WEBVTT

1
00:00:00.100 --> 00:00:14.600
On y'all. Welcome back to Blockspace Live, presented by CleanSpark. Charlie, it's earnings season, baby. So for today's show, we are kicking it off with Core Scientific and Applied Digital's latest quarterly earnings.

2
00:00:14.660 --> 00:00:25.500
Core's dropped this morning, Applied's yesterday evening after market close. So we'll go over the numbers, including Core's new $15 billion deal with AMD and an undisclosed NeoCloud.

3
00:00:26.140 --> 00:00:39.280
Following that, we have a story on Galaxy expanding its potential data center portfolio with an acquisition in Texas, and then we will cover Meta and BlackRock going in on a data center as well.

4
00:00:39.340 --> 00:00:54.560
We will end the stream with none other than AnchorWatch CEO Rob Hamilton to talk about open weights, Anthropic's recent letter on the open weight controversy, plus Jensen Huang's and friends' letter in defense of open weights last week.

5
00:00:55.840 --> 00:01:11.220
That's right. Blockspace goes live every weekday at 1:00 p.m. Eastern, featuring quick hits on AI, data centers, emerging tech, and markets. We are compute's daily live show coming to you hot, middle of the day.

6
00:01:12.760 --> 00:01:20.220
If you like what you hear, you'll love the rest of our content. Blockspace.media is the website with everything we write, everything we produce.

7
00:01:20.400 --> 00:01:27.000
You can find all of the post-live stream breakdowns, the podcasts, through the website.

8
00:01:27.200 --> 00:01:40.840
And if you don't like audio content or written content online, you can get it just straight to your email, newsletter.blockspacemedia.com. Sign up at newsletter.blockspacemedia.com.

9
00:01:41.560 --> 00:01:57.870
This show is brought to you by CleanSpark, NASDAQ-listed ticker CLSK. More on CleanSpark later on in the show. Colin, let's go to the earnings corner. The inaugural earnings corner here at Blockspace Live.

10
00:01:57.930 --> 00:02:14.180
We will be covering earnings as they come out for the data center companies and Bitcoin miners in our coverage orbit. We will most likely package them every other day into multiple reports on the earnings.

11
00:02:14.220 --> 00:02:23.679
So for today, kicking it off with Core's and with Applied Digital. And Core's, Charlie. Oh, go ahead. Core's, can I, can I interject here? Core's snuck this one in.

12
00:02:24.080 --> 00:02:31.280
This, the Core's Q2, uh, earnings was not posted on their investor relations. It was, there was not a press release.

13
00:02:31.760 --> 00:02:45.080
It was, uh, it was a press release about the AMD, uh, deal that we'll, that you'll talk about, and then they're like, "Oh, and we're doing an earnings," so to, to me they've not- Yeah, they didn't announce their earnings call, which is- Yeah...

14
00:02:45.120 --> 00:02:54.340
very odd. You know, most of these companies do. A part of me wonders if they did that to kind of create more shock waves for the AMD deal, kind of catch people- Yeah, I mean-... off guard...

15
00:02:54.380 --> 00:03:06.500
it's like, it's like Beyoncé dropping an album. Sometimes it's better if it just happens and you don't anticipate it. [lip smacks] So quick headline numbers for Core's before we get into the AMD deal.

16
00:03:06.560 --> 00:03:15.720
This is their Q2 earnings. Total revenue up from 115.2 million in Q1 to 164.2 million in Q2.

17
00:03:16.340 --> 00:03:34.220
Colocation revenue, their, their, their AI business line, up to 136.7 million, from 17-- from 707-- oh, uh, excuse me, from 77.5 million in Q1 and 10.6 million in Q2 2025.

18
00:03:34.320 --> 00:03:47.800
Clearly has slingshotted into the dominant business. They actually, the margins on their Bitcoin mining business were -56% gross margins as they continue to phase that out and hash price is in the gutter.

19
00:03:47.810 --> 00:04:02.320
Adjusted EBITDA came in at 41.1 million versus 8.9 million in Q1. Net loss was 1.155 billion, driven basically entirely by a 1.05 billion writedown.

20
00:04:02.400 --> 00:04:17.260
No cash, uh, charge for the change in fair value of warrants, stock appreciation, not their operations. CapEx more than doubled from Q1, which was 389.2 million to 797.5 million.

21
00:04:17.529 --> 00:04:33.170
Notes on that later, specifically how labor costs have ballooned since they signed their Watershed CoreWeave deal back in 2024. Liquidity stood at 1.82 billion, cash plus Bitcoin, up sharply from 1 billion in Q1.

22
00:04:33.600 --> 00:04:50.270
Now, the big piece of today's, uh, earnings call though, the thing that people glommed onto most was this AMD partnership, and it's pretty interesting, Charlie. So the TLDR of the partnership, [lip smacks]

23
00:04:50.280 --> 00:05:09.860
AMD has signed a 15-year deal for 377 megawatts, and a NeoCloud has signed a similar deal for 152 megawatts, for a total consideration of just over $14 billion in revenue over the total contractual period for the deals.

24
00:05:10.400 --> 00:05:20.840
There are two five-year extension options with these deals, and AMD holds a reservation right for an additional 1,925 megawatts.

25
00:05:20.980 --> 00:05:35.640
Yes, that is 1.9 gigawatts of critical IT capacity, exercisable, quote, "at certain times and under certain circumstances," end quote, through December 28th, 2028. So Core Scientific just booked

26
00:05:37.260 --> 00:05:54.120
530 megawatts across five sites. The total potential expansion, though, is 2.5 gigawatts of leased capacity. The initial delivery for these leases is early 2027,

27
00:05:55.060 --> 00:06:06.140
with the full leases to be realized by the end of 2028. Now, here's the interesting part. Again, this is a two-part deal here.

28
00:06:07.128 --> 00:06:17.238
And AMD is securing three hundred and seventy-seven megawatts, but an undisclosed NeoCloud is securing a hundred and fifty-two megawatts. And AMD has agreed to backstop that agreement.

29
00:06:17.988 --> 00:06:20.868
Similar to what we've seen from Nvidia in the past,

30
00:06:21.818 --> 00:06:33.808
they are going to their customer, and they are ensuring financial certainty for that project with these backstops in yet another instance of circular financing that has come under fire from analysts in the space.

31
00:06:34.968 --> 00:06:49.028
A few things to note here, Charlie, before I toss it to you for second thoughts, and then we move on to Applied. The construction costs have ballooned, my guy. According to CFO Jimmy, uh, Jim

32
00:06:50.028 --> 00:07:03.528
Ningard, re- uh, he reiterated that the math now stands at eleven to twelve million dollars per megawatt cost estimate for the initial five hundred and thirty megawatt AMD and NeoCloud commitment.

33
00:07:04.538 --> 00:07:16.828
That's roughly six billion dollars in capital that they plan to finance through project-level bonds. COO Matt Brown noted on the earnings call, because some analysts ask, "Whoa, why have costs gone up so much?"

34
00:07:16.868 --> 00:07:25.968
'Cause their CoreWeave build- builds were significantly less than this. Matt Brown said, quote, "The thing that's primarily driving cost of construction today is labor.

35
00:07:26.108 --> 00:07:37.908
Labor is very scarce in a number of markets across the USA, and a lot of general contractors, a lot of electrical subcontractors, the large ones, and a lot of the mechanical trades are just pretty saturated with work right now.

36
00:07:38.548 --> 00:07:46.098
What you're seeing is that labor constraint is naturally sort of driving up the cost of labor in some of the very, very competitive geographies across the United States."

37
00:07:46.568 --> 00:08:00.188
We've actually covered this, Charlie, because it's one of the, I think, lesser observed aspects of this entire AI CapEx boom. People talk about the financing. They talk about the deals.

38
00:08:00.208 --> 00:08:04.428
But no one really talks about the raw materials or the companies that are going to erect these data centers.

39
00:08:04.448 --> 00:08:13.808
And the fact of the matter is, the number of companies that have proven expertise in building these massive data centers, it's pretty, it's a pretty small pool.

40
00:08:14.168 --> 00:08:26.208
And so you're probably seeing some pretty crazy bidding wars for this labor. And as a result, the CapEx costs have surged significantly since CoreWeave signed its first deal with Core Scientific in twenty twenty-four.

41
00:08:26.928 --> 00:08:28.608
A, a few other notes here, Charlie.

42
00:08:29.768 --> 00:08:44.008
AMD received warrants up to, uh, may receive warrants up to thirty million shares at a twenty-three dollar forty-seven cent strike price, vesting at twelve thousand two hundred and twenty-two shares per megawatt of critical IT load under the lease.

43
00:08:44.748 --> 00:08:55.208
Since the five hundred, roughly five hundred and thirty megawatts was signed on day one, they have six point five million shares vested immediately. So AMD has a stake in Core Scientific as a part of this deal.

44
00:08:56.508 --> 00:09:08.268
Last few bits. Total leased customer power capacity is now at one point one gigawatts, with twenty-four billion in potential contracted revenue. They are billing four hundred and thirty-seven megawatts as of mid-July.

45
00:09:08.328 --> 00:09:16.588
That counts out to six hundred and thirty-five million annualized co-location revenue runway. [lip smack] Uh, their balance sheet has improved.

46
00:09:17.028 --> 00:09:23.488
It's moved from three hundred and eleven million in December twenty twenty-five to one point seven seven billion as of the end of this most recent quarter.

47
00:09:23.628 --> 00:09:34.608
Debt jumped to four point three billion, up from one point zero six. And again, the legacy mining segment is really not a consideration at this point. It's a very small piece of their revenue. It's down.

48
00:09:34.688 --> 00:09:48.128
It had negative fifty-six percent gross margins as of the end of last quarter. Charlie, what you got? So the, yeah, earnings and the finance, a little bit, uh, beyond me.

49
00:09:48.308 --> 00:09:57.828
But with some deft talking to the DI metrics MCP server and kinda scoping what's interesting about this deal.

50
00:09:59.088 --> 00:10:06.848
Um, this deal kind of reprices the Core Scientific book.

51
00:10:07.428 --> 00:10:21.248
So the other Core Sci lease-- or I'm sorry, the other CoreWeave leases are at around, uh, a hundred and twenty, low hundred and twenty dollars per kilowatt, uh, per month on the critical IT. However, this AMD deal

52
00:10:22.188 --> 00:10:33.468
at fourteen billion over five hundred and thirty megawatts in fifteen years implies a hundred and forty-six dollars per kilowatt per month, which is, like, twenty percent somewhat higher.

53
00:10:34.328 --> 00:10:43.658
And, uh, if we compare this to other NeoClouds, you know, the, uh, NeoClouds ten- NeoCloud tenants typically pay, like, a ten to fifteen percent premium, at least as far as the deals we're aware of,

54
00:10:44.708 --> 00:10:53.368
um, over, like, hyperscalers 'cause they're paying for the credit risk. But, uh, AMD kinda breaks that a little bit.

55
00:10:53.468 --> 00:11:08.448
Um, Core's got kind of like a near NeoCloud rate, uh, from an investment-grade counterparty, so pretty interesting. My other comment on this is that, um, these milestones are pretty crucial.

56
00:11:08.568 --> 00:11:26.808
I think Adam Sullivan was talking about this in, uh, the piece we published yesterday, the KBW piece, um, on just how important it is to, uh, not get in a position where you have capacity reserved and a counterparty or tenant who's not exercising that.

57
00:11:27.628 --> 00:11:33.648
So I'm interested to see what the milestones are. I don't know that we saw that they, they were published.

58
00:11:34.208 --> 00:11:45.518
But if I consider, like, the, a lot of the, um, the, the disagreement or just, uh, not everyone agrees about this next-- about the timing of the CapEx cycle.

59
00:11:45.938 --> 00:11:56.708
So if you're just-- If you got a fifteen-year agreement, then those milestones are gonna be very crucial when they happen, exactly what they are, and that's what my eye is on.

60
00:11:56.768 --> 00:12:07.988
I feel like the, like, milestones are a pretty important thing to get right. You mean milestones in terms of energization for this? Energization, yeah. Yeah. So the initial delivery is early twenty twenty-seven.

61
00:12:08.228 --> 00:12:12.508
They didn't really fully specify, I think, how much, how many megawatts exactly that's supposed to be.

62
00:12:12.608 --> 00:12:21.272
But by the end of twenty twenty-eight- The whole 530 megawatts for these two deals with AMD and the undisclosed NeoCloud should be online.

63
00:12:21.312 --> 00:12:28.042
Yeah, but you have to have the reserve capacity, so- Yeah, I, I see what you're saying... well, it's, yeah Yeah, de- definitely. And they've got a pretty big pipeline.

64
00:12:28.512 --> 00:12:39.292
Um, they've got, you know, a gigawatt plus that they're hunting right now. And this will be across five sites, is the, I think the other interesting part about this.

65
00:12:39.432 --> 00:12:47.492
If we actually look at their presentation, which I have in front of me, I won't throw it up on the screen 'cause I've got this stock chart here, and, which is a quick aside on that.

66
00:12:47.572 --> 00:12:59.142
To your point, Charlie, one of the things that became very apparent following the CoreWeave deal is that Core Scientific was not actually milking this for all it was worth. There was a kind of price discovery

67
00:13:00.312 --> 00:13:01.812
element to the CoreWeave deal.

68
00:13:02.012 --> 00:13:17.372
Core Scientific was pioneering for the Bitcoin miners in signing one of these deals, and they took a good deal for the time, at least it seemed, and then you saw Hut8, Wolf, Iron, and Cipher come out with their own deals, and the value per megawatt was significantly higher.

69
00:13:18.072 --> 00:13:29.192
Um, or if not significantly, it was, it was, uh, uh, enough to be notable, right? And now, as you said, with this AMD deal, Core Scientific has inched its way back up into a more favorable deal compared to its peers.

70
00:13:30.752 --> 00:13:39.852
But the, the stock's really not doing too well on this news, and I think that says less about the deal- It, it was up, like, right out the gate. It was up, like- Yeah... a few percent this morning.

71
00:13:39.872 --> 00:13:44.972
It was up, like, 4 or 5, and then it tanked. I think that m- says more about where we are in the cycle.

72
00:13:45.132 --> 00:13:59.632
If Core Scientific had struck this deal in the middle of last year, towards the end of last year, in the heady days of the AI bull run when everyone was, like, effervescent with excitement over all of these new deals getting slung around, the, the stock would've been up much more, no doubt.

73
00:14:00.412 --> 00:14:15.672
But, uh, going back to this expansion and, and going back to that timeline specifically, the AMD expansion will be across five sites. Pecos, Pecos, Texas, Dalton, Georgia- Pecos. Pecos, Texas. Pecos. Yeah. Yeah.

74
00:14:15.932 --> 00:14:27.912
Whatever. Whatever. [laughs] I gotta defend my southwest, my southwest knowledge. [laughs] Yeah. Pecos, Texas, Dalton, Georgia, Hunt, Texas, Muskogee, Oklahoma, and Auburn, Alabama. War Eagle.

75
00:14:29.132 --> 00:14:45.812
Uh, the Pecos, Texas has estimated delivery of first half of 2027. Auburn, first half of 2027. Muskogee and Dalton in the second half of 2027, and the Hunt County, Texas in, uh, estimated 2028, first half of 2028.

76
00:14:47.092 --> 00:15:01.292
So, and, and those are split according to this. 185 megawatts in Pecos, 120 in Dalton, 110 in Hunt, 82 in Muskogee, and 32 in Auburn, Alabama. So, uh, really stretching their,

77
00:15:02.492 --> 00:15:14.452
you know, stretching their capacity across these sites here, and they are currently looking at two plus gigawatts of new site opportunities according to their presentation. So Charlie, I think we can leave Core's there.

78
00:15:14.532 --> 00:15:32.392
Yeah. I'll go ahead and rifle through Applied Digital's earnings as well. So Applied Digital reported their fiscal Q4 and full year 2026 earnings, [lip smack] um, yesterday.

79
00:15:32.452 --> 00:15:46.152
Just a quick note, can we just please get on calendar year for these things? It is- I know... infuriating, man. No, man You know? The, every n- like, there are a few companies that do it this way.

80
00:15:46.712 --> 00:15:56.071
They either do the, like, Canadian reporting standards fiscal year. Iron does something similar, I think, with this, where it's like their fiscal year actually ends halfway through the year. It's crazy.

81
00:15:56.372 --> 00:16:01.772
I mean, thank God they le- they at least use the Gregorian calendar. [laughs] I mean, imagine.

82
00:16:02.892 --> 00:16:12.652
So I'll just focus on the Q4 numbers here, uh, with the main focus, and then also touch on 2026 full fiscal year for, for kind of bigger picture.

83
00:16:13.152 --> 00:16:23.572
So for the Q4 numbers for Applied, revenue was 258.7 million, up 107% year over year. Their adjusted revenue was 240.4 million.

84
00:16:24.082 --> 00:16:30.092
They, they portioned off their cloud services business, so they have co-location and they have a small cloud contingent.

85
00:16:30.612 --> 00:16:42.052
They actually did a kind of reverse merger and created this company called Kronos Scale, which is also publicly traded. So when you adjust for that, revenue, adjusted revenue came in at 20- 240.4 million.

86
00:16:42.332 --> 00:16:55.492
Uh, net loss was 110.6 million. Adjusted net income 12.9 million, so the adjusted diluted earnings per share was, uh, 4 cents per share. Actually beat expectations.

87
00:16:57.132 --> 00:17:12.792
Adjusted EBITDA at 42.4 million versus 1 million a year ago, and net operating income 39.9 million. Now for the fiscal year, revenue was 611.3 million, up 106- 167%.

88
00:17:13.232 --> 00:17:23.032
Adjusted EBITDA coming in at 107.2 million versus 19.6 million for the fiscal year 2025. Major announcements and strategic events.

89
00:17:23.392 --> 00:17:33.292
Again, they spun off their cloud services business and merged with e- Ekso Bionics, a robotics company, to form Kronos Scale. Nasdaq listed CHRN.

90
00:17:33.832 --> 00:17:44.632
Applied retains 96% ownership, but it ex- uh, but excludes it from non-GAAP figures going forward. Signed three new 15-year hyperscale leases in rapid succession.

91
00:17:44.732 --> 00:18:00.812
Delta Forge One, 300 megawatts at $7.5 billion for the deal. Polaris Forge, 300 megawatts at 7.5 billion. Again, uh, Polaris Forge 3, 300 megawatts, 7.5 billion per the last deal as well, similar terms.

92
00:18:00.892 --> 00:18:08.172
And post-quarter, Delta Forge 2, 210 megawatts for 502 billion total in contracted value, all with the same hyperscaler customer.

93
00:18:08.932 --> 00:18:30.948
Their total contracted HPC book is now 1.4 gigawatts, $36 billion in base term revenue- With 86 billion if renewals are exercised, and it raised 2.15 billion in senior secured notes at 6.75% interest due 2031, as well as taking out a $550 million revolver via Goldman.

94
00:18:31.608 --> 00:18:37.588
And post-quarter, they ha- issued another 1.59 billion in notes at 7% interest due 2031.

95
00:18:39.268 --> 00:18:48.668
Their balance sheet, restricted cash jumped to 4.2 billion, debt to 5 billion, total assets to 9.9 billion, up from 1.87 billion a year ago.

96
00:18:49.868 --> 00:19:02.048
And their legacy Bitcoin hosting business actually doing better than if they were mining outright. Uh, it remains, uh, a, you know, it, it, it remains somewhat cash flow positive, high margin cash generator.

97
00:19:02.088 --> 00:19:12.288
It's not affected by Bitcoin prices or difficulty because they're not mining outright. They're paid hosting fees. [lip smack] So overall, pretty decent quarter for Applied.

98
00:19:13.408 --> 00:19:28.228
Uh, I did see some questions about the remaining capacity at some of their leading sites, and I believe they mentioned on the call that they expect to sign a deal soon, so. [microphone thumping] You're muted, Charlie.

99
00:19:29.808 --> 00:19:39.388
It is the Colin show today, and we are gonna keep on rolling- [laughs]... with deals. We're gonna go to Galaxy next. Before we go to the Galaxy expansion, a word from our sponsor, CleanSpark.

100
00:19:39.428 --> 00:19:54.348
[gentle music] We are CleanSpark, America's Bitcoin miner. A publicly traded company with the largest operating hash rate. Powered entirely by self-operated infrastructure across four states.

101
00:19:56.188 --> 00:20:13.678
This is our proof of work, and we are setting the standard for what's next. Learn more about the intersection of energy and Bitcoin at cleanspark.com. Don't worry, Charlie, you'll get yours here pretty soon. I will.

102
00:20:14.607 --> 00:20:32.488
[laughs] We will hop on over to Galaxy, which has announced a new expansion for its AI play. This is coming to y'all from Blockspace. Galaxy Digital expands AI play with 74-megawatt Texas data center campus.

103
00:20:33.588 --> 00:20:45.908
The details, as reported by Blockspace, Galaxy has acquired 500 acres in McGregor, McLennan County, Texas, under a development agreement with the city of McGregor for a new AI and HPC campus.

104
00:20:46.588 --> 00:20:56.868
The initial development phase is 74 megawatts, [lip smack] which is notably just under ERCOT's 75-megawatt large load threshold.

105
00:20:57.648 --> 00:21:13.428
Galaxy out here trying to circumvent the new batching requirements by getting in just under that 75-megawatt threshold. Pretty smart. Uh, it, it should help them expedite energization for this site.

106
00:21:13.438 --> 00:21:19.388
They have more aspirations for it though. They won't stop at 74 megawatts, according to the press release.

107
00:21:20.488 --> 00:21:31.948
They are hoping to expand the site into a multi-100 megawatt campus through 2030, saying it has the potential for that, but it's contingent on the utility building out additional transmission infrastructure.

108
00:21:33.288 --> 00:21:47.858
[lip smack] Power is expected to start flowing in 2028. The land purchase generated 7.5 million in land sale revenue to the city of McGregor, Texas. And Galaxy will be building and funding its own private substation.

109
00:21:48.448 --> 00:22:06.588
Also, in accordance to some of the new rules we're seeing out of the Public Utility Commission in Texas, where they're telling data centers, "You're gonna have to build your own infrastructure, buddy, because we're not subsidizing that cost on the rate payers during this massive boom in building in the state."

110
00:22:06.628 --> 00:22:16.848
It will be a closed loop water cooling system, the same approach as Helios. So if you're worried about not having clean drinking water in Texas, don't worry, you will be fine.

111
00:22:16.908 --> 00:22:28.908
It'll be a closed loop system, and you can point to this when people tell you that the data centers are guzzling up all of the water and we're quickly approaching a Mad Max-esque apocalyptic scenario.

112
00:22:29.308 --> 00:22:38.528
S- it's important to note that this is just a land banking and early stage development play. There's no tenant here yet. They still have to build on the site, but it is

113
00:22:39.488 --> 00:22:46.287
Galaxy's first expansion outside of their Helios site, which has CoreWeave as its anchor tenant.

114
00:22:47.418 --> 00:22:58.068
[lip smack] That site, uh, very impressive, over a gigawatt of scale, and they are continuing to build phase two as we speak. Now, the other interesting part of this, and they didn't give a timeline for this,

115
00:22:59.228 --> 00:23:05.068
they're, they're, they're hitting on all of the FUD angles in this release, if you can tell. You know, they're like, "We're paying for the transformer.

116
00:23:05.468 --> 00:23:14.488
It's gonna be under 75 megawatts, so don't worry about it being too large just yet. By the way, we're using closed loop water." And they say that the estimated tax

117
00:23:15.448 --> 00:23:26.008
revenue for the city of McGregor will be 130 million in local property tax from this site. My question is, what, is that over the lifetime of the site? What timeframe are they using for that?

118
00:23:26.608 --> 00:23:38.428
'Cause it's clearly not 130 per year. But needless to say, they are hitting on all of the data center propaganda topic, talking points to make this seem like a big benefit to the area.

119
00:23:38.518 --> 00:23:47.568
And, you know, it does seem like overall the city of McGregor, McGregor consented to this, so they're more or less happy with what they see. Yeah, I mean, when you...

120
00:23:47.798 --> 00:23:51.748
I- it's, it's great that Galaxy is, uh, definitely skating where the puck is going.

121
00:23:51.868 --> 00:24:06.308
I would, I would argue the puck has already landed here and everybody else is trying to catch up because right now, um, you can buy transformers. You can acquire power at cost, and it's, like, difficult in a timeline.

122
00:24:06.368 --> 00:24:08.988
You, you can almost not buy goodwill anymore.

123
00:24:09.088 --> 00:24:36.296
You have to figure out how to navigate this, this, like, marketing and tactic, uh, because if 200 people show up at a city council meeting or, or in the, you know, in the wrong gym and are angry and have signs- Uh, those 200 people who you had no idea who they were, their names, or why they would possibly be a bottleneck, uh, they can shut down or indefinitely delay your entire site.

124
00:24:36.806 --> 00:24:45.256
When you're spending billions of dollars, you know, nine figures now, potentially multiple billion dollars, this is something you have to figure out.

125
00:24:45.316 --> 00:25:04.916
So in a way, it's almost like you-- it's like almost like your marketing spend has to be so precise and your, your, your, you have to s-send the right people to go knock on the right doors, uh, in the, in the small community in nowhere, Texas to, to be able to get these deals done.

126
00:25:05.006 --> 00:25:17.996
And in a way, Colin, that's how it should be. You know, in a way, this is my, this is my hot take, which is, uh, empower the American people to, uh, uh, you know, property rights.

127
00:25:18.216 --> 00:25:32.346
Be a, be a, be a-- let them-- give them the tools to let s- say what goes on in the community. However, unfortunately, sometimes the people can be kind of dumb. So, uh, shout out to Galaxy.

128
00:25:32.396 --> 00:25:47.436
They seem to have like identified this as like the constraint to getting permitting done and, uh, their, their press release certainly hits those points. So. Yeah.

129
00:25:47.516 --> 00:25:56.956
Hats off to the Galaxy team on the expansion, and we'll be eagerly awaiting updates for the construction of the site and whether or not they have a tenant in the barrel.

130
00:25:57.056 --> 00:26:09.936
I'm sure that if they're moving forward with it, there's probably some conversations underway already. I think that covers that, Charlie. Yeah. Perhaps we should- We're gonna roll to another deal in guess where, Colin?

131
00:26:10.296 --> 00:26:17.036
Texas. We're gonna go to the Meta and BlackRock deal. I'll lead this one after a word from our sponsor, Luxur.

132
00:26:17.516 --> 00:26:29.366
[fire sizzles] This episode of Blockspace Live is brought to you by Luxur's Commander, Bitcoin miner management software for enterprise operations.

133
00:26:29.996 --> 00:26:35.766
Luxur's Commander gives you real-time fleet monitoring, bulk remote commands across your fleet, and Intelligent Miner.

134
00:26:36.326 --> 00:26:43.946
That's an automated profitability engine that runs every five minutes and tests your fleet's power settings against live energy and hash rate markets.

135
00:26:44.596 --> 00:26:49.236
ERCOT backtests show a 10% improved profitability with intelligent mining over binary mining.

136
00:26:50.036 --> 00:26:59.796
Commander Pro is $100 per megawatt or a 25 basis point pool fee adder, roughly half the cost of competition, but you can also try it for free for 60 days.

137
00:26:59.876 --> 00:27:17.516
If you'd like to learn more, go to luxur.tech/commander to get started. Charlie, what are the spooks up to? BlackRock. BlackRock getting into Meta's business with a new deal. You're muted, by the way. Oh my goodness.

138
00:27:17.716 --> 00:27:35.336
Okay. So here it is. As on Blockspace this morning, Meta and BlackRock form $14 billion venture for one gigawatt El Paso AI campus. Not to be con-confused with the Core Scientific AMD $14 billion deal also this morning.

139
00:27:36.016 --> 00:27:51.816
Not a typo. This is actually the value of this AI campus in El Paso, Texas. You may know El Paso as a very southern city on the border of Texas, bordering Juarez, Mexico. So the deal is as follows.

140
00:27:52.056 --> 00:28:02.216
Meta and BlackRock announced a joint venture this morning to build and own a one gigawatt data center in campus, campus in El Paso, Texas. Total development cost is about $14 billion.

141
00:28:02.716 --> 00:28:15.216
BlackRock funds, BlackRock managed funds take about 80% of the venture, and Meta keeps about 20%, leases the entire campus, and runs it, construction manage it, services it, et cetera.

142
00:28:16.116 --> 00:28:21.916
Meta will contribute the land and construction, which is in progress, I believe, worth about $2.3 billion.

143
00:28:22.416 --> 00:28:40.236
BlackRock will put in $4.9 billion in cash, and Meta will get a one-time $1 billion payout up to, uh, to make it the 80/20 split between BlackRock and Meta. Um, BlackRock's side is funded by $12.5 billion in debt.

144
00:28:40.316 --> 00:28:47.516
JPMorgan and Morgan Stanley arranged the bond through a holding company tied to BlackRock's estate. So what does this mean?

145
00:28:47.856 --> 00:29:02.356
This means Black- this means Meta gets a gigawatt of compute, not necessarily o- like on its books, and a billion dollar cash back as they sign the deal. If this sounds familiar, this is basically the Hyperion playbook.

146
00:29:02.416 --> 00:29:20.396
Hyperion is the originally two, now five plus gigawatt Meta data center in southern Louisiana. Uh, the bond on the BlackRock side drew as much as $20 billion in orders,

147
00:29:21.296 --> 00:29:31.556
or was about 1.6 times oversubscribed. So when we're talking about the CapEx cycle, where does the money come from? What, where is the demand here? There's demand for this.

148
00:29:31.616 --> 00:29:45.076
So we don't see, y- you know, if you can get a creative financing, and you can get, uh, you can tap the money markets, they're still flowing. A little bit on the actual, uh, physical side of this.

149
00:29:45.176 --> 00:30:01.496
El Paso Electric wants approval from the McLeod Generation facility for 366 megawatts of natural gas, which by my calculations, I'm doing the mentat thing from Dune, it's like 400 million cubic feet per day of gas.

150
00:30:02.436 --> 00:30:14.216
Uh, 813 modular generators, uh, which is wild to say out loud. Uh, and another angle is that this is not on ERCOT.

151
00:30:14.376 --> 00:30:22.276
Even though El Paso is in Texas, it's not on the Energy Reliability Council of Texas, ERCOT.

152
00:30:22.416 --> 00:30:33.464
It's, it's, this is not pulling from ERCOT, so it doesn't have to jump through the hoops that ERCOT is starting to put- Uh, onto their grid, uh, for large new loads.

153
00:30:34.164 --> 00:30:43.164
Um, I got another couple comments or two, but I'm throw it to you, Colin. Yeah, I just think it's interesting for what you were saying about this is kind of,

154
00:30:44.304 --> 00:30:51.204
Me- Meta gets to add a gigawatt somewhat off the books, but they don't take on basically any of the financial risk.

155
00:30:51.924 --> 00:31:02.704
What really struck me was the bond that is being arranged by JPMorgan and Morgan Stanley for BlackRock that will be specifically tied to BlackRock and not Meta.

156
00:31:03.524 --> 00:31:12.824
You know, Bitcoin miners could only dream of having a counterparty like BlackRock. Maybe soon they'll get one somewhere. Maybe soon. But- But they won't be Bitcoin miners anymore. Yeah.

157
00:31:12.904 --> 00:31:24.124
[laughs] They'll be vessels for the deep state. Yeah. They'll be vessels for Larry Fink. [laughs] Oh man, we got, we gotta, we gotta cool it with our conspiratorial- Yeah, no... totally pivot to an AI show.

158
00:31:24.344 --> 00:31:38.264
Yeah I say that as a joke to all of the people who continue to use the whole data center and AI CapEx cycle as an ability to advance certain conspiracy theories regarding- Well, I mean, one, you know, one flock camera equals one data center.

159
00:31:38.464 --> 00:31:39.284
I don't make the rules.

160
00:31:40.184 --> 00:31:52.544
And the other interesting thing too is this idea of it being, there's like a five-year bridge period where it stays off the transmission system because of the 366 megawatts of natural gas that's being deployed on site.

161
00:31:53.124 --> 00:32:05.204
Yeah. That's a massive undertaking. I mean, just think about that honestly. That, that is a huge source of base load energy- Yeah... that is just being used for bridging this data center.

162
00:32:06.204 --> 00:32:13.904
You, the, um, the, the American mind cannot comprehend how much natural gas is in this part of the country.

163
00:32:13.944 --> 00:32:27.144
Now, this is a little bit outside of the, uh, this is actually a little bit outside of the, uh, uh, outside of like the West Texas, East New Mexico, uh, Permian Basin producing zone.

164
00:32:27.904 --> 00:32:32.644
But, uh, you get really close to where all the pipelines, uh, cross.

165
00:32:32.884 --> 00:32:48.004
Uh, so um, I'm not drawing off the top of my head to have exactly the, where all the pipelines are, but, uh, yeah, there's so much natural gas and it's, it's really a location and short, uh, short transportation.

166
00:32:48.104 --> 00:32:58.764
My question is, what, what happens to all of those turbines in that generating capacity once they get connected? Do you just keep them on site just as a backup? I don't know.

167
00:32:58.804 --> 00:33:10.544
Do you- I mean, you can, you can pull these turbines out and move them somewhere else. Uh, I don't know what their like lifecycle is. I know depending on the turbine, it's like an, an incredibly specialized,

168
00:33:11.464 --> 00:33:21.804
uh, piece of machinery. Elon Musk, like actually months ago, pretty presciently identified turbine blades as one of the obscure upstream,

169
00:33:22.724 --> 00:33:33.554
uh, bottlenecks as they're so precise and have a very specific type of engineering, and they kind of are a linchpin for all of like natural gas production. So I mean,

170
00:33:34.624 --> 00:33:51.124
yeah, uh, it's, it's one turbine in the hand is worth 10 d- you know, down the road, you could say. And they're going to Juarez, Mexico. I'm sorry, El Paso. And, uh, yeah, I don't know.

171
00:33:51.204 --> 00:34:04.604
If you wanna be, like I'm trying to think 10, 20, 30 years from now. Let's say data centers are not as like, they're not driving the physical infrastructure expansion of the American economy.

172
00:34:05.724 --> 00:34:21.564
You know, it's kind of bullish for the citizens of these, uh, of these cities. Like you're about to have at the, at the expense of capitalists, massive behind the meter generation built out.

173
00:34:21.984 --> 00:34:35.224
So how that plays out, probably very path dependent, we'll have the luxury of, of, you know, hindsight down the road, but for now, move aside Hoover Dam, like we are building natural gas. Get in loser.

174
00:34:35.264 --> 00:34:46.314
We're building- Yeah... nat gas. Yeah. You know, and last thing I'll say, the other thing that struck me, and you covered this though, is just that the four-year lease kind of gives Meta an out here.

175
00:34:46.384 --> 00:34:59.484
If, if the CapEx cycle turns, they can just bail on this project. Yeah. If it can turn. And compare that to the, the, the lease terms or the, the durations of the other neo clouds. Right. Like 15, 20. Yeah.

176
00:34:59.704 --> 00:35:06.304
I think there was a 30-year one. Like these guys are locked in, so I, I, flexibility is king.

177
00:35:06.424 --> 00:35:18.204
This is really where the hyperscalers get to come in and say, "Okay, maybe we didn't enjoy the insane upside to our stock that, you know, the Irons of the world or the Hud-8s of the world have gotten.

178
00:35:18.924 --> 00:35:35.724
Um, but we have agility." And they have big, big treasuries they've been building for years. So, you know, if they can not get two over their skis, like this has, uh, railroad baron,

179
00:35:36.884 --> 00:35:49.804
uh, money written all over it, so. We're gonna move on to our guest, Rob Hamilton, in the wings, and we're not gonna talk about Bitcoin that much. We're gonna talk about AI.

180
00:35:49.904 --> 00:35:53.004
Before we bring Rob up, a word from our sponsor, Ligos.

181
00:35:53.624 --> 00:36:09.104
[static sound] Bitcoin is struggling to hold 65K, and if you're using a Bitcoin-backed lender, you should be asking yourself, is your lender playing funny games with your precious coins?

182
00:36:09.964 --> 00:36:18.784
Do yourself a favor if you're using a Bitcoin-backed lender and look no further than Ligos Finance. 'Cause with Ligos, you don't have to ask yourself that question.

183
00:36:18.844 --> 00:36:29.724
You don't have to worry about becoming the next Celsius or BlockFi victim. With Ligos Finance, you can always be in control of your Bitcoin with Bitcoin native smart contracts. That's right.

184
00:36:30.064 --> 00:36:38.004
They are our preferred lender over here at Blockspace because they use Bitcoin native smart contracts to make sure that you always hold your keys. You heard that right.

185
00:36:38.484 --> 00:36:52.764
With Ligos Finance, there's no rehypothecation, no wrapping, no bridging. No funny business, just you holding your keys and getting competitive rates as low as 8.5 to 10% APR for their Bitcoin-backed lending services.

186
00:36:53.224 --> 00:37:07.764
If you'd like to learn more, go to lygos.finance. All right, let's talk shop with Rob. Robert Hamilton, welcome back to the show. Hello. Thanks for having me again. Absolutely. Always a pleasure, Rob.

187
00:37:08.904 --> 00:37:24.144
Um, okay, so let's talk AI. Let's talk open weights. This week, Jensen Huang made his first tweet ever, shook everybody to the core. What was, like, kind of a niche topic for those of us in AI out of open models

188
00:37:25.044 --> 00:37:36.164
thrust into the limelight. I feel like everybody else is kind of feeling this, with, like, Fable, and Mythos, and the Anthropic, you know, hand of God directing what you shall and shall not do. But

189
00:37:37.664 --> 00:37:54.064
I guess what I'm asking is to explain what the heck an open model is. Give us a little- [laughs]... 101, 201 explainer here. Absolutely. Yeah, so the term of art in the industry, uh, I guess maybe if you wanna start.

190
00:37:54.324 --> 00:38:05.644
Y- if you're- Yeah... familiar with the idea of open source, and if you're not, let's just start there. Uh, Linux is an open source operating system, right? Windows and Mac are closed source. What does that mean?

191
00:38:05.724 --> 00:38:16.224
It means that those companies, um, Apple and Microsoft, own that code, and they don't really share with you. You just buy the final software, you know, of the operating system.

192
00:38:16.964 --> 00:38:28.054
Open source, though, means that all of the code underneath it is present and available for someone to look at. Now, in the AI world, the way we kind of view this is, um,

193
00:38:29.184 --> 00:38:38.924
in the same way where you can have Anthropic and OpenAI, their models, you can use them, but you're not actually going to be able to, like, run that on your own computer, right?

194
00:38:39.314 --> 00:38:53.304
The open source movement, which is also commonly called open weights, is I can, in theory, download a model and run it on my own computer. Uh, the way I would think about this, there's kind of two buckets, actually.

195
00:38:53.464 --> 00:39:06.644
So open weights is what most people are normally talking about. Weights come from that term weights and biases, which is the, basically the math that undergirds all of these large language models. You can actually...

196
00:39:06.684 --> 00:39:12.324
It's linear algebra, like matrices, like really large grids of numbers. And the...

197
00:39:12.364 --> 00:39:18.304
If you have the weights and biases, if you have open weights, it means that you can load it on your own computer, and you can run it yourself.

198
00:39:18.404 --> 00:39:27.764
You may need very high-end commercial hardware to be able to run it, but the idea is that the data is available for you to look at, and you can run it yourself, which is...

199
00:39:27.824 --> 00:39:33.524
has a whole different context of different business applications, uh, that we can get into.

200
00:39:33.544 --> 00:39:46.744
But Jensen's letter that came out was basically an encouragement of we need to be nurturing and supporting these open source, open weight models for the just dynamism of American capitalism and just the strength of US hegemony.

201
00:39:46.784 --> 00:39:57.374
You need to be able to have these tools that are open and available so anyone, everywhere, is able to leverage these tools how they see fit. Rob, one quick question from me here. Mm-hmm.

202
00:39:57.424 --> 00:40:05.964
I think one of the large concerns about these open weight models is they are dominated in the Chinese market right now. Mm-hmm. And that's where all the advancements are coming from. So- Yeah...

203
00:40:06.004 --> 00:40:15.714
one of the kind of milquetoast criticisms, and I say milquetoast 'cause it's really easy to arrive at, and it kind of plays to the idea of the CCP boogeyman- Mm-hmm...

204
00:40:15.774 --> 00:40:23.564
is that, well, we want American companies using American models because we only want our government spying on what people are doing- [laughs]... not Chi- the Chinese government.

205
00:40:24.104 --> 00:40:35.024
Does the, does the fact that they're open weight and you can run them on your own hardware mollify that concern at all? Or is that not necessarily off the table if you're running these things locally?

206
00:40:35.104 --> 00:40:43.344
I think it largely mitigates that, right? In theory, you could have a data center that never talks to the internet that you can use for all of your analysis, right?

207
00:40:43.944 --> 00:41:02.803
Um, I would say w- with owning the hardware that these, uh, calculations are being done on, with competent IT infrastructure, making sure, like, your firewalls are closed off and it's very controlled in how you can access data into the system and pulling data out of the system, that you're able to mitigate those concerns.

208
00:41:03.423 --> 00:41:14.764
Uh, this also kind of dovetails into the next part of the conversation around China is the reason why China is leading the open source movement is it's very classic in any market.

209
00:41:14.774 --> 00:41:22.114
The second mover, the laggard, is going to go for open source. If you remember, this is what Meta did with their Llama models, right?

210
00:41:22.144 --> 00:41:37.784
They did open source models because OpenAI and Anthropic were so far ahead, your best shot as a second mover is to try and create an open source movement with a larger community base that's gonna engage, and you can get somewhat network effects building on that open layer to eventually be able to overtake it.

211
00:41:37.884 --> 00:41:53.744
That's a lot of the strength of what Linux ultimately did in the, in the hardware, in the, in the operating system and kernel environment, was they, as being open source, over time, more and more capital and resources get invested into it, which becomes a shared good and win among everyone who participates in that.

212
00:41:54.984 --> 00:42:07.754
Yeah, this second mover, uh, angle is actually really interesting. Um, so in a way, conspicuously... So conspicuously, Anthropic has not signed Jensen's letter. It's signed by everyone else. In fact, even- Mm-hmm...

213
00:42:07.764 --> 00:42:17.804
OpenAI signed it kinda, I don't know, later on. Yeah. But- They were, they were a follower. They were not one of the first ones. Yeah, hedging, hedging their bets, trying to curry favor, you know? Yeah. It's so...

214
00:42:17.984 --> 00:42:27.444
Anthropic is the, the only, like, premier model who has not signed this. I, did- Mm-hmm... did xAI or Grok? I forget. Anyway, Anthropic didn't. Um,

215
00:42:29.084 --> 00:42:36.764
one, is this, like, a de facto admission that the industry is kind of, uh, uh, not, not lined against Anthropic, but

216
00:42:37.704 --> 00:42:49.180
they, they f- they want to create some kind of opposition here 'cause Anthropic seems to have a direct line to the White House all out of nowhere. And- What does history tell us how this could play out?

217
00:42:50.520 --> 00:43:07.600
Yeah, this is interesting. With Anthropic's success, and you could say, I, personally, I feel GPT 5.6 is better than Fable 5 for what I use, um, the belief that... And Anthropic's not really coy about hiding this.

218
00:43:08.040 --> 00:43:24.030
They just believe that these models get so powerful that they need to actually be restricted with, uh, how they are used and leveraged for safety concerns, whether that is chemical, biological warfare, nuclear weapons, or cybersecurity hacking.

219
00:43:24.060 --> 00:43:33.820
They believe that these things should be gated. And now for the tension between the US and China in a geopolitical sense, these models, the big one that everyone's talking about right now is Kimi K3.

220
00:43:34.520 --> 00:43:48.420
Uh, and it's a model that comes out of China but is fully open weights. On Monday it came out. I just started using it, uh, y- today, basically, to start poking around. And a real, just...

221
00:43:48.570 --> 00:44:00.400
Th- there's several things or so to, like, to think about this where, uh, with GPT 5.6, and especially Anthropic, if I ever asked it to evaluate my code base for, like, security audits, it would shut down.

222
00:44:00.640 --> 00:44:07.200
It would not give me the latest model. But now with Kimi K3, it lets you do all of this stuff. W- what, is that, is that a liability thing, Rob?

223
00:44:07.260 --> 00:44:21.320
Why is that n- why will it not, why will the two frontier models in the US not give you feedback? I believe it is a strategic ploy to create distance of a regulatory moat.

224
00:44:21.920 --> 00:44:30.180
You could say partially it is because they want to be able to protect consumers. M- at the same time, though, I'm not so sure.

225
00:44:30.840 --> 00:44:39.460
Uh, information yearns to be free, and you have this really poor asymmetry now where black hat hackers, um, have good enough models to do all of this hacking.

226
00:44:39.520 --> 00:44:44.640
But if you're a white hat hacker and you're paying the full price, you still can't get protected. And that's the other thing, too.

227
00:44:45.000 --> 00:44:55.660
Kimi K3 is somewhere between 30, 20 and 30% of the price of, um, of Fable 5 and of GPT 5, 6 when you're using it at scale. So it's massively cheaper.

228
00:44:56.140 --> 00:45:01.800
It's way more helpful as software helping me try to do my job, right, in writing software.

229
00:45:02.460 --> 00:45:12.120
Um, and I think it just ultimately, there is a, a tension now, 'cause you saw when the Kimi K3 benchmarks started coming out maybe a week and a half ago, there was a lot of...

230
00:45:12.180 --> 00:45:18.380
Anthropic and OpenAI aren't publicly traded, but the, you know, they were getting kind of, like, pressure on their valuations.

231
00:45:18.400 --> 00:45:24.260
Whereas all of the data center companies, the stocks were ripping, like CleanSpark, Irin, Mara, when...

232
00:45:24.270 --> 00:45:34.780
'Cause the, ultimately, this is a part when you think about it from a geopolitical standpoint, the US versus China, the US still has all of the chip manufacturing infrastructure to be able to build all this stuff out.

233
00:45:34.860 --> 00:45:38.760
Kind of like the large scale data centers for this, at the moment the US is in the lead.

234
00:45:38.820 --> 00:45:52.140
All of these neo clouds, which are things that even if the software gets quickly commoditized down, there are hard capital constraints of what the US is much better at servicing at the moment to be able to kind of pr- pr- uh, protect and project hegemony.

235
00:45:53.660 --> 00:46:04.140
So- Rob- Go ahead. Sorry, Jordan, before you jump in, just how is it that Kimi K3 is so much cheaper? So- Is that because the Chinese government is subsidizing the tokens? Yeah.

236
00:46:04.460 --> 00:46:10.440
Is, is that- Well, so- Is it literally just that simple? So in theory, if you had the machinery, you could run it for free, right?

237
00:46:10.460 --> 00:46:21.120
But to actually run Kimi K3 at reasonable scale, it's like a $1.5-$2 million specialized basically server rack to be able to do it at, like, a reas- at a quick pace.

238
00:46:21.880 --> 00:46:28.530
Uh, so the cost is because that's just at the moment what they're able to charge and profitably, like, to keep these machines running 24/7.

239
00:46:29.160 --> 00:46:34.020
And this is where it goes into st- distillation, where the US puts out this frontier model.

240
00:46:34.880 --> 00:46:48.060
China over the past year and a half has gotten incredibly adept at basically being able to, uh, what's called a distillation, use those frontier models, store all of the outputs, and use that to train their own model.

241
00:46:48.140 --> 00:46:57.480
So they're basically able to take the intelligence from the top model and colli- collapse it down. But this is back, you know, uh, I was listening to, uh, All-In podcast.

242
00:46:57.520 --> 00:47:06.260
David Friedberg said, "When I worked at Google, we would go to Yahoo all the time and see what Yahoo returned in their search engines and use that to improve our product," which is basically the same thing happening here.

243
00:47:06.719 --> 00:47:16.860
Additionally, a lot of enterprises require what's called zero data retention, where if I'm using a vendor, they don't hold anything on their side. Anthropic refuses to let you do that with Fable.

244
00:47:17.120 --> 00:47:28.360
They say, "If you're going to use this model, you must let us have full access to the queries that you're running for the next 30 days." Kimi K3 now just lets you not have to worry about data retention at all, right?

245
00:47:28.400 --> 00:47:36.760
So there's, like, whole classes of enterprises that are able to now leverage the latest frontier or close enough to the frontier tech without being handicapped.

246
00:47:36.780 --> 00:47:49.210
And I think the, the gap, it, for a while it was like six to nine months is what the frontier labs were, that everyone else has gotten incredibly efficient at getting the distillation out of these models to where Fable now is starting to hide what are called the reasoning traces.

247
00:47:49.220 --> 00:47:57.800
Which is if you've ever used these high-end models, you have a, you ask a question and then it ultimately gives you an answer, but you can click to see the thoughts, and that's like the detailed showing your work.

248
00:47:57.810 --> 00:48:05.480
They stopped showing that for a lot of queries 'cause they realized that that was exactly what was being pulled out for the distillation. So, um, it's interesting times.

249
00:48:05.540 --> 00:48:10.770
I think it's ultimately good for capitalism, freedom of markets and information to be able to have this.

250
00:48:10.800 --> 00:48:18.000
There's no reason why Anthropic and OpenAI should be like a government protectorate kingdom to kind of protect their shareholder value.

251
00:48:18.060 --> 00:48:22.860
It's better if these models are widely distributed everywhere, and the model companies are gonna have to just find other revenue streams.

252
00:48:23.900 --> 00:48:33.640
Which, uh, gets kind of to the heart of the stuff we cover, data centers, the value thesis. Yes. They're basically tokens. They sell tokens. Yeah.

253
00:48:34.300 --> 00:48:44.900
Um, and so we care a lot of the price of tokens, and this is kind of determined by the software and, uh, what happens- Mm-hmm... up the stack. Mm-hmm.

254
00:48:44.960 --> 00:49:00.782
So what, in your view, what does this mean for, uh, the, the sh- m- short and medium term- For token vendors, like Neo clouds, like the inference providers, what does this mean? Oh. Uh, yeah.

255
00:49:00.832 --> 00:49:07.312
I think it's incredibly bullish, because before the only high-value token vendors that you had were Anthropic and OpenAI, right?

256
00:49:07.352 --> 00:49:12.932
Those were the only ones that charged the most premium high rate, and most of that value went to Anthropic and OpenAI.

257
00:49:13.192 --> 00:49:18.612
But now if you have a model that's 25 cents on the dollar, cheap, like it's, it's, you know, you're saving so much money.

258
00:49:19.072 --> 00:49:30.012
I just ran a query earlier today that was f- $35 in Kimi K3 that would've been probably somewhere around $250 if I used the Anthropic, and I now am gonna make many more queries.

259
00:49:30.182 --> 00:49:39.192
And the person- the Neo cloud that's hosting the data center is gonna be able to maintain that margin rather than it being it going all to Anthropic and OpenAI's pocket, right?

260
00:49:39.232 --> 00:49:52.492
So for data centers, massively bullish for the o- revenue opportunities that are able to accrue to them, 'cause to your point, they're the ones that have done the, the hard CapEx, uh, and OpEx overhead to be able to administer these data centers at scale.

261
00:49:52.732 --> 00:50:03.312
That's where the margin's gonna be able to go to for this, like, metered intelligence. I think in the longer run, metered intelligence is not gonna be the way these LL- like, these AI companies are gonna make money.

262
00:50:03.912 --> 00:50:09.852
Um, the data centers can all day 'cause they're, they're literally the gas station in the metered intelligence. They're, they're, they're sitting right there at the tap.

263
00:50:10.232 --> 00:50:21.452
But if you're one of these frontier labs, you're gonna have to find other things to do rather than releasing a frontier model and charging, you know, s- five to six times what the rest of the market's able to do for almost comparable intelligence.

264
00:50:21.752 --> 00:50:26.622
What do you think that might be? Like, what do they do? Uh, this weirdly [laughs] enough- Uh, yeah...

265
00:50:26.632 --> 00:50:33.492
this gets, like, to, like, the fat app, like, fat protocol thesis that we in the world of Bitcoin and crypto have been, like- Yeah... thinking about for forever. Yeah.

266
00:50:33.632 --> 00:50:41.732
Where does the va- where might the value accrue, Rob? I think the value accrues, I think they're gonna do things like, um, working on healthcare. They're gonna work on frontier physics.

267
00:50:41.852 --> 00:50:56.732
They're gonna work on patenting, like, bleeding edge, you know, bioinformatics and, like, uh, whether it's can- curing cancer, things that they're going to be able to specifically narrowly point on with leaders in the field to be able to get this access.

268
00:50:56.812 --> 00:51:05.212
I've talked to people, um, casually that are within the, I would say the life sciences and working on longevity and life research, and that...

269
00:51:05.232 --> 00:51:11.622
What happening is these frontier labs are setting up big partnerships to, like, let us help you work on this next clinical trial, right?

270
00:51:11.652 --> 00:51:22.342
And I think it's gonna turn into that's something that they uniquely are gonna be able to do versus an open source model. Maybe you'd need that extra lift of 10% more intelligence to get the really big breakthroughs.

271
00:51:22.632 --> 00:51:30.172
If you're seeing all these breakthroughs of people, um, taking 80-year-old math theorems and proving or disproving them, like, that's where you're gonna get it.

272
00:51:30.332 --> 00:51:38.012
In the physics, the life sciences, like, there's gonna be... Even maybe cybersecurity, right? Having the ultimate frontier thing, like, y- I used to...

273
00:51:38.072 --> 00:51:46.332
You know, you could pay $50,000 for a code audit, or I can pay Kimi $100 and get the full, like, readout at way better detail than anything I was able to do for before.

274
00:51:46.652 --> 00:51:57.732
But maybe if I have OpenAI, maybe I'm willing to pay them $20,000. It's better than the old cybersecurity vendor at $50,000, and they can still chart, you know, spend $500 of tokens and charge me $20,000.

275
00:51:57.772 --> 00:51:59.912
That's a really great margin, right, 95% margin.

276
00:51:59.981 --> 00:52:10.672
Like, there's other fields and things they're gonna be able to do, but this ambiguous metered intelligence I think is gonna start, uh, falling apart, 'cause I'm guessing the open source models are gonna get even tighter and tighter to that.

277
00:52:10.692 --> 00:52:18.992
Like- It kind of makes sense, too, when you think about not every query is as valuable as another one. Bingo. Right? Like, me asking ChatGPT is, like, how long should I braise my lamb shank?

278
00:52:19.322 --> 00:52:29.722
[laughs] Isn't quite in the same, is, is, isn't quite in the same category as someone trying to find a cure for lung cancer, right? Right. Exactly. You know, Rob, la- last question from me- Sure...

279
00:52:29.732 --> 00:52:32.212
before I kick it back to Charlie for final thoughts.

280
00:52:33.112 --> 00:52:46.232
Do you think this letter from NVIDIA, OpenAI, Microsoft, et cetera, does anything to move the needle for the federal government in terms of banning these open weight models? Do you think that it's basically a...

281
00:52:47.072 --> 00:52:56.332
Do you think this puts this, the, the problem to rest, I should say, in the eyes of the government? I think it's a very strong nudge for Jensen to come out publicly, first time ever on Twitter, to kind of drop this.

282
00:52:56.852 --> 00:53:09.391
I will say if everyone's looking at each other's incentives, the open source model world greatly benefits NVIDIA, 'cause they're the ones with the differentiator, which are the GPUs themselves, right, and the fabrication of these semiconductor chips.

283
00:53:09.401 --> 00:53:10.772
They are a unique place.

284
00:53:11.951 --> 00:53:23.372
Um, I think in the long run we'll actually see the ASICification that we saw in Bitcoin, but we're gonna see it for AI chips, where you're gonna be able to buy, like, a Groq 4.5 heavy ASIC that'll be a fraction of the cost and run really fast.

285
00:53:23.472 --> 00:53:29.152
Uh, but that's still gonna take a lot of time to play out. But I think to directly to answer your question, yeah. I think this is...

286
00:53:29.212 --> 00:53:43.672
Jensen was putting his slight thumb on the scale, and I, I'm grateful for it because to be against the open source nature and the open weights of these models is to ban math, and we don't wanna do a rerun of the '90s where people were trying to ban prime numbers because of cryptography.

287
00:53:44.312 --> 00:53:46.592
Y- you're basically banning, like, matrices. That's what it is.

288
00:53:46.632 --> 00:54:00.092
You're banning math, and there's not a free society if you're banning math, and I think it's healthier and better for the American experiment and for all of our lives to try and wrestle with what it looks like in a new world rather than trying to put the toothpaste back in the tube.

289
00:54:00.152 --> 00:54:09.392
Yeah, the 250 long American experiment. Jensen says, "You can all compete for my product." [laughs] That's right. That's right.

290
00:54:09.972 --> 00:54:23.782
I guess my last question, um, and this is probably worth a long, a longer-term dedicated interview and... But, um, you mentioned, like, you could pay 50K for a code audit from a great top-tier company. Mm-hmm.

291
00:54:23.812 --> 00:54:32.982
You'd maybe pay, you'd be down to pay 20K from a premier OpenAI, you know- Yeah... model and, and product. Maybe even less than that.

292
00:54:33.592 --> 00:54:49.332
Um, but this gets into, like, normal small businesses, startups who have to now think about how you hire people, where do you spend money, where do you allocate resources, and this gets into the token spend question.

293
00:54:49.342 --> 00:55:01.142
Mm-hmm. Token spend and, and employees. Basically, a lot of what you're doing is figuring out where to allocate money towards intelligence. Yeah. How do you think about this, Rob? Is it too early to have a clear view?

294
00:55:01.456 --> 00:55:13.276
What is your current view on, I would say, agile tech companies- Yeah... who are building things and trying to figure out where to allocate resources? Yeah, so I think the stage I'm at right now...

295
00:55:13.336 --> 00:55:15.676
Well, there's a gradient, right?

296
00:55:15.736 --> 00:55:27.896
Where to I don't use this stuff ever, to maybe I have ChatGPT in the browser, I ask questions, whether code or operational, like strategy things, to leveraging Claude code and Codex and open code.

297
00:55:28.516 --> 00:55:37.236
Um, I, I use open code for the Kimi tool. It goes right in there. All the models are hosted in America with zero data retention. Like, it's a very clean setup. Uh,

298
00:55:38.136 --> 00:55:50.456
and then what I'm trying to do now is I feel that my engineers have a pretty good grasp for their individual silo tasks, and the next step is trying to go one level higher to, like, a company brain, if you wanna call it that.

299
00:55:51.036 --> 00:56:03.256
And being able to then enable non-engineers to be able to get, like, a 90, 80/20 rule where they can get 80% of all of the minutiae and all of the skills here without having to be a programmer themselves.

300
00:56:03.716 --> 00:56:10.396
And having that be a reinforcing agent model of, like, building this company brain and strategy and having different things in there.

301
00:56:10.476 --> 00:56:23.076
I'm working on that right now with building own, our own basic internal cloud that runs all of this information that could be behind VPNs and be secure, so it's not just anyone who walked in could be able to get all of the sensitive information.

302
00:56:23.856 --> 00:56:30.316
I think it's time-consuming, and I've talked with a couple other founders and entrepreneurs in the space, and I think everyone's working on the same exact problem right now.

303
00:56:30.336 --> 00:56:34.296
They're trying to figure out how to set up, like, the secure internal open, like, company brain.

304
00:56:34.896 --> 00:56:43.756
And then from there, I think once that's done, you can start making intelligent decisions on incremental spends on intelligence and tokens. The open source models make it a lot more justifiable, though.

305
00:56:43.996 --> 00:56:54.916
Something that costs me 250 bucks just costs me 30 bucks, right? So, um, as the open source models and weights expand and intelligence gets cheaper, w- we can put more effort into our team.

306
00:56:54.956 --> 00:57:04.336
We're pretty lean on being able to get more productivity out of each employee because they have basically a massive team of... It used to be interns, but now it's kind of like coworkers.

307
00:57:04.416 --> 00:57:12.416
Like, it's getting to the point where, like, they are meaningfully intricate parts of the entire organization. So, it's exciting for sure. Happy to come on sometime and talk about that. Yeah.

308
00:57:12.496 --> 00:57:23.596
We'll de- we'll definitely... This will be an interesting, uh, ongoing conversation. Rob Hamilton, thank you so much. Co-founder of Anchorwatch. Enjoy having you on the show. We'll get you back on.

309
00:57:24.036 --> 00:57:34.216
I'm glad we got you to say the word neo clouds. There you go. Uh, welcome to the show. Thank you for having me. Cheers. See you, Rob. See ya. Love Rob. Anchorwatch, folks.

310
00:57:34.856 --> 00:57:45.876
Uh, we had a packed day of deals, so if you liked what you hear, all of these clips will be available on our various social media. Make sure to check out our website.

311
00:57:46.076 --> 00:57:59.016
Sign up for the newsletter, newsletter.blackspacemedia.com. And hit the website for the longer form stories. You can find those also syndicated through channels like Yahoo Finance.

312
00:57:59.696 --> 00:58:13.176
This show is brought to you by CleanSpark, NASDAQ listed ticker CLSK. I'm Charlie. I'm Colin. And we'll see you tomorrow. [upbeat music]