points that reinforce the trend:

• Waymo, Waabii, and Zipline raise massive rounds to scale autonomy

• NVIDIA releases Physical AI with “common sense” reasoning

• DeepMind - adapts its robot training for playable worlds

• Our portfolio advances intelligence for the built world

💰 Recent Investments Physical AI

Recent funding rounds reflect autonomy companies moving from regional pilots into broader market deployment.

Waabi autonomous freight

Waabi raised $750M to advance its Physical AI platform across autonomous trucking and passenger vehicles, alongside a partnership with Uber. While consumer autonomy receives most of the attention, freight remains one of the largest cost centers in global commerce.

Waymo raised $16B on a $126B valuation to expand into new markets and scale its operations. The autonomous ride-hailing scale-up has already driven the distance “of the moon and back 290 times.” Now it’s adding 20 new cities in 2026, including the complicated roadways of London and Tokyo.

Zipline raisesd$600M at $7.6B valuation. The company began commercial operations in 2016 delivering blood to hospitals in Africa. It now operates in multiple U.S. cities and partners with major brands for on-demand delivery. Deliveries are reportedly growing a staggering 15% week over week. It may already be one of the largest autonomy companies operating at scale.

Robotics and autonomy are the tip of the spear in Physical AI today. They attract funding because they are ambitious, capital-intensive, and platform-oriented. They have also been under development for more than two decades. The current funding cycle builds on years of technical iteration, hardware refinement, simulation tooling, and regulatory groundwork.

This is the first large wave of Physical AI to reach meaningful commercial scale. But there are many more to come.

Did someone forward this to you? Want more? 

💡 Research Continues to Advance

Funding momentum is supported by steady technical progress. Here are a few of the recent announcements:

NVIDIA released Cosmos Reason 2, a vision-language model focused on spatial and temporal reasoning. The goal is to improve how machines understand how environments evolve and how actions unfold over time. For instance, if a ball rolls out into the road while driving, assume that a child might jump out to chase it.

DeepMind launched Genie 3 last summer. Genie 3 is the first interactive general purpose world model that enabled users to conjure up physically plausible and photorealistic 3D environments from text. We wrote about the implied impact on robotics last summer. Now Waymo announced a new world model built on Genie 3.

DeepMind launches Project Genie. DeepMind has also released Project Genie, which allows users to generate playable, physics-based 3D environments from text or images. While lighter in tone, it demonstrates continued progress in generative world modeling.

DeepSeek is expected this v4 of its model this week. Last year, China-based DeepSeek drew attention after releasing a competitive model trained under tighter GPU constraints, relying on architectural efficiency to close the gap with Western models. Researchers are watching closely to see how the next iteration may translate into practical gains for robotics and other physical systems.

Portfolio Signals

Physical AI is expanding across the systems that support the physical economy. Autonomy and robotics are the most visible today. Infrastructure and asset intelligence are advancing more quietly, but address markets of comparable scale.

Here is how a few of our portfolio companies are adding intelligence to the physical world.

The CEO of Vizonare, showcasing their “Intelligence for Commercial Real Estate” Platform

Vizonare placed 2nd at SF Tech Weekend, showcasing its intelligence platform for commercial facility management. Large commercial properties often operate with fragmented systems for maintenance, energy optimization, compliance, and capital planning. A recent BCG report outlines how AI and digital twins could reduce billions in long-term infrastructure and real estate management costs.

ConeLabs was named one of the Top 50 ConTech startups. Our physical infrastructure (e.g. bridges, buildings, and roadways) are aging. Conelabs addresses this with AI-powered inspection automation, enabling our tax dollars to be more efficiently redirected towards fixing assets. This week it is showcasing its AI platform for physical infrastructure inspection at GeoWeek.

4Point CEO showcasing early customer locations

4Point advances its subsurface intelligence for mining. Leveraging its Spatial Intelligence Platform (SIIM), 4Point turns sparse, biased geological data into accurate and stable 3D predictions of mineralization and material behavior. Recent benchmarks show strong predictive accuracy and reduced error compared with classical methods.

Final Thoughts

Robotics, autonomy, and world models are the most visible applications of Physical AI advances. They provide a clear demonstration of machine perception and reasoning interacting directly with the physical world.

They are not the only markets where these capabilities will compound.

Large portions of infrastructure, facilities, logistics, agriculture, and asset management remain early in digitization. Compared to autonomy, these sectors have seen less concentrated investment in Physical AI. The economic incentives, however, are substantial. Even incremental improvements in uptime, maintenance planning, or asset visibility can translate into meaningful cost reductions across large portfolios.

The first wave has been building for years. Additional waves are likely to follow as tools mature and deployment costs decline.

– Doug Griffin, Spatial Capital

Two of the biggest names in AI left their long-time roles to start world model companies. Investors are responding with very large checks.

Yann LeCun’s new company, Advanced Machine Intelligence, reportedly raised capital at a $3.5 billion valuation. Fei-Fei Li’s World Labs raised $250 million in its first round and is reportedly talking about raising again at around $5 billion.

World models are a core building block of Physical AI. They allow software to understand, navigate, and reason about the real world, moving beyond just text, images, or code. This capability is critical for robotics, autonomy, industrial systems, and any AI that must operate safely outside a browser window.

We have followed Physical AI and world models for years and are happy to see investors funding the sector. These valuations do raise questions, though. The technology direction seems sound, but the investment math might be flawed.

Let's take a closer look.

Who Are These Founders?

If you don’t follow AI research closely, you might not grasp how influential these two people are.

Yann LeCun is one of the architects of modern deep learning. His work supports much of today’s computer vision and neural network research. He spent over a decade leading AI research at Meta.

Fei-Fei Li is a central figure in computer vision and perception. She led the creation of ImageNet, which helped start the deep learning boom. For years, she has focused on spatial intelligence, which is how machines perceive and interact with the physical world.

These are serious researchers with long track records. Their interest in world models strongly validates the importance of this direction.

Credibility alone does not guarantee a successful investment, however.

Did someone forward this to you? Want more? 

Why Capital Is Moving Here

World models are getting attention because today’s dominant AI systems have limitations. LLMs excel at generating text. Intelligent behavior in the real world requires different capabilities: the ability to predict outcomes before acting, an understanding of cause and effect, and a basic grasp of physics and time. Without these capabilities, autonomy fails.

Put a system into a factory, a warehouse, or onto a street corner, and any gaps in this knowledge show up quickly. Bad decisions damage equipment, interrupt operations, and put people at risk.

As AI moves into physical environments, investors need the next foundational layer. World models are a clear candidate for that role.

What World Models Actually Do

A world model answers one question: Given the current state of the world and an action, what happens next?

This capability is essential for planning, learning, and control. Without it, systems tend to work only in narrow, scripted scenarios, require massive amounts of data, and fail when they encounter common edge cases.

This is why world models are constantly discussed in robotics and autonomy research. They are necessary for creating truly capable systems.

From Perception to Action

Fei-Fei Li and Yann LeCun are making the same core point: perception is not intelligence. Real intelligence must perceive the world, take action, observe the result, and adjust. That demands an understanding of space, time, and physics. It needs to understand cause and effect.

World models are one way to enable that. They are an important piece of Physical AI, but they also require different assumptions and longer timelines than most software systems.

The opportunity may be large. Whether the returns match is a separate question.

The Timeline Question

Multi-billion-dollar valuations suggest investors believe this technology is near deployment. The founders themselves are more cautious.

Both LeCun and Li openly discuss missing pieces, evolving architectures, and the need for breakthroughs that will take time to translate into products. Research progress will happen long before commercial reliability.

From an investment perspective, this usually means slower revenue ramps and fewer winner-take-all outcomes.

But Physical AI may deploy more slowly and unevenly than a typical SaaS startup. Early systems will appear in controlled environments where risk is managed carefully. Edge cases must be generated and tested repeatedly. Trust will be earned over time.

This is the correct way to deploy technology that interacts with the real world. It is not a fast track to massive revenue.

The Venture Math Problem

Here’s the uncomfortable part: venture funds need outcomes that move the needle. At valuations like these, that becomes difficult.

A $100 million investment at a $5 billion valuation buys roughly 2% of a company. After future rounds, that stake may end up closer to 1%. Even in a strong outcome—say a $50 billion exit—that 1% delivers a 5x return. That’s a good venture result, but it won’t return the fund.

Venture portfolios are rarely that balanced. When large amounts of capital are concentrated in a small number of high-valuation bets, the downside becomes harder to offset. If one of those bets fails to produce a meaningful return, the impact on the fund can be material.

Investors making these deals understand the math. They are underwriting much larger outcomes, assuming world models become a foundational layer of the next computing era.

That assumption may prove correct. At these entry prices, the risk is that it needs to.

Final Thoughts

The recent funding rounds around world models confirm the importance of this space.

But belief in a technology is not the same as belief in a price. At these valuations, returns depend not just on whether world models work, but on how quickly, how broadly, and where value actually accrues.

World models will unfold over the next decade. From a venture perspective, the challenge will be understanding which bets translate into venture-scale outcomes.

– Doug Griffin, Spatial Capital

CES is one of the largest technology conferences in the world. Every January, more than 100,000 attendees gather in Las Vegas to see what companies have been building and how they’re thinking about the next phase of technology.

The scale is hard to overstate. Some booths are the size of football fields. The main halls are typically filled with familiar consumer tech pushed to extremes—wall-sized TVs, thinner laptops, smarter appliances—while more experimental work tends to live in side halls or private rooms.

The floor opens later today, but based on advance announcements, we expect that Physical AI will be hard to miss this year. Here’s what we’ll be watching for.

🚗 Autonomy: Reasoning Behind the Wheel

Autonomous vehicles have already moved beyond the demo phase. In cities like Phoenix and San Francisco, people regularly take rides in fully driverless Waymo vehicles.

At CES, attention is expected to shift toward how autonomy systems make decisions in complex, real-world situations. Similar to how reasoning models reshaped large language models, reasoning-focused Vision-Language-Action (VLA) models are starting to play a larger role in autonomy.

In his keynote yesterday, NVIDIA CEO Jensen Huang announced Alpamayo (link), an open-source reasoning model for autonomous vehicle control. Alpamayo introduces an interpretation and reasoning layer that helps autonomous systems handle uncertainty and choose between multiple possible actions. 

Alpamayo, NVIDIA’s autonomous driving model, navigating urban streets in San Francisco.

Driving involves negotiation, timing, and tradeoffs that are difficult to capture with fixed rules. Models like Alpamayo are designed to sit closer to control, shaping how perception inputs translate into driving behavior.

We expect autonomy discussions at CES to focus more on decision-making and system behavior.

Did someone forward this to you? Want more? 

🤖 Robots Leave the Lab

Robotics will be another major area of focus on the show floor. Here are two systems that were announced leading up to the show:

Hyundai and Boston Dynamics are expected to showcase humanoid robots designed for factory environments (link). These systems target industrial settings with repetitive tasks and defined workflows. We expect to see additional announcements in this area, highlighting continued progress toward real deployments on factory floors.

Atlas, Boston Dynamics’ humanoid robot, designed for work in factory settings

LG will also show a household assistant robot called CLOiD, designed for domestic tasks. This will likely attract attention, but widespread home deployment remains further out. Home environments are far less predictable than factories, and the range of tasks required is difficult to train for and execute safely. Still, these systems offer a glimpse into how robotics companies are thinking about longer-term consumer applications.

LG’s CLOiD robot, designed to assist with household tasks in home environments.

Across both factory and home use cases, robots are being presented as systems meant to operate alongside people, using human-scale tools and navigating human spaces. At CES, conversations will focus on reliability, safety, and how these machines fit into existing environments, rather than one-off demos.

📸 Sensors at Scale

Sensors will underpin much of what’s on display. We expect to see several announcements around new lidar and sensing platforms tied to autonomy and robotics. For instance, in advance of the show, lidar manufacturer Hesai announced plans to double its production capacity, citing increased demand from both markets.

We expect conversations around sensing at CES to focus on availability, cost, and readiness for real-world use.

Hesai lidar reflects growing demand for sensing hardware in Physical AI systems.

Physical AI is Arriving

Across autonomy, robotics, and sensing, CES will show how Physical AI is moving from concept to deployment. We’ll follow up after the show with what mattered—and what actually held up on the floor.

Doug Griffin, Spatial Capital

Meta launched its new Ray-Ban “Display” glasses last week — their first pair with a built-in screen. Alongside them came a neural wristband, based on their CTRL Labs acquisition, that translates small electrical signals from your hand into gestures.

The hardware is clever and reasonably stylish. But could this form factor become the next computing platform after the smartphone? Mark Zuckerberg seems to think so.

🕶️ Why Glasses Matter for AI

Attempts at smart glasses go back more than a decade. With cameras to see the world and microphones to hear it, the form factor captures the wearer’s perspective more naturally than any other. Yet efforts to scale smart glasses have never succeeded.

This time may be different — not only because components are smaller and more powerful, but because glasses now have AI. This could enable real-time contextual understanding of the world around the user. That means what you see and hear can be analyzed instantly and turned into guidance, memory aids, or translation without breaking your flow. And unlike the phone form factor, which forces you to stop and look down at a screen, smart glasses with a display can place that context directly into your field of view.

Mark Zuckerberg put it plainly: “glasses are the ideal form factor for personal superintelligence.” Although this may sound aspirational, it’s why Meta and others are spending billions to make glasses work.

Did someone forward this to you? Want more? 

What’s Missing Today

The Ray-Ban Display is still an early product. The user interface interface is experimental, is only visible in one eye, and there’s no developer ecosystem for third parties to build on. The Display is more of an incremental step toward than a full fledged platform.

But big companies are making big bets on this future. Meta took a $3.5B stake in Essilor Luxottica, Ray-Ban’s parent company, while Google invested $150M in Warby Parker to accelerate its own AI glasses. These aren’t little side projects. They’re early indicators that a new platform race is underway.

The Real Hurdles

Technical specs still matter — battery, weight, style — anything we wear on our face has to look good, stay cool, and last long enough to be useful. But large-scale adoption may depend even more on privacy. Glasses that constantly see your every move and hear every word raise questions for both the wearer and everyone nearby. Who controls the data? How visible are the signals when the device is active? Without clear answers, the technology risks being technically capable but socially rejected.

The Road Ahead

So, will AI-enabled glasses replace smartphones? Not yet. For now, they’re products for enthusiasts. But the trajectory is clear: if computing shifts toward devices that see and understand the world with us, the glasses form factor may take the lead.

We unpacked this more fully — including what would make glasses a true platform — in this week’s blog.

👉 Read the full post →

– Doug Griffin, Spatial Capital

Hi All,

First held in 1974 with about 500 participants, SIGGRAPH grew into a conference that at its peak drew nearly 50,000 people. Its influence runs deep: architecture, engineering, gaming, and visual effects—industries worth trillions today—trace key breakthroughs back to research first presented here.

I’ve been attending since 1996 and have presented several times. SIGGRAPH has always been the place where research makes the jump into the real-world. This year, much of the energy was around how graphics is fueling a new wave of robotics, autonomy, and what many now call Physical AI.

Here’s what stood out at the conference:

🧠 NVIDIA’s Bet on Physical AI

NVIDIA’s special session underscored how central synthetic data and simulation have become to training real-world intelligent systems. The company stressed that progress will depend on scalable platforms and tools, highlighting its own environments like Omniverse and Isaac Sim. It also announced new work in neural rendering—evidence that even the largest players see Physical AI as a long-term growth market.

The takeaway is hard to miss: scaling Physical AI will require an ecosystem of environments, tools, and workflows. NVIDIA is putting down some of the foundation, but the space is far bigger than any single company—leaving significant room for new approaches and new teams.

NVIDIA Omniverse for autonomy and robotic simulation (Image: NVIDIA)

Did someone forward this to you? Want more? 

🤖 Disney: Simulating Personality With Physics

If NVIDIA showed the scale of the infrastructure being built, Disney showed how those same ideas can delight people in practice. At SIGGRAPH, the company presented how its park droids were designed to move with personality—animations inspired by ducklings, refined in simulation with reinforcement learning. The result was memorable: small robots that feel alive to the people who see them.

But the showcase also revealed the limits of today’s systems. These droids aren’t autonomous; they rely on remote control and operate only in carefully managed park environments. It’s a reminder that while the technology is advancing, there’s still a wide gap between controlled demonstrations and robots that can handle everyday life. That gap is exactly where new ideas—and new companies—will emerge.

A behind the scenes look at the development of the BTX Droid (Image: Disney Research)

🏞 CAST: Reconstructing Worlds from a Single Image

While Disney’s work showed how robots can move with personality, CAST tackled how to create the data they need to learn. Presented by researchers at CMU, it’s a reminder that many of the most important advances start in universities.

CAST takes a single photo and reconstructs a physically consistent 3D environment—an approach very different from NVIDIA’s physics-based simulations. It highlights that there’s no single path to building the synthetic data these systems require. As the demand for training environments grows, we should expect a range of novel approaches, each creating new opportunities for innovation and company formation.

🌀 Genie 3: Generating Worlds Through Inference

CAST focused on reconstructing environments from a single photo; Genie pushed further, creating interactive worlds from a single image. Released just before SIGGRAPH by DeepMind, it shows how companies are exploring inference-based methods as an alternative to physics-driven simulation.

Genie takes a single image and turns it into a navigable space, constructed by inferring likely outcomes from prior data rather than simulating physics. It’s a fundamentally different approach, and one that points to a broader trend: synthetic data won’t come from one method alone. As the market grows, inference-based systems like Genie could complement physics-based ones, offering new ways to generate the scale and variety future AI systems will need.

For more insights, read our blog →

Final Thoughts

Physical AI is a fast-emerging market with broad potential. SIGGRAPH underscored how graphics research continues to provide the building blocks for this shift. While large players and academics are advancing the field, there’s still meaningful room for startups to translate those advances into real systems—and we’re keeping a close eye on what’s next.

– Doug Griffin, Spatial Capital

Hi All,

I spent most of July off the grid—hiking the full John Muir Trail with my 17-year-old son. We covered 220 miles through granite passes and past alpine lakes. After that much time immersed in real terrain, it’s hard not to reflect on how far we are from truly modeling the physical world—and how quickly AI is trying to catch up.

Doug and his son catching the sunrise on Mt Whitney.

This month’s updates trace that arc—of how AI is learning to See, Map, and Build. From planetary-scale models, to infrastructure intelligence, to design-to-manufacturing automation, there’s real momentum toward systems that better understand and assist our physical world. Let’s take a look

🌍 DeepMind: A Foundation Model to See the Planet

DeepMind unveiled AlphaEarth Foundations, a new AI system trained on decades of climate and remote sensing data. It produces high-resolution environmental maps that can model droughts, predict land use change, and support conservation or agricultural planning.

For researchers, policymakers, and operators, AlphaEarth represents a new kind of tool: a generalizable model that can interpret complex geospatial patterns—and help link environmental perception to real-world decision-making. It’s a strong signal that foundation models are moving beyond language and images, and into physical systems.

Separately, the Google Earth Engine team released a public satellite embedding dataset—another sign of growing momentum in AI-powered Earth observation.

Did someone forward this to you? Want more? 

🏗 Looq: Mapping Physical Infrastructure

While DeepMind builds planetary-scale models from orbit, Looq.ai—founded by astrophysicist-turned-roboticist Dominique Meyer—is focused on critical real-world infrastructure—what cities depend on but rarely see clearly.

After working on camera systems for near-Earth object detection, Meyer turned his attention to urban  infrastructure. He founded Looq AI to combine LiDAR, computer vision, and AI to scan and map utility poles and related urban assets—making our aging infrastructure as machine-readable as the road ahead.

This kind of perception is critical to maintaining and upgrading the infrastructure cities depend on—especially as it ages, expands, and becomes more complex to manage.

Read more about Looq AI’s founding story

🔧 Drafter: Turning Design into Built Reality

Seeing the world is one thing—building for it is another.

Manufacturing drawings are a necessary—but often painful—step between CAD and production. Drawings convey critical tolerances and specs essential for bridging design intent and real-world fabrication. Yet few engineers are trained in producing this documentation accurately or efficiently.

To solve this, Drafter has integrated its solution directly into Solidworks, a leading modeling program for mechanical design. Drafter automatically generates annotated part and assembly documentation, ready for manufacturing. It’s a big step toward streamlining design-to-fabrication workflows in industries where documentation is still mostly manual.

Market & Events

📈 Computer Vision Is Gaining Ground A recent forecast put the computer vision market at $34.3B by 2033, driven by AI advances and physical-world automation. It’s a broad category—but the most interesting developments, to us, are where perception meets action.

📍Exploring Vision and Robotics at SIGGRAPH Doug will be in Vancouver next week for SIGGRAPH—tracking work at the intersection of computer vision, simulation, and robotics. Let him know if you’ll be there.

📍Joining the Conversation at Step SF Later this month, Doug will join Step SF—a new event focused on startups and venture capital in San Francisco.

Signing Off

We learn a lot from the physical world. The hard part is teaching machines to do the same— in ways that move the world forward and create value along the way.

More soon,
– Doug and the team at Spatial Capital

Hi folks,

Robots are getting better at reading the room—and the real world.

June saw two big releases that push physical AI from lab curiosity to field-capable software: Meta’s V-JEPA 2 and Google’s Gemini Robotics On-Device. Different goals, same theme—training machines to think and act in the real world.

Meta: Making Machines Imagine

Meta’s V-JEPA 2 is a world model. This sounds philosophical, but here it just means: it can watch the world and make predictions about what happens next.

Trained on over a million hours of video, it lets robots reason through new environments, anticipate outcomes, and plan actions without needing environment-specific training.

V-JEPA 2 both recognizes what it sees and imagines what might happen next. If that sounds like a toddler who’s watched too many surveillance videos, well, that’s not entirely wrong.

Did someone forward this to you? Want more? 

Google: Getting Hands Dirty

While Meta trains minds, Google is making hands move.

Gemini Robotics On-Device is a compact model that runs on the robot, not in the cloud. It lets robots fold, unzip, pour, and generally handle real-world messiness—with minimal lag and maximum independence.

Trained for dual-arm systems, Gemini can adapt to new tasks with just 50–100 examples. And it’s not picky: it works across robot types, from ALOHA to humanoid platforms like Apptronik’s Apollo.

Brains, Biceps, and the New Stack

Taken together, these two launches sketch a new kind of robotics stack:

• Meta trains foresight. Robots imagine possible futures and chart a course.

• Google handles follow-through. Robots execute—directly, locally, and with dexterity.

Robotics is beginning to look more like scalable infrastructure than sci-fi.

From Fixed Routines to Flexible Robots

These models let robots reason through chaos, adapt to unfamiliar tools, and act with minimal supervision. That’s the heart of physical AI—learning how to operate in a world that rarely repeats itself.

Early days still—but these models offer a clear template for perception, planning, and physical action.

Best,
Spatial

Hey Everyone,

We just got back from AWE—the world’s largest AR conference—where the theme was unmistakable: AI is starting to move beyond the screen and into the physical world.

That same shift showed up at Apple’s WWDC and Google’s launch at InfoComm. Across a variety of platforms, computing is becoming situational—shaped by space, context, and environment.

Physical AI

Most of what we call “AI” today is still glued to a two-dimensional screen. It lives in chatbots, code editors, and web-based copilots—generating helpful text that feels more or less like technology always has. But computing is starting to migrate. Out of the browser. Off the desktop. Into real space.

This shift is already playing out across some of the most valuable platforms in tech. Apple, Google, Meta, and Niantic are pushing their interfaces outward—into glasses, rooms, objects, and shared environments. They're building toward what many are starting to call “Physical AI”.

Physical AI understands where and what the user is doing—then responds accordingly. It’s the beginning of something that feels very different from a chatbot. It’s situational computing.

Apple: UX that Anchors to Your Room

Apple’s recent visionOS update introduced spatial widgets, collaborative 3D experiences, photoreal “personas,” and spatial images that respond to your movement. It’s an evolution of the UI into actual space. Imagine browsers that sit on your kitchen table, or widgets that stay stuck to your wall instead of floating in a virtual void.

Apple calendar widget embedded in a wall

Niantic: Goodbye Games, Hello Map of Everything

Niantic—yes, the Pokémon Go company—quietly did something huge. It sold its entire gaming division for $3.5B and launched a new company: Niantic Spatial. They raised $250M to map the real world as a shared, AI-readable environment.

At AWE, they announced a partnership with Snap to start building a semantic 3D map of the physical world. It’s a global layer for spatial computing that can incorporate fun and games as well as infrastructure.

Niantic and Snap pair up to visualize the world

Meta, on the other hand, is chasing the platform. With Llama (AI), Orion (AR), and a pipeline of smart glasses, they’re building a vertically integrated stack: language, vision, interface, and device. The goal is to transition from being an app company to the operating system and platform. 

Their second-gen Ray-Ban smart glasses already offer AI-powered image capture, voice interaction, and audio streaming—all without a screen. Next up: collaborations with Oakley and Prada. The goal is to make physical AI wearable, casual, and socially acceptable.

Meta’s goal is to be the new spatial platform

Google: High-Fidelity Presence, No Goggles Required

Google is playing the long game on spatial interaction. Last year they announced Android XR, a foundational OS for spatial devices. This year, they unveiled Google Beam (formerly Project Starline), launching soon in partnership with HP and Zoom.

Beam uses AI, depth sensors, and light-field displays to deliver photorealistic, glasses-free 3D video calls. It’s spatial computing as communication—compressing space, adding nuance, and letting people feel like they’re in the same room without putting on hardware.

Did someone forward this to you? Want more? 

The Interface Is Moving

What ties these stories together is the proximity of the compute to the physical world. Interfaces are moving closer to the environments where decisions get made: job sites, operating rooms, factory floors, or your own kitchen. The interface is becoming situational. And that’s where Physical AI shows up.

Instead of delivering knowledge through fixed screens, these systems adapt to their surroundings. They observe, interpret, and offer assistance within the flow of real-world activity—often without asking for direct input or attention.

Physical AI, Real-World Stakes

Physical AI could become the bridge between information and action in everyday life. Not as a screen to consult, but as a layer of quiet intelligence that supports decision-making in the moment: on the job, in motion, or mid-task.

Of course, a world where AI understands physical context raises real questions about trust and privacy. Systems that process location, audio, and ambient data need clear boundaries—not just for user comfort, but for safety and accountability. It’s critical that we  build infrastructure that respects consent, limits data exposure, and makes its behavior legible by default.

Stay situated,

Spatial Team

As headlines focus on tariffs and domestic manufacturing efforts, the conversation has turned to reviving industry, rebuilding supply chains, and protecting jobs.

But we’re having the wrong conversation.

Tariffs aim to support a 1950s-style economy—one built on labor-intensive manufacturing and domestic supply chains. But the world that created that economy no longer exists. We should be planning for something far more transformative.

The next 10 to 20 years will bring changes on par with the last 300 years. It’s a staggering rate of change.

In 1700, there were no automobiles, no electricity, no factory lines, no national banks, no stock markets as we know them. Most people farmed. The idea of a “job” in the modern sense didn’t exist.

Nearly every system we rely on today—transportation, finance, healthcare, logistics, education—was the product of 300 years of compounding technology.

Now, we’re about to run that same curve again—compressed into a single decade. And our systems aren’t ready.

Now, we’re about to run that same curve again—only faster. And our systems aren’t ready.

This might sound dramatic—but when viewed on a macro scale, it’s inevitable.

Our economic playbooks were written for an era where labor was the primary input. Where industries scaled by hiring more people. Where public policy focused on full employment and slow, steady productivity growth.

That’s not the world we’re heading into.

The last major transformation took centuries. The next one will unfold within our lifetimes.

What Venture Sees That Others May Miss

Venture capital doesn’t operate on quarterly timelines. We place bets on what the world will look like a decade from now—and who’s already building toward it.

• At the early stage, that means getting unusually close to the edge.

• It means reading research before it spins out of the lab.

• Meeting founders when the idea is still a sketch.

• Reviewing thousands of startups.

• Spotting patterns in pitch decks—months or years before they show up in headlines.

• Watching concepts become prototypes, and prototypes become companies.

From that vantage point, a clear pattern emerges: 

Founders aren’t focused on restoring legacy systems or defending the status quo. They’re building around a different set of assumptions—where automation is the baseline, not the breakthrough. Where work is distributed, not centralized. And where digital systems increasingly operate without human mediation.

They’re building AI agents that can take on entire functions once considered too complex to automate. Vision systems that allow machines to perceive, interpret, and act in the physical world. And robotics platforms that operate in warehouses, farms, and construction sites—bridging software and hardware in real-world environments.

Did someone forward this to you? Want more? 

The Change Has Already Begun

AI today helps us write emails, generate code, and summarize calls. But that’s just the starting point. It’s already replacing individual tasks and workflows. Soon, it will replace entire operational stacks. From support desks to strategic planning, AI is transitioning from assistant to operator.

This isn’t speculative. Some of these products are in-market now. Many more will be within a year.

Over the next decade, we’ll see a profound shift in how white-collar work is executed. Functions long considered immune to automation will be rebuilt from the ground up.

Meanwhile, robotics is reshaping the physical economy. Prices are falling. Capabilities are improving. What once required a $500K robot and a team of PhDs can now be done by a $50K system with off-the-shelf sensors and a trained foundation model.

Factories, farms, fulfillment centers, construction sites, commercial kitchens—these are already being transformed by scalable, low-cost automation.

We’re approaching a world where AI agents analyze and design, and machines build and ship.

The labor model that powered the last century—built on human scale—no longer applies.

Labor is shifting from task execution to systems orchestration and oversight.

The Labor Model Is Changing

Much of the current conversation assumes we can bring jobs back by reshoring manufacturing. That tariffs will protect workers, revive industry, restore the balance.

But that vision doesn’t align with today’s reality.

Manufacturing may come back. But the jobs likely won’t—at least not in the form most people expect. A 1950s factory was bustling with workers on the line. A modern factory is quiet, dark, and filled with machines. Where once 5,000 people operated the floor, today 10 specialists oversee automated systems.

Take Figure AI, a leading robotics company. They’ve announced plans to ship up to 100,000 humanoid robots over the next four years.

This isn’t a prediction. It’s a deployment.

Automation isn’t theoretical anymore—it’s practical, cheaper, and increasingly superior.

We’re not heading back to the post-war 1950s economy. We’re entering a new system where labor is no longer the primary input. Where productivity scales without people.

We don’t need a tariff strategy. We need a strategy for an economy where jobs are no longer the organizing principle.

Modern factories operate with minimal labor—powered by scalable, increasingly autonomous systems.

Preparing for a Post-Labor Economy

If we want to talk policy, we should be talking about:

• How to reskill millions of people as automation reshapes traditional career paths

• How to rethink education, not just for digital literacy—but for adaptability, creativity, and systems thinking

• How to design infrastructure, healthcare, and governance for a population that will increasingly not work in the way we once defined work

• How to retool economic metrics that were built around labor participation and wage growth

We need to focus on economic policy to best support the world we’re heading toward.

Designing new systems—across both policy and infrastructure—requires foresight, not nostalgia.

The Next Economy Will Scale Differently

The economic engine we’ve relied on—one powered by full employment, career ladders, and steady wage growth—is winding down.

The next economy won’t scale by adding people. It will scale by expanding capability—across systems, tools, and organizations.

That means:

• AI systems that handle analysis, coordination, and decision-making across thousands of tasks.

• Robotic platforms that perform physical labor with precision, consistency, and no downtime.

• Modular organizations that operate with fewer people, powered by autonomous systems and dynamic software.

This reflects a fundamental shift in how value is created—and how quickly it can compound.

If we want to build a thriving future, we need to design around that shift. Because it’s already underway.

The Future Is Already Being Built

This transformation is already underway. But the planning hasn’t kept up. Our systems and frameworks are still rooted in a labor-first paradigm.

But labor is no longer the constraint. Imagination is.

The blueprint for the next economy is already being drafted—in code, in automation systems, in the quiet routines of factories, hospitals, and classrooms.

We can’t plan for 2035 using the mindset of 1950.

We can’t build what’s next by trying to preserve what used to work.

If we want a future that’s prosperous and widely shared, we need to design for it—intentionally, urgently, and with a longer view.

A thriving future won’t just be automated—it will be designed for human flourishing.

Doug Griffin
General Partner, Spatial Capital

The last major computing platform shift—the smartphone—didn’t just change how we interact with technology. It created entirely new industries, reshaped trillion-dollar markets, and minted a new generation of billionaires. Many of today’s largest companies didn’t exist before mobile computing took off.

So what’s next?

If Meta is right, AI-powered, spatially aware wearables could be the next major platform. And if that’s the case, we’re at the very beginning of another massive transition.

Last week, Meta announced Aria Gen 2, the second generation of its AI-powered research glasses. The device features on-device AI processing, advanced sensors, and real-time spatial awareness. It’s not aimed at consumers, but rather, aimed at academics and industry partners exploring machine perception and contextual AI.

Meta Aria Gen 2, courtesy of Meta

This wasn’t a flashy product launch. It wasn’t widely covered in the news. But it might signal one of the biggest shifts in computing since the smartphone.

Officially, Aria Gen 2 is for academia and industry partners. But let’s be honest—Meta isn’t pouring resources into this because it’s a fun research project. They’re building the foundation for what could be the next dominant computing platform.

AI That Understands the World

The core of Aria Gen 2 is machine perception. Technologies like SLAM (Simultaneous Localization and Mapping), eye tracking, hand tracking, and spatial audio turn its passive cameras into an AI system that processes the world in real time. Aria’s AI can understand the physical context of daily life, making interactions more intuitive and responsive.

Did someone forward this to you? Want more? 

Real Time Privacy

A system like this could easily veer into dystopian territory, particularly if your surroundings are transferred into the cloud—where they can be analyzed, stored, or even misused. Meta knows that, which is why real-time AI elements actively block personalized information before it’s stored. 

Meta’s approach here is admirable. Real-time filtering of personal information—blurring faces, removing license plates, distinguishing voices—shows they’re thinking about privacy from the start. It’s a smart move, and it makes us far more optimistic about a future where perceptual AI enhances our world without compromising it.

Example of real time face obscuring, courtesy of Meta

The Use Cases

Once real-time, spatially aware AI exists in a consumer device, millions of new use cases will emerge—just as they did when the smartphone became mainstream. Many will be hard to predict, just as few could have imagined Uber or Snapchat before mobile GPS and high-quality cameras became ubiquitous.

Here are a few obvious examples where real-time, context-aware AI could provide seamless, immediate assistance without pulling out a device:

• Social Memory: A subtle AI reminder of past conversations—“This is Jordan from Acme Robotics. You met last year at CES. She just raised a Series A.” 

• Emergency Response: A paramedic wearing AI glasses gets triage guidance, with AI relaying critical patient details—like blood pressure, medication history, or response to treatment.

• Assistance for Flight Attendants: A flight attendant receives real-time updates on passenger needs, pulling from the flight itinerary to display names and pre-submitted requests—like meal preferences or accessibility requirements.

• Outdoor Navigation: A mountain guide gets AI-driven environmental insights, identifying avalanche risks and suggesting safer routes based on terrain conditions.

Meta showcased another use case—leveraging SLAM and spatial audio to provide real-time navigation for blind and low-vision users. (see below)

Major Tech Players Are Laying the Groundwork

Meta isn’t alone in this effort.

Google is investing heavily in becoming the OS for AI-powered wearables. See our recent blog on their work on Gemini 2 and Android XR.

Apple has also been building toward spatial awareness in its devices. While its current focus is on mixed reality, their next move could bring AI-powered AR into the mainstream. Apple’s 4nm processors are already pushing the limits of on-device compute—a key ingredient for AI wearables.

Qualcomm is another critical player. Their Snapdragon chips already power AI-driven mobile devices, and their Snapdragon XR series is well suited for wearable AI. As AI moves on-device, Qualcomm’s role in low-power, high-performance AI processing will only grow.

The takeaway? Major tech players aren’t just experimenting—they’re making a long-term bet on AI-first perceptual computing. Billions are being poured into this shift, and history suggests that when this much capital and talent converge, something fundamental is about to change.

So, What’s Next?

The first iPhone launched with no third-party apps, no ecosystem, and limited functionality—and completely changed the world. If AI-powered wearables become the next dominant platform, we’re watching the earliest stages of a new industry take shape.

The next wave of applications, services, and startups? They don’t exist yet. But soon, they will.

And that’s why this matters.

Cheers,

The Spatial Team

Happy (Belated) New Year from Spatial Capital! 🎉

Yes, we know—late January isn’t exactly the frontlines of New Year’s greetings. But sometimes it's ok to arrive fashionably late. Our excuse? We kicked off 2025 with a bang, an exciting new investment we made this week (details coming soon!).

We hoped to get our 2025 predictions out before most people abandoned their New Year’s resolutions. But then we realized: all good things come to those who wait. Still, we couldn’t help but share a glimpse of 2025 trends along with the places we’ll be rubbing elbows with innovators this year.

Trends The Past Few Years

Every year brings transformative advancements in AI. Recent years laid the groundwork:

• 2022: NeRFs and Splats, novel ways to represent the real world in 3D, took visual computing to new heights. (Cool tech, funny acronyms. read more)

• 2023: LLMs went mainstream, tackling text tasks with ease. Many were caught off guard. Others recognized this was only the beginning.

• 2024: Generative AI hit its stride, speeding up workflows and sprinkling creativity across video and 3D content.

Predictions for 2025

2025 promises to be the year AI breaks free from its “text” roots and steps into the visual world. Here’s what we’re expecting, as well as some relevant blogs:

AI Will “See” The World: 2025 will be the year Large Vision and World models prove their worth. These systems started making waves in 2024, but this year they'll move beyond labs and demos into real-world applications. Expect systems that understand context, navigate spaces, and power autonomy in ways we've only imagined.

Synthetic Data Will Unlock Vision: To make vision models adaptable and reliable, synthetic data will become indispensable. Scenarios that are rare, risky, or just plain impractical—like navigating icy roads or hauling ore out of underground mines—will now be trainable. This unlocks autonomy across unique industries like agriculture, construction, and mining.

Small Models Will Expand Opportunities: Smaller AI systems may start to steal the spotlight this year (actually this is more likely in 2026-27, but we’re writing it now anyway). While massive models grab headlines, the real opportunity lies with compact models running on everyday devices. Google's work with AndroidXR is just the beginning—AI in AR is getting poised for its breakout moment.

Agents Will Extend Our Ability: Finally, AI agents are ready to get real work done. From scheduling meetings to conducting research, this is the year they stop being theoretical. We haven’t written a blog about this just yet... but perhaps we should. Inspiration is brewing for a future newsletter. 😅

Conferences: Where We’ll Be in 2025

We’re not the ones building these innovations—we’re just here to fuel the sparks of genius. But first, we need to find these rocket builders. 🚀 Here’s where we’ll be touching down in 2025:

• GeoWeek (February, Denver): Michel takes the stage to showcase how geospatial tech reshapes the built world.

• GDC (March, San Francisco): We'll explore where gaming tech meets non-gaming innovation. We’re hoping to host a gathering.

• GTC (March, San Jose): A hub for the latest in AI and computer vision. Yes, it inconveniently shares its dates (and most of its name) with GDC!

• CVPR (June, Nashville - Tentative): The latest in cutting edge computer vision research. Where PhDs present.

• SIGGRAPH (August, Vancouver): Known for its iconic 3D teapots, this conference has evolved into a premier destination for AI and graphics—with Jensen Huang leaving his mark last year.

• Slush (November, Finland - Tentative): The coolest conference (both figuratively and temperature-wise) that we've never been to. And the home of one of our latest investments, Ray!

Events - Let’s Connect

We love hosting events. Our first kicks off this Friday in Mill Valley! More are on the horizon, so keep an eye out! And if you’ll be at any of the conferences we mentioned, let’s grab a coffee (or something stronger).

Thanks for sticking with us this far—especially given how late we are with this email. Maybe it’s your New Year’s resolution to read all the way to the end of newsletters? If so, you’ve absolutely crushed it!

Cheers,
Doug & the Spatial Team

Can AI finally understand and navigate the real world? At CES this week, NVIDIA introduced Cosmos, a "world foundation model" designed to help AI make sense of the unpredictable physical world we call home.

What is Cosmos? 

It’s a generative model built to simulate and master the madness of our messy, marvelous planet. By synthesizing environments with object permanence, physics, and real-world dynamics, Cosmos equips AI to navigate, predict, and interact effectively with its surroundings. As Jensen Huang quipped, "Cosmos is for the physical world what ChatGPT is for words and text."

Cosmos is trained on 20 million hours of video (yes, hours). If you’re wondering, that’s like starting a video feed in Ancient Rome and letting it roll through gladiators, guillotines, and global revolutions until today.

Did someone forward this to you? Want more? 

Cosmos Core Use Cases

But Cosmos isn’t just binge-watching history. It’s generating synthetic data to train AI systems for the messy, unpredictable world we live in. Let’s face it—relying on real-world data alone is like trying to teach a toddler to walk in a crowded mall: chaotic, slow, and often impractical.

• Need to train autonomous bulldozers to handle a mining landslide or falling trees? Cosmos can simulate it.

• Preparing disaster response vehicles for wildfire containment or earthquake rescues? Cosmos fakes disasters beautifully.

• Missing that ladder-falling-off-a-truck scenario for your autonomous vehicle model? Cosmos has you covered, no drama required.

Autonomous bulldozer navigating harsh environments.

Who Can Use It

And here’s the kicker: Cosmos is open-source. It’s free. It’s tweakable. It’s brilliant. In a remarkable feat only seen by the boldest of companies, NVIDIA is unleashing this model into the wild to spark the next wave of startups and groundbreaking use cases. Think LLMs, but for the physical world: practical, powerful, and potentially game-changing.

Other Use Cases

Cosmos isn’t confined to robotics or autonomous systems. Its flexibility opens doors to innovative applications. For instance, urban planning teams could simulate traffic scenarios in bustling cities, helping design better roads and intersections. Environmental researchers might use it to model deforestation impacts or predict flood responses in real time. With Cosmos, the possibilities are as vast as the challenges we face. We’ll likely see a wave of new startups and groundbreaking solutions emerge, powered by Cosmos’ versatile capabilities.

In Closing

Cosmos might just be the foundation AI needs to finally make sense of our messy, chaotic, and occasionally ridiculous world—and maybe help us clean it up while it’s at it.

Cheers,

Doug & The Spatial Team

Hey there,

Will AI’s future form factor be AR glasses? Google seems to think so.

For years, we’ve been slaves to our screens—heads down, thumbs tapping, and necks craning like flamingos in a tech-fueled flock. Last week, Google unveiled two groundbreaking developments that could transform how we interact with technology: Gemini 2.0 and Android XR.

If you thought AI was just about chat windows and endless autocomplete fails, think again. This is AI with eyes, ears, and spatial smarts—and maybe a new role as your pocket (err, glasses-mounted) assistant.

The Dynamic Duo: Gemini 2.0 and Android XR

Gemini 2.0: Google’s newest AI marvel does more than just chat—it can see, reason, and remember. Need to translate a sign or identify a plant? Gemini’s got you covered. Its smaller "Flash" variant is even faster and smarter, proving good things come in small packages.

Android XR: An operating system designed for AR and XR devices. As the first new OS Google has released in over a decade, the team at Google is building the foundation of a rich ecosystem of apps and devices to interact with the real world. Need a nudge to tackle that dreaded 'honey-do list'? Android XR, powered by Gemini, can swoop in and guide you with laser-sharp precision.

Why It Matters - A Whole New Ecosystem

Google’s play isn’t just about gadgets; it’s about an ecosystem of devices and apps that could one day be the successor to today's phone app stores. With support for OpenXR, WebXR, and Unity, Android XR is a playground for developers building everything from productivity tools to immersive entertainment.

But here’s the twist: building an ecosystem like this isn’t just technical—it’s cultural. Can Google convince developers, consumers, and manufacturers that AR glasses are more than a novelty? Will these tools spark a tech renaissance, or just become fossils in the gadget graveyard? And when will devices arrive that we actually want to wear?

Did someone forward this to you? Want more? 

The Big Picture, Seeing Beyond The Hype

Let’s not forget: Google isn’t the only one making big bets on AR/XR. Apple, Meta, and others are elbowing their way into the race, tossing money and moonshots at the space like they’re competing in a billionaire science fair. With Gemini’s visual smarts and Android XR’s developer focus, Google might just have the secret sauce: a rare mix of practical utility and moonshot ambition.

The Big Question: Will AR Glasses Go Mainstream?

The potential is huge, but so are the hurdles. Mass adoption of AR glasses will hinge on power, price, battery life, and whether they’re stylish enough to avoid becoming the next Segway of wearables. On top of that, developers need to create those 'must-have' apps.

But the concept is bold: an AI-powered assistant that sees and reasons about the world around you—unshackling us from our screens and letting us fully engage with the people and places in front of us.

One thing’s clear: Google isn’t afraid to bet big on a big vision—and they might just have the tools to pull it off. Are we looking at the next evolution of AI, or another overhyped experiment? Stay tuned.

Best,
The Spatial Capital Team

Hey there,

While much of today’s AI is busy reading, writing, and generating 2D art, there’s a whole new frontier quietly taking shape: spatial intelligence.

It’s not just about teaching machines to “see” the world (aka “perceptual computing”) or even interacting with it in 3D (“spatial computing”) - we defined both of these in a previous blog. Spatial intelligence takes it further—allowing us to understand visual context and act intelligently.

Imagine robots in a warehouse that don’t just move boxes but dynamically adapt to shifting tasks, or design tools that let you interact with your 3D ideas as if they were right in front of you. It’s about making tech feel as seamless and natural as working with a human.

Why Spatial Intelligence Matters

Spatial intelligence builds on three core abilities:
1️⃣ Perception – Seeing and sensing the world.
2️⃣ Understanding – Building context and meaning from what’s perceived.
3️⃣ Action – Taking smart, meaningful action based on that understanding.

Who’s Leading the Charge?

World Labs

Fei-Fei Li (aka the “Godmother of AI”) is leading a team that includes the co-inventor of NeRFs and other top-tier researchers to build a "Large World Model." The goal? Create tech that can perceive, generate, and interact with the physical world.

While their work is still under wraps, Fei-Fei outlined her vision in a TED Talk earlier this year. It’s an inspiring look at how AI can move beyond recognition to true understanding and meaningful interaction.

Niantic

Known for blending the real and digital worlds through AR (yup, they’re behind Pokémon GO), Niantic has been leading the way in real-world integration for over a decade. Their tech stack includes tools like Scaniverse for 3D scanning and VPS for anchoring digital objects in the physical world. Their new geospatial platform takes this expertise to new heights, enabling everything from AR-enhanced collaboration to immersive logistics solutions.

Courtesy of Niantic

Did someone forward this to you? Want more? 

Oh, and Have You Seen Blendbox Yet?

Two weeks ago, Blockade Labs dropped blendbox.ai.

Here’s why it’s worth checking out:

🖼️ It lets you generate images as controllable layers. Tweak a single part of the image without regenerating the whole thing.

🎨 And this week the team dropped more features. Now you can upload existing photos, remove backgrounds, and add layers for custom creations. (Perfect for whipping up AI-powered holiday cards in no time!)

Your Weekly AI Fix

Learn AI in 5 minutes a day

This is the easiest way for a busy person wanting to learn AI in as little time as possible:

1. Sign up for The Rundown AI newsletter

2. They send you 5-minute email updates on the latest AI news and how to use it

3. You learn how to become 2x more productive by leveraging AI

Sign up to start learning.

So, what’s more exciting to you—robots with spatial smarts or AI-generated holiday cards? Either way, we’d love to hear your thoughts. Hit reply and let us know!

Best,
The Spatial Capital Team

Hey there, AI Enthusiast!

As AI continues to evolve, we’re witnessing significant advancements in “perceptual commuting” - how computers “see” and “understand” the physical world. This update highlights several industry developments, including recent breakthroughs in robotics, mapping, surveying, and visual effects. Let’s explore the cutting edge of computer vision and AI.

Industry Spotlights

Enhancing Autonomous Systems in Warehouses

NVIDIA recently rolled out generative AI models for OpenUSD, supercharging the creation of 3D data for training autonomous systems such as robotic warehouse pickers. This text-to-3D solution is important for creating synthetic training data for robotics, ensuring safety and reliability. Even better, it's available now. Read more here.

Courtesy of NVIDIA

Innovations in Mapping with Gaussian Splats

Recent research has shown the potential of Gaussian splats for mapping. This particular paper showcases how splats can be used to create training data from the real world, replacing cars and other identifiable features from collected data. These types of techniques could advance how we visualize and navigate complex environments. Dive into the details.

Surveying Physical Infrastructure

Gaussian splats aren’t the only solution for understanding the real world. Looq.ai got some love from GeoWeek, which praised its secret sauce for surveying. Looq makes converting 2D images into a survey-accurate 3D point cloud look easy—but the details are anything but. Check out how they do it.

Courtesy of GeoWeek & LooqAI

Did someone forward this to you? Want more? 

Segmentation Precision for Film and VFX

Some video projects require removing unwanted elements, known as 'segmentation' in AI and 'rotoscoping' in VFX. Electric Sheep launched an upgrade to Spotlight, its professional-grade rotoscoping tool for film and VFX. To put it to the test, they compared it to Meta’s Segment Anything 2 (SAM2), the state-of-the-art segmentation tool. The verdict? SAM2 is faster, but Spotlight nails the details when it comes to quality. Curious about the complete comparison?

Wrap-Up

Looking ahead, the momentum in these fields shows no signs of slowing. We’re thrilled to support the trailblazers pushing the boundaries of possibility. Stay tuned for more thrilling updates as we navigate this ever-evolving landscape together.

Stay tuned for more insights in our next update!

Best,
The Spatial Capital Team