Why AI Demand Is Outrunning Compute Supply
- Gavin Baker has spent the summer asking every AI leader “Can you tell me one quantitative data point in your business that’s getting worse? Just one” — and through July and August, nobody could. AI broadly accelerated both months even as some AI names fell into significant drawdowns; the index calm is misleading because “you can drown crossing a river that’s on average two feet deep.” His caveat: Anthropic is in an IPO quiet period.
- Both men reject zero-sum framing: this cycle is “not an or thing, it’s an and thing,” where frontier labs, open source, neoclouds, applications, and Nvidia all win. Every LP conversation starts with “How is this all gonna go wrong?”, but the supply-side data shows a roughly nine-to-ten-month payback for Nebius ($50B/gigawatt, 50-60% prepaid by customers), Blackstone/KKR/Apollo financing at low cost, and useful lives extending — true equity payback “might be way inside of a year.”
- The demand side is “absolutely nowhere”: the companies’ roughly $180B of revenue rests on maybe sub-10 million heavy users against 1.5 billion knowledge workers. Baker’s fund Atreides grew internal token consumption 100x from March to August; some AI-native companies already spend 10%+ of human compensation on tokens versus ~1% at old-economy firms. Both worry more about undersupply than overbuild through ’28 — Dwarkesh’s scenario of token prices rising 10x is “the opposite direction of where everybody thinks this is gonna go.”
- Public markets will have to digest lab revenue as a dial, not a stream: a lab monetizing 8 gigawatts of inference at ~$60B per gigawatt per year could cut revenue from $480B to $120B in this example by reallocating to training — “and I actually think they would do that.” Satya “blinked” on capex and regrets it; Dario chose bankruptcy-avoidance over share, “and OpenAI was aggressive, and now OpenAI is back in the game.”
- Baker and George argue that the AI industry must tell its own truth: data centers are “probably the best thing that has ever happened to working-class Americans.” Town tax revenue “10Xs,” Loudoun County pairs America’s highest income with its highest data-center density, and cheap natgas ($2-3 vs ~$20-25 in Europe/Asia) is reindustrializing America — while Baker alleges “an organized CCP-funded campaign... laundered through TikTok” against data centers. His favorite Dario line: “stop talking about curing cancer and actually cure cancer.”
- Orbital compute flips on Starship reusability: of $50B per gigawatt, ~$35B is IT either way, while the $15B of terrestrial power/cooling/labor is inflationary — and reusable launch takes the space alternative under $1B. Elon and Jensen have co-designed a Reuben rack targeted for a Q4 ’27 launch; even two quarters late, “that’s 2028,” and per Brad Gershner it’s “happening in plain sight.” Training stays on Earth — latency and speed of light are real.
- The endgame is an ensemble of models behind routers, and the “arbiter of intelligence” abstraction layer is the most vied-for position “in the history of business.” Enterprises will RL open-source base models (soon likely Nvidia’s, via Nemotron and the Poolside acquisition) on their own data rather than hand context to frontier labs; Fireworks Nexus is the best instantiation today; Kirkland & Ellis’s $500M in-house build validates the category but understates the difficulty.
- On Nvidia, Baker’s rule for semiconductor CEOs is “the only thing you should ever say is ‘Thank you, Jensen’” — while George’s rule of thumb is that every 1% of accelerator share is worth ~$100B, so plug into Jensen’s ecosystem rather than tugging on Superman’s cape. Baker estimates Jensen has locked up roughly 70-80% of the supply chain; his data centers are the most financeable ($15B equity on $50B), and George’s deal hierarchy reads true customer preference — equity investments beat RVGs beat token-priced warrants beat naked warrants.
1. Nobody can name a single worsening data point — while AI stocks sell off
- Baker’s standard question all summer: “Can you tell me one quantitative data point in your business that’s getting worse? Just one.” In July and August he found no takers. OpenAI has “clearly accelerated,” open source more so, and Groq saw “a pretty dramatic acceleration” after GroqBot; the hedge he volunteers is that Anthropic is in a quiet period, “so maybe they’ve slowed down a little bit.”
- The disconnect he’s trading around: some AI names in “pretty significant drawdowns” over two months while fundamentals broadly accelerate. Little action at the index level, but “you can drown crossing a river that’s on average two feet deep.”
2. “Maybe everyone wins” — and Anthropic’s pre-IPO gamesmanship
- Borrowing Eric Fisher’s line from the Patrick O’Shaughnessy podcast — “maybe everyone wins” — Baker lists Anthropic, OpenAI, SpaceX, Meta, Google selling TPUs, open source, neoclouds, and inference clouds. George’s LP version: every conversation starts with “How is this all gonna go wrong?”, and his answer is “this is not an or thing, it’s an and thing” — with Nvidia “at the center of all of it.”
- Baker’s hypothesis on Anthropic’s apparent slowdown: they trued up accounting to be comparable to OpenAI on revenue added, tested the waters, and the next disclosure is likely a re-acceleration. Plus checkpoint gamesmanship: Anthropic is “clearly waiting” for OpenAI to release Astra, so Fable 5.1 “magically” ships hours later.
- One culture flag: Anthropic interviews now ask “How would you feel if the equity went to zero?” Baker wants missionaries too, but “you can’t afford the compute you want for your mission if the equity goes to zero.” George’s tag: they’re “the accidental enterprise company” — Baker: “the accidental everything.”
3. Lab revenue is a dial public markets haven’t priced
- The illustration: a lab with 10 gigawatts of power, 8 on inference at ~$60B per gigawatt per year, is a $480B-revenue business (conservative — “people seem to think Anthropic and OpenAI are both monetizing at a hundred billion dollars a gigawatt today”). A research breakthrough could flip 8 gigawatts to training and revenue would fall to $120B — “and I actually think they would do that.”
- The contrast with Meta and Google is that their businesses had no comparable massive cost or infrastructure-to-serve-revenue trade-off; labs face a different dynamic.
- The spending-doctrine scoreboard: Satya blinked — “I know I’m good for my eighty billion” at Davos, then slowed down, and “really regrets that.” Dario reasoned publicly that overspending risks bankruptcy, which is worse than losing share, and was conservative — “and OpenAI was aggressive, and now OpenAI is back in the game.” SpaceX was aggressive too.
4. Sub-one-year paybacks, financed by people who underwrite for a living
- From Nebius — called “Nebulous” once in the transcript — and CoreWeave disclosures: a gigawatt costs ~$50B, customers prepay 50-60%, leaving $25-30B to recover — a nine-to-ten-month payback, faster on spot. SpaceX is faster still because it brings big clusters on quickly; Baker now prices per megawatt, not per GPU.
- His career-level framing: “there haven’t been that many opportunities where you have companies that could deploy tens, hundreds of billions of dollars and get sub-one-year paybacks.”
- On circularity fears: the financiers are Blackstone, KKR, and Apollo at relatively low cost, and one reason is useful lives keep extending while monetization per gigawatt rises — so “the true equity payback might be way inside of a year.”
5. Diffusion is “absolutely nowhere” — and GrokBot is another ChatGPT moment
- George’s demand math: the monetization of these companies, roughly $180B “or something in that direction,” rests on maybe 30 million heavy-paying users — Baker takes the under, George concedes “it might be sub ten” million. Inside a16z companies, the top engineers spend 10-100x the median on tokens; some AI-native firms spend 10%+ of human comp, while old-economy companies doing a good job spend about 1%. Against 1.5 billion knowledge workers, “it feels like we’re nowhere on the demand side, and we’re massively supply constrained.”
- Baker’s own tape: Atreides’ internal token consumption is up 100x from March through August, and Grok Enterprise with two users looks like another 10-20x in a month. His personal test: things that took hours with Claude Code — podcast, Substack, and X summarizers, a sentiment tracker — “each took seven to twelve seconds with GrokBot. And it’s better.”
- The next leg is action-taking: a bot that answers “What are the recommended actions?” from everything the other bots learned. George is “horse racing” GroqBot, Codex, and portfolio company Town on “make me better at my job” — “just wait till everyone does this stuff... it feels like that’s sort of endless token consumption.”
6. Yes, every real technology gets a bubble — but physical constraints are the governor
- Baker concedes the pattern: railroads, steel, autos, radio, internet — “you get a bubble because the markets get really excited... that overvaluation leads to an overbuild,” and debt-funded buildouts “demand immediate ROI,” so “you can’t be off on the timing.” Mitigant: a majority of this buildout is still funded from operating cash flow. He also corrects himself on record — he’d claimed the South Sea Bubble was tied to longitude and sailing; “turns out it was not.”
- The buildout is straining raw productive capacity — wafers, copper (“everybody in copper, there’s an AI thesis”) — with only several million people driving “a crazy global compute shortage. What happens when that’s five hundred million?” He thinks these constraints slowing things down is “actually good for society.” New headwinds: real rates rising (“it just is what it is”) and regulation — “it’s shocking what’s happening in America.” George: “We’re in a really bad place.”
7. The industry has to tell its own truth: data centers are reindustrializing America
- On doomerism, Baker recounts his X exchange with Sholto and Dario: writing one positive and one negative essay isn’t balance when the negative is existential — against Yudkowsky’s “If we build it, everyone will die,” his favorite Dario line is “stop talking about curing cancer and actually cure cancer.”
- The untold story: data centers are “probably the best thing that has ever happened to working-class Americans” — college may now be “significantly NPV negative” versus electrician/plumber/HVAC wages; with behind-the-meter power, town tax revenue “doesn’t double. It, like, 10Xs.” Loudoun County — highest-income county, highest data-center density — is the rebuttal to relocation critics: “we’ve done that. And it worked out really well.” The water objection, in George’s words, is “totally debunked”; Baker calls the water use “nothing.”
- Baker alleges “an organized CCP-funded campaign, I think, against data centers here in America... laundered through TikTok.” Meanwhile George argues that the Strait of Hormuz closure leaves US natgas at $2-3 versus $20-25 in Europe and Asia — a structural manufacturing-cost advantage compounding the data-center boom: “we are reindustrializing America, and it’s awesome.”
- The playbook is Sheryl Sandberg’s: name specific small businesses transformed, such as the Des Moines cake bakery. Every AI company — SpaceX, Anthropic, OpenAI, Nvidia, AMD, Broadcom — should do it: “the truth shall set you free, but only if you tell it.” George’s pushback on the standard message: “we need to stay ahead of China” is “correct but ineffective” — too abstract when voters care about affordability.
8. Undersupply through ’28, possible price spikes, and the compute-inequality trap
- Both take the undersupply side: no capacity is available through ’28 and forecast builds will probably slip on politics. Baker: “Everybody’s worried about oversupply. I’m more worried about undersupply.” Consequence: prices could rise for access to intelligence — Dwarkesh posited token costs up ~10x — “the opposite direction of where everybody thinks this is gonna go,” plausible only because frontier tokens carry enormous user surplus today.
- Baker’s ironic warning to “data center degrowthers”: the consequence may be “real compute inequality, where big companies and wealthy people can afford compute... and it’s like, well, that happened because of you.” Baker also makes the mass-market point that advertising takes years to build, creating a dangerous gap before low-cost products can be broadly supported.
- Open-source myth-busting: open tokens aren’t free — they require roughly the same compute per token as a comparably sized frontier model, with the difference coming from margins charged on top. And the Kimi license stipulates a 30% revenue share (“because it’s open weights, not open source”), while the model is far more token-hungry per task.
9. Orbital compute: “a solved problem” that flips on Starship reusability
- Not the Death Star: a rack of 72 chips roughly airplane-sized, solar wings, sun-synchronous orbit, and a radiator always in the rack’s shadow. Baker’s favorite anecdote: a physics-PhD investor friend insisted it was impossible, visited SpaceX, and said “Well, I was wrong.” His meta-point: your hours of thought versus 10,000 SpaceX engineers with hundreds or thousands of hours each — “it’s a solved problem,” simpler than a Starlink satellite.
- The math: of $50B per gigawatt, ~$35B is IT either way; the $15B of power, cooling, and labor is inflationary on Earth (electrician compensation, copper, materials). With Starship reusability, launch drops under $1B and “the economics just instantly flip.” Kept caveats: you always train on Earth — “speed of light limitations are a real thing” — and terrestrial data centers are “not going anywhere.”
- Timeline: Elon and Jensen have co-designed a Reuben rack to launch in Q4 ’27 — “let’s just say he’s off by two quarters... that’s 2028.” Per Brad Gershner, “nobody’s really paying attention to this, and it’s happening in plain sight.”
- Baker’s framing of the Elon companies is “heads you win, tails you win”: first-party AI caught the frontier fast, and any overbuilt capacity earns a sub-six-month compute payback. Baker stacks the TAMs: Starlink mobile is another $800-900B of wireless, roughly $2T with broadband, plus fast-growing AI ARR, the neocloud, and growing X ads — expect a “Starlink, GrokBot, X advertising bundle,” Google-style. Starbase Louisiana: infrastructure for thousands of launches a year, two per pad per day — probably conservative.
10. The ten-year moonshots: asteroid Psyche, Earth “zoned residential,” Optimus on Mars
- Baker’s most futuristic call: “asteroid mining is gonna be a very real thing” — Psyche holds more gold, silver, platinum, and every precious metal “than exists in the Earth’s crust”; capture it into stable orbit over an American-owned Pacific atoll, work it with Optimus robots, and “delivery to Earth is free.” He pairs it with Bezos’s line: “Earth is going to be zoned residential” — heavy industry moves to space, answering the pollution objections.
- Mars, “at the outside, eight years away”: a fleet of Starships lands, a ramp comes out of the Pez dispenser, “Optimus robots holding American flags” walk down and deploy solar, batteries, and racks of compute — 4K video across Mars, then humans. Bigger than the moon landing, and the century’s frame: “this will be like the age of Elon and Jensen,” who are “fundamentally altering the fabric of human society and civilization.”
11. The ensemble future and the war for the abstraction layer
- Microsoft failed at frontier models — Satya said, by Baker’s recollection, that they would have competitive in-house models roughly 18 months earlier and doesn’t have them — but the world got friendlier: the future is “an ensemble of models” on a Pareto curve behind routers. Even GrokBot, in Baker’s understanding, is Gemini 3.7 Flash, Groq 4.6, and some Opus behind a router, with Elon surely pushing to make it all first-party.
- The enterprise pattern: take a strong open base model — soon likely an Nvidia one, via Nemotron and the Poolside acquisition — and RL/fine-tune on your own data rather than share your context with a frontier lab, which “may be hazardous for your financial health.” Chip companies can fund open training (“it’s trivial to do a fifty to a hundred billion dollar training run for Jensen”), and Baker wonders if Google’s long game is selling TPUs and letting cash flow decide the winner.
- The prize is being “the arbiter of intelligence for global enterprises” — George calls it the most vied-for position “in the history of business,” contested by labs, Microsoft, Databricks, Snowflake, Palantir, inference providers, Fireworks (whose Nexus product Baker calls the best broad instantiation today), Harvey, Legora, Salesforce, and Workday. Kirkland & Ellis’s $500M in-house build is “massive validation of the category” — but it’s not a one-time build. Baker’s retail analogy: run 1,000 clean, well-stocked, well-staffed stores across 50 states and you’re worth $50B — sounds easy, almost nobody in history has done it.
- Baker’s point on Cursor: while everyone else was “creating a digital deity,” Cursor “just wanted to make great product” — the most product-focused at the frontier, now part of SpaceX, suiting Elon’s engineering mindset. Coding is uniquely verifiable and perfectly documented; Baker says the broader knowledge-work pie “is gonna be very messy to go get.” Long-run winner is the low-cost provider — hard without vertical integration — hence George’s hyperscaler lens: EV to net PP&E, “kind of an AI version of price to book.”
12. Nvidia: be nice to Michael Jordan
- Baker’s rule for semiconductor CEOs: “the only thing you should ever say is, ‘Thank you, Jensen.’” George’s rule of thumb is that every 1% of accelerator share is worth ~$100B, so pick a niche and plug into an ecosystem of nine chips — accelerators, CPUs, Ethernet switches, two DPUs, scale up/out/across/in — rather than going head-on. The Jordan game-tape warning: talk trash in game 50 “and he just looks.” Sometimes Superman “just flies away — that’s what happened to the TPU team.”
- Financeability is the moat: a $50B Nvidia data center needs only a $15B equity check, with Blackstone/KKR/Apollo underwriting the rest — and it’s not circularity, because a residual-value guarantee below Jensen’s gross profit is “super NPV positive with very little risk,” plus a revenue share. TPUs are second most financeable at roughly double the equity and higher rates. RVGs also democratize compute against an Anthropic/OpenAI-dominated world — the neocloud playbook again — and Baker asks whether Jensen has locked up 70-80% of the supply chain: fab, DRAM, NAND, lasers, capacitors. Per Dylan at SemiAnalysis, he’s “the central bank of AI.”
- Hardware humility from a scarred semis investor: sometimes the chip comes back from the lab, they plug it in, “and it doesn’t work at all” — and you’re back for another billion. (He corrects George on Cerebras: the chips worked; they lacked product-market fit for two generations.) Credit where due: the “Halapeno” ASIC, as spoken, is the first good internal chip he’s seen outside TPU and Trainium — but it’s competitive with one of Jensen’s nine, and Elon partnering rather than building his own was “a very high Elon move”; George says history will judge it a wise decision.
- George’s closing analytical tool: in a supply-constrained world you can’t infer preference from sell-outs (even old H100s resell high), so read the deal hierarchy — chip-maker equity investments in customers (Amazon and Google’s TPU/Trainium investments in Anthropic: unable to lose if dollars invested are less than gross profit) beat RVG-financed deals, which beat warrants tied to a fixed price per million tokens (good only “as long as the performance of your chip outruns the performance of your stock”), which beat naked warrants that can be negative NPV. On Nvidia’s deals, George says “there’s a reason that people I consider smart are investing in their deals.”
Full transcript
When the history of the 21st century is written, there was the Victorian Age. I think this will be the Age of Elon and Jensen because they are fundamentally altering the fabric of human society and civilization.
What happens if there's a massive supply shortage?
Every time you've had a really profound new technology, you get a bubble because the markets get really excited and get ahead of themselves. Things get overvalued. That overvaluation leads to an overbuild.
One of the things that I think has been correct but ineffective is this idea that we need to stay ahead of China.
You're opposed to data centers. Well, you know what? It's probably the best thing that has ever happened to working-class Americans. We are reindustrializing America, and it's awesome.
Assume that you're right. There's not a physics reason why this can't work.
An increasing fraction of the world's compute is going to be in orbit. This sounds crazy, but asteroid mining is going to be a very real thing. It has more gold, silver, platinum, and every precious metal in it than exists in the Earth's crust.
Every LP conversation that we have starts with, “How is this all going to go wrong?” Gavin Baker has spent the summer asking AI leaders one question: Can you give me a single quantitative data point in your business that's getting worse? So far, the answer has been no.
In this episode, a16z general partner David George sits down with Gavin to take a fresh look at the economics of the AI boom. They discuss why AI may be a positive-sum market where frontier labs, open source, applications, clouds, and chip companies can all win, and what today's compute economics tell us about the sustainability of the build-out. They also tackle the bubble question head-on. Every major technology shift has produced overinvestment at some point. But with compute already constrained and AI usage still concentrated among a relatively small number of people, what happens when that demand spreads across the broader economy? From data centers and reindustrialization to orbital compute, open source, and Nvidia, this is a wide-ranging look at what happens if AI demand keeps outrunning supply.
Gavin, you've been hanging out on the West Coast over the summer, and you've been talking about how you're trying to find someone to give you a bearish case to make your sentiment more negative. Have you found anybody?
1. AI Keeps Accelerating
No, and I ask everyone. My standard question is, “Can you tell me one quantitative data point in your business that's getting worse? Just one.” That's my standard question, and at least in July and August, I haven't been able to find a single person.
Now, if we're being honest, Anthropic is in a quiet period, so maybe they've slowed down a little bit. But I do think the rest of the world has accelerated. OpenAI has clearly accelerated. Open source, I think, has accelerated more. And then I do think Groq, particularly after GroqBot, has had a pretty dramatic acceleration.
And so AI overall accelerated in July, it accelerated in August, and it can't keep accelerating forever. But it's just kind of wild that public stocks have fallen out of bed over the last 2 months. You can drown crossing a river that's on average 2 feet deep, and so there's not a lot of action at the index level.
Right.
But some of these AI names are in pretty significant drawdowns, and they bounced a little bit in August, but they're still in pretty big drawdowns, and things are broadly accelerating.
Yeah.
2. The AI Ecosystem Can Win
Our friend Eric Fisher did a podcast with Patrick O'Shaughnessy, and he said, “Maybe everyone wins.”
Yeah.
Anthropic wins. OpenAI wins. SpaceX wins. Meta wins. Google wins by selling a lot of TPUs. Open source wins. Neoclouds win. The inference clouds win on top of the neoclouds.
Applications win.
Yeah. Maybe not all applications—applications that I think execute well and navigate this. But that feels like a very possible scenario to me, and there's so much zero-sum thinking in the world.
And by the way, on Anthropic, my hypothesis would be, one, I think they probably trued up and cleaned up some accounting.
Yes, definitely.
You'd rather do that.
Yes.
So you rebased.
Yeah.
And now you're comparable to OpenAI.
Yeah, in terms of revenue added. In terms of the definition, and now, I think, kind of revenue added.
Exactly.
Yeah.
So you kind of rebased, and then they did their testing of the waters. Because they've executed well, I would hypothesize that the next disclosure is a reacceleration. And then there's always this kind of funny game between the frontier model companies. They always have more advanced checkpoints. Anthropic is clearly waiting for OpenAI to release Astra.
Yes.
And then it's like—
The next bit will be—
The next day, here's Fable five point one.
Yes, exactly.
Magically, it just happened to be available several hours after GPT-5.
Yeah.
So I think they're being thoughtful heading into this IPO, and everyone is shooting at them.
Yes.
Everybody's shooting at them, and they're in a quiet period, so they can't really shoot back. So there's a lot of gamesmanship, but I do think having OpenAI and Anthropic be public companies is going to be helpful for the market because it's such a powerful force. A lot of public investors hear, “Sarah Friar said this at an all-hands meeting, and it's on the cover of The Wall Street Journal. Okay, we're going to put that into our model.”
Yeah.
And I think it'll be better for them to be public. I am a little concerned. Anthropic is now asking in its culture interviews, “How would you feel if the equity went to zero?”
Yeah.
Because we're looking for people who are mission-aligned.
Yeah, mission, not mercenary.
And that's great. We want missionaries, but we also want people to make money.
Yes.
And at the end of the day, you can't afford the compute you want for your mission if the equity goes to zero.
Yeah. Yeah, yeah.
I'm no expert, but I'm pretty sure on that. And then I do think they are—
They're the accidental enterprise company. Enterprise is just a byproduct of the mission, the objective at the end.
Oh, for sure. They're kind of the accidental everything.
Whereas I think OpenAI is a little more commercial, and obviously SpaceX is a little more commercial.
3. Labs Trade Profits For Training
But all of these companies, let's just say they have 10 gigawatts of power, and they're allocating 8 to inference. Let's say they're monetizing that inference at $60 billion a year. So that's $480 billion a year in revenue, right?
Which, on a revenue-payback basis, would be a 1-year payback on a revenue basis, not a gross-profit basis.
Yeah, on a revenue basis.
Yeah.
Yeah. And I'm trying to use conservative numbers. People seem to think Anthropic and OpenAI are both monetizing at $100 billion per gigawatt today.
Yeah, yeah.
Let's say they have a big research breakthrough, and they decide, “Wow, it is to our long-term advantage to go from 8 gigawatts allocated to inference and 2 gigawatts allocated to training to 8 gigawatts on training.” Then your revenue just went from $480 billion to $120 billion, and I actually think they would do that.
They would make that decision, yeah.
And this is just something that public markets are going to really have to get used to.
Yeah.
As you say, OpenAI may be a different animal, and I do think the reality is that everybody has these ideals about how they're going to manage their business. Then they go public, and the stock is volatile, and it really impacts employee morale, recruiting, and retention. So I'd be surprised if they did such a dramatic cut.
But a lot of the revenue is kind of under their control based on what checkpoint they release—
Yeah.
—where they price along this kind of Pareto curve, and then how much they allocate between training and inference. So it's going to be Meta and Google and these kinds of internet companies. It was pretty smooth fundamentally—
Smooth enough, yeah.
—even if the stocks were volatile.
Well, there was no massive trade-off they had to make in terms of the cost or infrastructure-to-serve-revenue side. They were totally separate.
Yeah.
A hundred percent.
Yeah.
It's fascinating. If you go back to Eric's point that it's all going to work, I actually think that's a great point. I describe it differently. I've had this conversation with LPs a lot, because every LP conversation that we have—it's probably the same for you—starts with, “How is this all going to go wrong?”
Yeah, it's a bubble.
And it's like, “What's going to crash?” The large model is screwed, or the lab is screwed because of open source. And I'm like, this is all wrong. This is not an or thing; it's an and thing, right? This is an and thing.
Frontier's gonna work really well. N-1 models are gonna work really well. Open source is gonna work really well.
There's gonna be a bunch of application companies that work really well. The clouds are probably gonna be fine. They're probably gonna work really well. The 5 lab companies are probably gonna do really well.
Yeah, and NVIDIA is cent—
It's the center of it all.
…is at the center of all of it.
Yes, yes.
Yes.
They're probably gonna do pretty well.
Yeah. The last 26 years have taught me not to bet against Jensen.
Yeah, he's in a pretty good position here. I wanna come back to that. The point that you made about training versus inference is an interesting one. It seems to me like the labs will decide to take all incremental profits, and probably much more than their profits, and invest them in training for a long period of time. Would you think that's fair?
It's very different than the clouds, because the internet companies and the clouds just end up being supply-demand driven, and they generate tons of profit, and they can still grow a certain amount. But they don't have some—maybe with the exception of Meta—big, long-term bet that's like a multiyear payoff.
Yeah, I think it's important to be precise. I definitely don't think they will generate free cash flow anytime soon. I think they're gonna generate a lot of operating cash flow, and then they'll use that to buy a lot of GPUs, XPUs—whatever we're gonna call them.
Or maybe they subsidize heavily. We do know that that's happening at the labs.
Subsidize what heavily?
Their first-party products.
Oh, yeah, yeah.
So, token consumption of their first-party products.
Oh, yeah, yeah, yeah.
So they're doing all this research, and they're spending a lot on data and compute.
Yes.
And their first-party products are heavy-subsidy products today, right?
Yeah, so it's 8 gigs of inference—
Yes.
…and 2 gigs is for internal research.
Yeah.
And then 2 gigs is actually training.
Yeah, exactly.
Including probably the inference that goes into post-training. Yeah, I don't… I think, given the belief systems that they all seem to have about scaling laws, which continue to hold, I don't think any of them are gonna be that focused on generating free cash flow. And you've seen—right, we saw Satya blink.
Yes.
And Satya really regrets that, I think. He kind of blinked. I think it was last year, you know, he gave that great interview at Davos, and they asked him about all the CapEx, and he said, “I know I'm good for my $80 billion.”
Right.
And I think they blinked a little. They slowed down. They regret that. And Dario famously went on a podcast and said, “Listen, some people are being super irresponsible with their spending, and it's a hard decision because if you don't spend enough, you could lose a lot of share. But if you spend too much, you could go bankrupt, and those are both bad things, but bankruptcy is worse than losing share, so I'd rather be conservative.” And he was conservative.
And now—
And OpenAI was aggressive, and now OpenAI is back in the game.
And xAI was aggressive.
And SpaceX was aggressive.
And so there are clear, high ROIs on those, independent of the supply-demand mismatches that are happening. Clearly, that seems to be the right decision, short term and long term.
Yeah, absolutely. We calculated that Nebulous and CoreWeave both gave some interesting disclosures. You can kind of get to a 9- to 10-month payback for Nebius because, okay, you bring on a gigawatt, it costs $50 billion. You can get an upfront payment for 50% to 60% of that from customers.
Yes. Yeah.
So now you're talking about $25 or $30 billion, and then you can monetize it if you put it into the spot market—
The spot, yeah.
…at spot. Paybacks are probably much faster than 9 or 10 months on that basis.
Yeah, you could assume a smoothed-out level, like $2 or $3. Even with that, it's a very high payback.
Yes, it's a really good payback.
And now you can get, like, $5 or $8, and yeah.
And then SpaceX, because they build these really big clusters, and I think a really important point is they bring them on fast.
Yes.
They have an even faster payback. I've tried to shift to thinking of pricing per megawatt rather than per GPU—
Yeah, yeah, yeah.
…because it seems like that's where the world is. But xAI, the payback feels well inside of that.
Yes.
And in my career as an investor, there haven't been that many opportunities where you have companies that could deploy tens, hundreds of billions of dollars and get sub-one-year paybacks. It's kind of crazy.
And then also, particularly if you're buying NVIDIA GPUs—to a lesser extent, TPUs—you can finance these.
Yes.
And there's a very sophisticated—
Yeah, it's a very low cost of capital to finance them today.
Yeah, and everybody's worked up about circularity, and it's like, well, I don't know. I know a lot of smart people who work at Blackstone, KKR, and Apollo, and they're the ones that are financing it.
They're the ones who are financing it at a relatively low cost, right?
At a relatively low cost. And I think one reason that's happening is useful lives just keep getting extended. As these models get better and better and better and the ROI on token spend goes up, the monetization rate per gigawatt goes up. So the true equity payback might be way inside of a year.
Yeah, exactly. Exactly. And look, there's a case you could make that the prices actually of all this stuff go up, which could make the supply-side economics even more compelling, right?
So on the supply side, that's the dynamic today. It just is what it is. There's a ton of data points out there that paybacks are within a year.
Yeah.
4. AI Demand Has Barely Started
I think it's actually interesting to think about the demand side too, because the knock would be, well, in all these cycles you get some overbuild, and then that destroys the economics of the supply side.
The demand side today—what are we monetizing? The monetization of these companies, which are doing, call it, a hundred and eighty billion of revenue or something in that direction, is on the back of what? Like 30 million actual heavy-paying users getting real value. I'm talking about developers. It—
I might take the under on 30 million, man.
Yeah, actually, what we see inside our companies is that, obviously, there's a power law in which companies are spending a lot on tokens. Old banks are probably spending 1%. Very tech-forward companies are spending high single digits.
But if you actually look at the power law of what's happening with the actual engineers in those companies, the highest-spending engineers are spending 10 or sometimes 100 times more than the median engineer. And so, yeah, your 30 million is probably way overstated. It might be sub-10 million.
And so there's this question of where are we at in diffusion? There's 1.5 billion knowledge workers. It feels like we're nowhere on the demand side—
Yeah.
…and we're massively supply-constrained.
And what are—I’m just curious: across the a16z portfolio, what are your best companies spending on tokens per month relative to human compensation? What rough range?
Oh, high single digits, some at 10%. Some of the very AI-native ones are at 10% plus. And then old-economy companies—the ones that are probably doing a good job—are spending, like, 1%.
So it feels to me like, when I look at the supply-demand characteristics, supply-side stuff—people say, “Is that sustainable?” Well, when you pair it with the demand stuff, it feels pretty simple. There could be things that disappoint us in terms of diffusion into the real economy. But over a 10-year stretch, it feels like we're nowhere, right?
Yeah, absolutely nowhere. At Atreides, our internal token consumption has gone up 100 times from March through August—100 times our token spend.
We just got access to Grok Enterprise, and with 2 people using it, it looks like token spend goes up 10 or 20 times in a month from August.
Yes.
I think—
But it's actually extremely valuable. We have some heavy Grok Enterprise users here, and it is very productive use.
This is not wasteful token spend.
Yeah. And listen, I try super hard. When I use AI, I always remember when I was trying to get my parents to shift to an iPhone and an iPad, get them used to it. They did a good job, and I give them loads of credit. But I’m 50 years old—how old are you, David?
Forty-two.
And you see these 23-year-old kids, and the way they use AI—they’re just fluent and native in it. I feel like maybe, no matter how hard I try, I will never be as fluent and native in it, and I’m trying really hard. We got Claude Code, and I built some stuff and did some cool stuff. Then, in three minutes of creating GroqBots, I had much better versions of everything I created.
I went on Patrick O’Shaughnessy’s podcast about 5 months ago, and I said, “I love having a podcast summarizer.” Everybody asked, “How’d you do it?” I said, “Just use AI and do it.”
Yes. Pretty simple, yeah.
It takes ten seconds in GroqBot.
Yes.
It’s amazing, and it’s so good.
Yeah.
Then there’s a Substack summarizer, an X summarizer, and an X sentiment tracker for topics and stocks.
Yeah.
All of those would have taken me hours working with Claude Code, and they each took 7 to 12 seconds with Grok.
Yeah.
Yeah, yeah, yeah.
Yeah.
And it’s better.
Yeah.
So, to me, Grok does feel like another ChatGPT moment, at least for me. With Claude Code, I could see in the data that it was powerful. I did some really cool stuff with it that was empowering.
Yeah, yeah, yeah.
And this is neat—family calendar apps, things like that.
Yeah.
But this is just 10 seconds, and it’s way better than what I was able to do.
Yeah.
The Claude Code thing was obviously the shift in coding. Our most sophisticated engineers went from doing 20% of their code with AI to 90%-plus.
Yeah.
So now I think everything you described that you built with Claude Code or Codex is still kind of reactive, in a way, right? It’s still summarizers and preparation. It’s all knowledge-enhancing, which is part of your job, but it’s not actually doing the work for you.
You have a GroqBot that says, “What are the recommended actions?”
Yes, exactly. I’ll go do that.
Based on everything the other bots have learned today: “What recommendations do you have for me today?” That, for sure, is a different thing. It was so easy to build.
I now have it. I’m horse-racing all of these: I have GroqBot doing it, Codex doing it, all the action-taking for it.
Yeah, yeah.
Because I want to know: make me better at my job. Look at everything I do. Give me recommended automations you can do. I have Town doing it as well, which is one of our companies that has been very good at it. We’re kind of on the bleeding edge of trying to do this stuff.
Yeah.
Just wait till everyone does this stuff.
Yeah.
And then, when we actually click, “Yes, go just automate this,” it feels like that’s sort of endless token consumption.
Yeah. But we should acknowledge the history of financial markets, dating back to the South Sea Bubble. Whenever you get this transformational new technology, I actually went on a podcast and said I thought the South Sea Bubble was connected to the invention of longitude and the ability to sail. It turns out it was not. It was more of a Tulip Mania episode.
But every time you’ve had a real, profound new technology—whether it’s the automobile, television, radio, the internet, the PC, railroads, or steel mills—you get a bubble because the markets get really excited and get ahead of themselves. Things get overvalued. That overvaluation leads to an overbuild, particularly if you’re funding it with debt. Even today, a majority of this is still being funded out of operating cash flow, which I think is really helpful.
Debt-funded build-outs demand immediate ROI, not an ROI in 3 years.
Exactly.
You can’t be off on the timing.
Yeah, yeah. You can’t be off on the timing.
But I’m more concerned about this. I talked to Jazz and Patrick about how Watson wafers are these fundamental constraints, and the build-out is so big, and we’re so early that it’s impacting the raw productive capacity of so many industries.
Yeah.
Now everybody in copper has an AI thesis.
Exactly.
We’re going to have to think about how to fill the gap. If 10% of what we just talked about comes true, we’re in this acute shortage, with several million people driving a crazy global compute shortage. What happens when that’s 500 million? How many copper mines do we need to build to support this?
Yeah, yeah, yeah.
It’s kind of a wild thought. These fundamental constraints are slowing us down, and I actually think that’s good for society. I would now say rates and regulation—real rates are going up.
Yes.
It is what it is, which makes sense because we’re investing a lot. It makes sense that real rates are going up. And then regulation—it’s shocking what’s happening in America.
We’re in a really bad place.
Yeah. I had this exchange with Schulto from Anthropic and Dario on X last weekend. Dario said, “Hey, I don’t think I’ve been negative. I’ve written 2 essays. One was positive, one was negative.”
Being 50% negative is particularly significant when it’s a terrifying negative.
Like an existential—
An existential negative. Everybody might be out of a job. Eliezer Yudkowsky says, “If we build it, everyone will die.” How about if we build it, we’re going to cure cancer? We’re all going to live forever.
One of the best things Dario said was, “What we need to do is stop talking about curing cancer and actually cure cancer.”
And actually cure cancer and actually make breakthroughs, like—
But somebody needs to tell that story. My favorite line in the Bible is, “The truth shall set you free.” The only group that can tell the AI industry’s truth is the AI industry. They need to just start telling the truth.
5. Data Centers Reindustrialize America
If you’re opposed to data centers, well, you know what? It’s probably the best thing that has ever happened to working-class Americans. Going to college might be significantly NPV-negative now, because you can go learn how to be an electrician, a plumber, or an HVAC tech and make ungodly amounts of money.
Yeah.
So this has been amazing for working-class Americans. We now have a lot of data that, particularly with behind-the-meter power generation, when a data center goes in, it transforms a town. Tax revenue doesn’t double; it 10Xs, and it is revitalizing all these dying small towns all over America.
We’re getting much better at addressing the environmental issues. Generally, they use natural gas, which is a pretty clean fuel.
The water-consumption thing—
The water is nothing.
It’s totally debunked. It’s totally debunked.
It’s nothing.
Yeah.
So these are really, really, really good, and they’re having a really positive impact on the world. That’s without even considering things like curing cancer, but somebody needs to tell that story.
And I think the problem now is that the burden of proof is on not just talking about curing cancer, but actually delivering some real, tangible, everyday American benefits beyond using ChatGPT or Grok to answer your questions or substitute for a search engine, right?
It does feel like we’re pretty close to that.
Yeah, it does. And, by the way, one of the things that I think has been correct but ineffective is this idea that we need to stay ahead of China. It is true.
Yeah.
I'm very much a patriot. I believe that.
Of course.
But it's way too abstract.
Yeah.
The abstract, for the average American, doesn't do anything.
Nobody's worried about China invading America.
Yeah, exactly.
I'm pretty sure the Pacific Ocean is really big.
Yeah, what they care about is affordability and how this is going to change my life for the better first.
Yeah.
Right? And so I think there's a pretty immediate impact you could feel. My favorite is Loudoun County, Virginia, which is the highest-per-capita-income county in the US.
Yeah.
And it has the highest density of data centers.
Yeah.
And they make a tremendous amount of tax revenue from data centers. We should do this everywhere.
Yeah. Well, it's actually very funny. Someone very opposed to data centers said, “Oh, you're for data centers. I'd like to see them put in the highest-income zip code and the highest-income county.”
Yeah, yeah, yeah.
And they're like, “Actually, the highest-income zip code in America and the highest-income county have the highest per-capita concentration of data centers.”
Highest proportion, highest proportion of data centers.
“So we've done that.” “And it worked out really well.”
It worked out really well.
Yeah, but, you know, “Hey, don't bother me with the details.”
Yeah, exactly.
“I'm onto my next talking point.”
That's good. That's good.
And all those talking points—it's tragic. There is an organized CCP-funded campaign, I think, against data centers here in America. I think a lot of it gets laundered through TikTok, and it's just tragic because the other thing that's happening is this is reindustrializing America. The combination of having the Strait of Hormuz closed, which is amazing for America—
Yeah.
—you know, natural gas here—
Yeah, of course.
—is 2 or 3 bucks. It's now 25 bucks—
If you want it, yeah, sure.
—in Europe and Asia, or 20 bucks—
Yeah.
—or whatever it is. And natural gas is an important input to the cost of electricity, which is an important input to almost all manufacturing—
Yeah.
—processes. And so we have a huge cost advantage for that basic input now, and you have that happening, and you have this kind of data center boom happening. We are reindustrializing America, and it's awesome. This is what everyone in both parties has wanted for a long time.
Yeah, exactly.
Bring industry back. Small towns that were left behind by the steel mills closing—well, data centers are bringing them back.
Yeah.
But somebody has to tell that truth. I mean, I try to do it on every podcast, but I'm just a dude.
Yeah. And your audience is the tech audience that already believes you. You're preaching to the choir, if you will. But, yeah, the story—Meta's probably doing the best job of telling that story, I would think.
Yeah.
It seems.
You know, I think one reason is it's really wired into Meta's DNA. One of the first things they started doing as a public company—I don't remember if it was on their first earnings call—but Sheryl Sandberg would run through 10 or 15 very specific small businesses that had started using Meta's advertising products and the impact it had on that business.
Yeah.
This cake bakery in Des Moines started working with Meta, and it was 2 women who were single mothers working by themselves. Now they have 15 locations and employ 50 people.
Yeah.
This has been amazing for Des Moines, and it's been transformative for them.
Yeah.
And they would just run through that every time. I do think the entire AI industry—I'd love to see everybody, SpaceX, Anthropic, OpenAI, Google, and Meta, say, “Hey, here are real businesses and real Americans.” Either name the business or, if you can, give permission to name the American, or anonymize it. This is a really positive thing it did—
Yeah, of course.
—it had on their life.
Already very tangible, yeah.
Yeah, same with Nvidia, AMD, and Broadcom—all of them. Just run through specifics, because the truth shall set you free, but only if you tell it.
Yeah, exactly.
Yeah.
Exactly. Yeah. So it seems more likely, then, given that fact pattern, if you go back to just the sort of macro situation that we're in, that we underbuild on the supply side through 2028. And, by the way, there's no capacity available with all the forecast builds that will happen through 2028, which are probably now going to be delayed given the political dynamics that we have.
Yeah.
So—
Everybody's worried about oversupply. I'm more worried about—
Being massively—
—undersupply.
—massively undersupplied. Exactly. Okay, so then if that's the scenario, you could see a scenario where you see big price increases—
Oh, yeah.
—actually to access the intelligence.
Yeah, well—
Which is the opposite direction of where everybody thinks this is going to go.
Yeah. Well, Dwarkesh Patel had a wild point. I forget what it was, but he was positing—
The cost of a token could go up 10X or something like that.
Yes.
Yes.
Yeah.
Which is crazy, but we do live in a supply-and-demand world. It's conceivable if the demand goes massively up. And, by the way, the whole premise of what's happening so far is that there's a massive amount of consumer or user surplus being generated, right?
Yeah.
So why do people select the frontier tokens when they could use the cheaper tokens to do most tasks? There are many reasons why, but the biggest one is because there's a tremendous amount of surplus, even if you're using the frontier tokens.
Yeah, absolutely.
And so what happens if there's a massive supply shortage?
Well, I think the kind of funny consequence of these data center degrowthers may be real compute inequality, where big companies and wealthy people can afford compute. Then, 2 years from now, they'll be on about that, and it's like, “Well, that happened because of you.”
Yeah. Yeah, yeah.
That happened because you wouldn't let us build data centers.
Yeah, and, by the way, we've seen this, right? The path to a low-cost product delivered to consumers in a mass market is advertising. It takes a long time to build an advertising business, as we've seen with all the consumer internet businesses that we've invested in over the years. And so there may be a disconnect during the period when you can't actually offer that.
Yeah.
And that would be a terrible outcome.
That'd be a terrible outcome for the world. Nobody wants that, so we need to build a lot of data centers.
Yeah, exactly.
Yeah.
Exactly. Yeah.
A compute-inequality future—that's not a good future for anyone, which is another reason open source is so important. One of the things—I had Grok make me a meme of that 3-headed dragon, and one of the heads is kind of confused about all of the really stupid bearish AI narratives. But people have this idea that open-source tokens are free.
They're not.
And it takes the exact same amount of compute—
Yeah.
—to make an open-source token as a frontier token for a comparably sized model. Now, there are a lot of nuances there, but that's broadly true. It's just a question of what margins are—
Yeah, of course.
—charged on top of that.
That are being captured, yeah.
And even then, something that I don't think a lot of people appreciate about the Kimi license is that it stipulates a 30% share of any revenue.
Yeah. Yeah, yeah.
So Kimi has taken a 30% cut of all the revenue generated on its—
And this is because it’s open weights, not open source.
Yeah, exactly. Yeah.
Yeah.
But it’s also extremely token-hungry too, right?
Oh, yeah.
So it’s far more token-hungry. Even if we’re talking on a token basis, on a task basis, it’s far more inefficient.
Absolutely.
And so it’s very costly.
Yeah. I always like Jensen. He’s a great patriot, a great American. We’re so lucky to have him and Elon. I think when the history of the 21st century is written, there was the Victorian age. I think this will be the age of Elon and Jensen.
Yeah.
They are fundamentally altering the fabric of human society and civilization with AI, SpaceX making humanity multiplanetary, and Starlink bringing low-cost internet access to the poorest communities in the world, which is amazing. It’s something people don’t talk about, but it’s an amazing surplus. You talked about consumer surplus; that is an amazing surplus.
There was never going to be an economic case to build internet access in those places because of the cost and the willingness to pay, and now you can.
Yeah.
Any incremental internet capacity is not going to be built in a traditional sense on Earth.
Yeah.
It’s going to come from space, and so—
Well, yeah.
That is a huge unlock. I agree.
It’s a good thing, but we should all be grateful for them because I do think they’re making the future as exciting and as inspiring as possible.
6. SpaceX Moves Compute Into Orbit
Let’s say we are in this supply crunch. It’s so funny whenever I talk about SpaceX, which is obviously near and dear to both our hearts. I say, first of all, the orbital data center stuff is not like big buildings in space. It’s helpful to actually think of it as—
Yeah.
—the size of an airplane.
Yeah.
Yeah, it’s like a big rack—
People are picturing, like, the Death Star.
Yeah, exactly. It’s not that.
Or the Pentagon—
Yeah, yeah.
—floating around in space. That’s not what it is at all.
Yeah. It’s, you know, whatever, the size of an airplane, right? A rack of 72—
Yeah, but even—
—whatever chips, whatever.
Yeah, it’s like 5 of us standing together is roughly the—
Yeah, and the airplane is the wings—
—and then you have the solar—
The solar arrays.
—solar wings.
Yeah.
Then you keep it in a sun-synchronous orbit, so you have the radiator—
Yeah, the back end, yeah.
—that’s always in the shadow of the rack. That’s how you cool it. I can’t—it’s very hard for me to engage. There are all these people on X, and they’re like, “I am a physics PhD, and this is impossible.”
There’s a friend who’s another investor and actually is a physics PhD, whom I had many arguments with. He’s like, “I am a PhD, and this is impossible.” Then he goes to SpaceX Day, talks to the SpaceX engineers, and he’s like, “Well, I was wrong.”
If, let’s say, you’re an astrophysics PhD, you’re brilliant, and you’re hanging 100 IQ points on me, have you thought about this for 1 hour? Have you thought about it for 10 hours? Have you thought about it for 5 hours? You have 10,000 of the world’s smartest engineers at SpaceX who’ve thought about this each for hundreds, if not thousands, of hours. The sum of that, working with very sophisticated engineering tools, is that it’s a solved problem, and in their minds, it’s dramatically simpler and easier—
Yeah.
—than a Starlink satellite because the Starlink has to have the phased arrays and move around.
Yeah.
Yeah.
I think—so, okay. Assume that you’re right. There’s no physics reason why this can’t work. Cost-wise, it seems really imposing, but the history of the Elon companies is that the cost curve gets dramatically better. When we first invested in SpaceX, Starlink was not commercially available, and we had all these questions about how the economics would proceed over time—the same on the launch side, the same with the Model 3. I just have to think that will get solved, paired with the fact that we’re going to have massive undersupply, self-inflicted, on Earth.
Yeah.
It feels clear to me that, at a minimum, it will be swing capacity.
Yeah.
In the fullness of time, maybe it will be larger.
Well, no, it’s really simple. The question people should be asking about orbital compute, which is the one SpaceX is focused on, is Starship reusability.
Because the math is: let’s just say it’s $50 billion per gigawatt, and let’s just say $35 billion of that is IT. So that’s the same, and maybe it grows a little because it’s going into space. The rest is power, cooling, labor, and all sorts of things that you don’t need in space because you have the solar panel and the big radiator.
Yes.
And that’s $15 billion, and it’s probably inflationary here on Earth—
Yeah.
—because labor fundamentally feeds into that. We just talked about what’s happening to electricians—
Yeah, compensation.
Yeah.
Materials are all going to go up.
Yeah, all of it. We’re going to run out of copper. The copper bulls are focused on copper shortages.
Yeah, optics. Yeah, all of it.
So that $15 billion is inflationary, and what you have to compare it to is the cost of launch. With Starship reusability, that goes to under $1 billion, so the economics just instantly flip.
Now, you’re always going to train on Earth. There will always be advantages to having GPUs right next to each other. There are speed-of-light limitations; that’s a real thing. Latency matters. So data centers on Earth are not going anywhere. I think they’re going to continue to be very, very valuable, but an increasing fraction of the world’s compute is going to be in orbit. Elon said that he and Jensen have co-designed—
Yeah.
—a Reuben rack, and it’s going to launch in the 4th quarter of 2027.
Yeah.
And let’s just say he’s off by 2 quarters.
Yeah.
I mean, that’s—
Still fine.
—that’s 2028.
Yeah.
Yeah.
That’s still okay. That’s pretty soon.
As Brad Gershner says, nobody’s really paying attention to this, and it’s kind of happening in plain sight.
Yeah.
And it kind of, to me, solves for something—
Trust. Yeah.
—mid-single-digit billions today, which, by the way, is just like keeping share constant of what’s happening with coding.
Not presuming it takes any share from Grok bot. Yeah.
Yeah, from $3 billion. And by the way, man, I’d probably take the over with Grok bot.
Yeah.
I bet it’s changing by the day, just based on my own usage.
Yeah. Yeah.
And the number of people who are hitting their usage limits—you’re starting to get messages from GrokBot like, “Hey, our servers are overloaded,” every once in a while.
Yeah.
And they have a lot of compute. So it’s just like, okay, you don’t want to debate orbital data centers. No problem. Starlink Mobile has a pretty clear, credible plan for how that’s going to work, and wireless is, call it, another $800–$900 billion of revenue that they can address.
So your mobile plus your broadband, whatever it’s called, is close to a $2 trillion market.
And then you have a really rapidly growing AI ARR base.
Yeah. AI ARR. You’ve got the cloud, you know, the sort of—
Yeah, the cloud—
—the neocloud business.
Yeah. So I don’t think—great, you’re an orbital compute skeptic. No problem. It doesn’t matter.
Yeah, exactly.
We don’t even need to—we can just look at things that are happening today with—
Yeah.
—terrestrial compute, with Cursor, with Grok, with GroqBot. By the way, I think X ads are—
Yeah, they’re going up.
Yeah.
They're also growing. You know, I would expect at some point you'll have a Starlink–GrokBot–X advertising bundle. One of the ways Google built their cloud business is they bundled it with ads. Maybe you're bundling the ads with AI, but why not do that?
Yeah.
I actually like the AI position they're in because it's heads you win, tails you win, in the sense that their first-party business is growing very fast, and they caught up to the frontier very quickly.
Yeah.
They've made very aggressive compute investments to enable that first-party work.
Yeah.
And that's the kind of heads-you-win, tails-you-win setup. Say they overbuilt their capacity for what they need for inference or training, they have a—
Sub-six-month payback.
—a very compelling six-month payback—
Yes.
—on the compute side, with massive scarcity of supply. And so I think that's a really good setup.
And there was a bear case that, okay, in an OpenAI–Anthropic maximalist view, where they're the only 2 companies and they're designing their own chips, what's the room for anyone else? Well, I don't think they're going to have a reusable Starship and multiple spaceports anytime soon. If the economics of compute are such that orbital is where it increasingly makes sense going forward, because Starship should be deflationary, while terrestrial cooling and power should be inflationary, even in a world where they fumble the ball with their first-party AI applications, they do still have a—
Yeah, then they're a massive infrastructure business.
Yeah.
Yeah, I'm so fired up about Starbase Louisiana.
Oh, yeah.
It sounds so cool. I was reading about it last night, and it's sort of like—they now have the infrastructure for thousands of launches a year.
Yeah. And eventually, I think you will see these Starbases in multiple places, on multiple coasts all over the world. At some point, you'll probably see one somewhere in the Middle East. You'll see whatever European country is the least bureaucratic at the time. You'll see one there. For sure, I think you'll probably see one in Japan or South Korea.
Yep.
Who knows?
Yeah.
Yeah.
No, it's pretty exciting.
Yeah.
The capability to do—call it 5,000 launches a year—that feels very futuristic.
Yeah. I mean, it's wild. And I do think a distinction that SpaceX really tried to hammer home during their IPO is that there's a difference between reusability. In China, they did catch a rocket using this—
Yeah.
It was actually kind of ironic. It was this jury-rigged system of wires—
Yes.
—that had actually been suggested on the SpaceX subreddit—
Yes.
—before they landed the first Falcon 9. So it was more than 10 years ago. China's clearly paying close attention—
Yeah, exactly.
—to the SpaceX subreddit. But that's very different from catching that thing from what they're trying to do with Starship, where the booster gets caught with the things, then it gets moved, the Starship gets caught, then it gets stacked—
Yeah.
—it gets fueled, and it's sent right back up.
And ready to go. 2 launches a day per pad—those numbers add up pretty fast.
Yeah. And I do think they're engineering the pads for more than 2 launches a day.
Yeah, I think that's a conservative assumption.
What's the most futuristic thing that you think about with SpaceX? You and I were at this conference together, and there was this whole debate among a small group of public investors about what's going to be the first $10 trillion company. I think what you said was, “I have no idea, but I know which one's going to be the first $20 trillion company.” What's the most futuristic product, market, or technology thing about SpaceX that you can think of?
Look, this sounds crazy, but asteroid mining is going to be a very real thing. We're going to capture—there's asteroid Psyche. It has more gold, silver, platinum, and every precious metal that exists in Earth's crust. At some point, particularly with Starship, you will be able to capture these asteroids. We may need that lunar base to make this happen.
You'll bring them into a stable, geosynchronous orbit over some American-owned atoll in the middle of the Pacific. No humans within 50 miles. You can imagine Optimus robots—
Yeah, doing the work.
—doing the work. And then delivery to Earth is free. For sure, some of it is going to burn up, but I think that's going to happen.
Jeff Bezos said something very interesting. He said, “I think in the future, Earth is going to be zoned residential.” Somebody asked him—
Mm.
—this was about 15 years ago—“What do you mean by that?” He was like, “All heavy industry will take place in outer space.” And then this addresses the pollution concerns.
Yeah.
It addresses everything. People always get really worried about, “Oh, will we still be able to see the stars?” I think it's hard for the human mind to understand how big space is.
Yeah. Yeah.
How big outer space is.
Yeah.
Yeah. So I think that is—
That's probably the most futuristic thing.
But in terms of an economic application, I do think in the next few years you're going to have a fleet of Starships land on Mars. By “the next few years,” I mean, I don't know. At the outside, let's say this is 8 years away.
Yeah.
They're going to land on Mars. A little ramp is going to come out of the Pez dispenser, and it's going to be a modified Starship, the Mars Colonial Transporter, and it's going to be wild. You're going to have Optimus robots holding American flags walk down, and then they're going to pull out a bunch of solar panels and batteries and racks of compute, and they're going to set all of that up. They'll be deploying Starlinks, and maybe the orbital mechanics don't allow this, but I think they'll figure out a way to provide capacity.
Just think how crazy it is to watch the views from Mars Pathfinder, or whatever these different Mars rovers are, and get 4K video through Optimus robots all over Mars. After that, there will be humans.
Who can inhabit it. Yeah.
Yeah.
That is crazy to think about.
And that's going to be an amazing moment for America. Think about the moon landing.
Yeah.
You know?
This is a little bit bigger.
Yeah.
Yeah, that's a good one. There's not a lot of chatter about that one out there.
Yeah, but I think it's highly likely to happen.
Yeah. So you mentioned Microsoft and the bet that they made. Apple made the extreme bet against the future, and Microsoft is kind of a gradient of that. What's your outlook for their decisions?
Well, I do think the world has gotten a lot friendlier for their strategy. They clearly tried to make a frontier model. They failed. Satya said, “We're going to have our own models that are very competitive.” I think he said that 18 months ago. They don't have their own models that are competitive. But I think the future is an ensemble of models. There's a Pareto curve. No one model is going to be the best at everything.
And I think the future, certainly, for the global 1,000 biggest companies is that you're going to take whatever the best open-source model is. I think, probably in the very near future, that's going to be an NVIDIA model.
Yep.
The labs making ASICs create very interesting—
Incentives for each—
Incentives—
—to get into each other's business. Yeah.
Incentives for Jensen and everybody: “Well, in a world where open source wins, who funds the training?” Well, the chip companies could fund the training.
Yeah.
It's trivial for Jensen to do a fifty to a hundred billion dollar training run. Maybe soon. I do wonder if this is Google's super-long-term play. They seem to have opted out of the frontier race for now: “We're going to monetize our compute at high rates, and we're going to sell TPUs externally.”
But that generates so much cash flow, and open source is getting closer and closer and closer to the frontier. It may be that the winner is ultimately whoever has the most cash flow to fund these big training runs. I do think you're going to see American open-source models, led by NVIDIA, get really close to the frontier.
Yeah.
They paid for that Poolside acquisition for a reason.
Yeah.
Poolside actually had a lot of really good American open-source talent. I think they're doing a lot of smart things, but that is really good for Microsoft and, at some level, almost every application-software company, because what you can do now is take a base model. Nemotron, to date, has not had a lot of post-training. It's been a—
Yeah, exactly. Yep.
—good pre-trained model—
Yep.
—that you could do what you want with.
So if you take a really good pre-trained base model, then instead of sharing your own enterprise context—which is truly your IP, truly the value of your company, the context embedded in all of your data—with a frontier lab, that may be hazardous for your financial health.
Yeah, certainly with the shift in the ZDR policy, yes.
Yes. And so you take a really capable open-source model, and you do a lot of RL and supervised fine-tuning on your own data, so you own it and it's your model.
Yeah.
And then, if intelligence is a super-important input into your business, you want to own and control your intelligence—its capabilities, its cost, and then what we've seen from a lot of companies. GroqBot, my understanding is, I think it's Gemini 3.7 Flash—
Yeah.
—Groq 4.6, and some Opus.
Yep.
And behind a router.
Yeah.
And I'm sure Elon is very focused on having it all be Grok as soon as possible.
All be first-party—yeah, yeah, of course.
But I think what you'll see these companies do is have their own model on their data, and it will work with 1 or 2 other frontier models. Not necessarily all the time, but just checking each other. It'll be a kind of transparency for—
Yeah, you could use the most frontier one for planning and then have execution run by everything else that's lower cost or so.
Yeah.
Yeah.
Absolutely. I think that feels like a very likely future to me, and that's a much more Microsoft-friendly future than one in which there are only 2 dominant frontier models. It certainly looks like there are going to be at least 3 with Grok. I do think you've got to give Meta a lot of credit.
They've done a great job.
Yeah, and I mean, they were out of the game, and they got back in the game.
Yeah.
It's kind of amazing.
It's very impressive.
Who could have imagined a year ago, when Gemini was ascendant, that this is the scenario we're in—that Gemini wouldn't even be in the conversation, and Muse and Meta would be significantly ahead of them from a capability perspective?
Yeah, exactly.
It's just the highest-stakes game of corporate chess ever played. Some people have made bad moves, and they've made good moves. You've seen some people come out of the game and others come back in.
I don't know if we're going to call it multi-model or a hybrid model. I don't know what terminology the world is going to settle on, but I think that's the future.
Yeah.
And I'm actually surprised. I think the best broad instantiation of that today, outside of GroqBot and Cursor—Harvey has done some cool things with it—is actually just the Fireworks Nexus product.
Yeah, yeah, exactly.
You can choose your frontier model. Let us take whatever open-source model you want, RL it for you, for your data—for Goldman Sachs, for Morgan Stanley, for J.P. Morgan, for Fidelity—
Yep.
—for a16z. You have all your own data, you control your intelligence, and we make it transparent behind a router.
Yeah.
I think that is a very plausible future, and that's clearly what Lynn from Fireworks was the first one to say. Then Alex Karp and Satya both took their own versions of it—
Yeah, they've taken their own version of it.
Yeah. Satya's essay on specialized intelligence—I think it's very plausible, but this stuff is really hard to do. That—
Yes.
—that sounds easy.
It sounds easy to describe. The way I describe it to people is: who gets to be the abstraction layer for the organization and the users with intelligence?
Yeah.
It's the most vied-for space or position that you could imagine in business, in the history of business.
Yeah, for sure.
Right? I think the answer is—
Yes, and for sure. Who's the arbiter of intelligence—
Yes.
—for global enterprises—
Yeah.
—and probably consumers? I was a retail analyst, and everybody thinks running one of these big chains is easy, but there's a lot to it. It's like, well, it's really easy: start an American retailer in any category, because America is so big, that's worth over $50 billion. Almost any category.
Yeah.
All you have to be able to do is have a fleet of 1,000 stores in 50 different states that have very different climates and consumer preferences. You need to have them stocked with the right products at the right time for that region and at the right prices. They need to be staffed by friendly and knowledgeable employees who don't steal from you.
Who turn over at 100% a year.
Who turn over at least 100% a year. The stores need to be clean and well lit. And if you can do that—presto, $50 billion.
Yeah.
In the history of American business, you can—I mean, it's more than one hand, but you don't have to—
That's like 10.
—go through many.
Yeah.
It's really hard to do.
Yeah.
Having that abstraction layer, having it work, having it be seamless, is way harder to do than people think. Something I think is very interesting about Cursor, and I'd love your opinion on this, is that everybody else in the lab space had this idea that we're creating a digital deity, AGI, and ASI.
Yeah, yeah, yeah.
The Cursor guys were just like, “We want to make a great product.”
Yes, exactly.
In a strange way, of everybody at the frontier, Cursor was probably the most product-focused.
Yes.
I'd say now they're part of SpaceX, but that suits Elon and his mindset really, really well.
Yeah.
Let's make it an engineering problem: create the model factory, and then we need to have a really good product.
Yeah.
The Tesla cars are amazing. I don't know if you drive one—
Yeah, yeah, I do.
—but it drives—
Yeah, yeah.
—everywhere.
Yeah, yeah. But what Cursor figured out is that they had, I would say, a similar end-state vision to what—
Yeah.
—those other guys had.
Yeah.
It was just a different path to get there. It's sort of like: be practical, meet the customer where they are, and meet the technology—
where it is. I think they'll sort of... They have already demonstrated that they can edge their way up into autonomy from that starting point.
Yeah.
Coding is unique compared to everything else in knowledge work. This would be in support of the point that Microsoft is in a good position because it is verifiable and perfectly documented, and nothing else in enterprise is verifiable and perfectly documented.
Yeah.
And so it will be messy. That leads you to a good bull case for something like Microsoft as that abstraction layer.
If they execute, but it's really—
It's really hard to execute.
Really hard to make it really simple: click my Copilot, link to all my stuff, train a model on our data, convince me that you're not going to share it with anyone else, and then put it behind a router that's seamless for me and continuously upgrade that open-source model.
Yeah.
Yeah, yeah.
It's not just some middleware. It's very hard to do.
Yes.
And, by the way, they're going to be competing with not only the labs to be that abstraction layer, but Databricks—
Oh, yeah.
Snowflake, Palantir, the inference—
Fireworks.
—the inference providers.
The inference files.
The application companies, right? Harvey has done an incredible job of this.
Yeah.
Legal has been taking off, and I think they can see the future of how to be that abstraction layer and do the work. But legal is also unique because it's very documented.
Yeah, and tax.
And it's somewhat verifiable.
We'll see tax.
Tax.
Similar.
We'll see that. We'll see things like that. But the $1.5 trillion—the really appealing, broad pie—is going to be very messy to go get.
Yeah, although I do always think, and I think probably in their heart of hearts, Harvey and Lagora think, “Oh, if we solve this, we could be that abstraction layer for everyone.” I think probably in their heart of hearts, Cognition thinks something like that, too.
I think everybody thinks it. And, by the way—
Everybody thinks it at this point.
This is massive validation of the category.
Yes.
Because Kirkland & Ellis said, “We're going to spend $500 million to build this ourselves.” First of all, good luck. That's going to be very hard.
Yes.
But that actually tells you that the pie is really big, right?
Oh, huge.
Yeah, it's massive.
And I'm sure they have a very smart head of AI, but it's not like a $500 million one-time build. That model—
No, it has to be constantly—
Continuously updated—
Constantly built.
Switching out the base model, and all of that has to happen transparently. But I think you're going to have this huge collision between products like Fireworks, Nexus, these legal agents, coding agents, and big companies like Microsoft.
Databricks.
Databricks.
Palantir.
And Snowflake coming up.
Yeah.
Salesforce, I think, is going to—Salesforce and Workday and all these companies are going to go after it, and it's just going to come down to who executes the best and who has the lowest costs.
Yes, exactly.
And it's going to be very hard over time. If you're not vertically integrated, you have to be so good to emerge as that abstraction layer.
Yeah, yeah, yeah—to be the low-cost provider.
Yeah.
Very hard.
Because you're simply not going to be the low-cost provider if you're not vertically integrated, if you don't own your own compute over the very long term. And that's another reason I increasingly look at these hyperscalers on EV to net PP&E.
Yes.
Because net PP&E is compute, and that is just what the market thinks you're going to monetize your fleet of compute at. You can look at them, and there are some pretty obvious inefficiencies, too.
Yes. Yeah, yeah, yeah.
Yeah. Kind of an AI version of price-to-book.
Yeah, I like the price-to-book. Okay.
Exactly.
7. Nvidia Controls The Ecosystem
So, okay, you mentioned Jensen. I share your sentiment—he's carrying this industry forward. Tell me your thoughts on NVIDIA.
I think he's in a very, very good position with his strategy of being vertically integrated but horizontally open. Let's say there's some accelerator that emerges that is really good. Almost certainly, it will be better if it can plug into—and this is why I know you have an accelerator investment—my number one thing is, if you're a semiconductor CEO, the only thing you should ever say is, “Thank you, Jensen.”
Yes. Yeah.
“Thank you for creating this opportunity. Thank you. How can we work with you? We want to enable you. Sure, we're going to compete with you on the edges.”
Yeah.
But my rule of thumb for accelerators is that every 1% of share today is probably worth $100 billion.
Yes.
So there's no need to go head-on with NVIDIA.
Yeah.
Just pick a niche and get your 1%.
That pie is very big.
He has 9 chips. He's got multiple flavors of accelerators. He's got CPUs, Ethernet switches, and 2 kinds of DPUs. We've gone from just scale-out networking being a thing. We have scale-up, scale-out, scale-across, and now scale-in.
Yeah.
So just try to find a way to plug into his ecosystem.
By the way, this is not foreign. His biggest customers all have competing products across various of those 9 chips.
Absolutely.
Yeah.
Just try to find a way to plug in, but just be nice to him. Be nice. Be nice. It's all personal, you know? Sometimes you hear some of these stories, and it's like, have you ever seen game tape of the Chicago Bulls when Jordan was playing? It's game 50 of the season.
Yeah.
And he's a little bored.
Yeah.
And the Bulls are down because they're up 8 games over the number 2 team in their conference, and he's a little bored.
Yeah, yeah, yeah.
And then somebody—
Somebody talks shit.
Somebody who's young decides, “I'm going to talk shit to him because we're beating him.” And then he just looks.
And it's like—
And it's like—
It's the best. Those are my favorite.
Yeah, it's amazing. Oh, yeah. We've all seen The Last Dance. Just don't do that.
Yeah, exactly. Exactly.
You know?
Yeah.
Just like, “Hey, Michael, man, I'm so happy to be on the court with you.” That's the move. But the reason it's particularly important is because Jensen's data centers are financeable.
Yes.
And it goes back to that point. Let's say it's $50 billion. For an NVIDIA data center, you need a $15 billion equity check.
Yeah.
You can finance the other $35 billion.
Yeah.
And it's not circular financing. I have a lot of respect for the people I have met from Blackstone, KKR, and Apollo.
Yeah.
And they're underwriting each of those.
Yeah.
They finance it, and then there's a residual value guarantee. As long as that residual value guarantee is less than the gross profit dollars he's getting from selling the chips into that data center, it's essentially super NPV-positive with very little risk for him.
Yeah, makes sense. Yeah.
And then he gets a revenue share. His data centers are the most financeable.
Yes.
A good case for probably TPUs is that they're the second most financeable.
Mm-hmm.
It probably takes, I don't know, at least double the equity check.
Right. Yeah.
And then the rates on the rest of it are higher.
Yeah, exactly.
And so cost of capital is a huge advantage, and that's why you just want to be part of his ecosystem. You can see he has all these chips.
He is acquiring land, power, and shell companies now, matchmaking them with offtake agreements. I think one reason he's doing these RVGs is that, if he doesn't do them, it's kind of an Anthropic and—
Yeah, of course. Yeah.
—OpenAI-dominated world—
Yeah.
—because they can pay the most for compute. He can effectively help other people—
Yeah, exactly.
—compete with Anthropic and OpenAI.
Yeah, in the same way that he stood up the neoclouds in the first place.
Yeah.
Yeah.
It's just democratizing compute, which is good for the world. Again, I think he's a patriotic American.
His interests are aligned with that, though, with the patriotic American ones, right?
Of course, yeah.
Fragmentation, right? Yeah.
Yes, fragmentation, no dominant AI.
Yeah, exactly.
Which is really good because he's a ruthless competitor.
Yeah.
It's awesome that his incentives around fragmentation of AI, fragmentation of models, and fragmentation of power are completely aligned with what's good for America. Going back to open source, I just can't take it that people think that Jensen is the world's biggest advocate for open source, and that it's somehow a giant risk to his business.
Yeah, exactly. No, it's great for his business.
Yes.
It's great for his business. Yeah.
It's amazing for his business because it means that instead of having a 90% margin on top of a token made with an NVIDIA GPU—
Yeah.
—maybe it's a 40% margin, so more of those tokens are going to be consumed, which means you need more compute.
Yeah, exactly.
In a supply-constrained world.
In a supply-constrained world.
And let's just say, what percentage of the world's supply has he locked up? 70%? 80%? Somewhere in there.
You're talking about fab capacity?
All of it.
Yeah.
All of it. It's just because he saw this coming before everybody else.
Yeah, and all the system supply chain. Yeah.
Yeah, he's got the fab capacity locked up. He's got DRAM capacity locked up. He's got NAND capacity, laser capacity, and capacitor capacity. He has what you need to make the racks. He used to say, if I go back 15 years, "Listen, I'm making a $2 billion or $3 billion bet every 2 years, and I'm moving really, really fast."
Yeah.
Now he's making these multihundred-billion-dollar bets, bringing the supply chain alongside him. He's bringing the financing alongside him by standardizing it, making it easy for the very smart people at Blackstone, KKR, Apollo, Goldman Sachs, Morgan Stanley, and JPMorgan to finance. That is hard to compete with.
Yeah.
My firm, Atreides, has a pretty big portfolio of private semiconductor companies. Elon said a lot of people are going to learn a hard lesson in hardware, and I would just say I've learned a lot of hard lessons in semiconductor investing. You can bet on the best team, and you tape the chip out. You feel great. "Okay, we've taped it out, and it—"
Yeah, yeah.
It's happening faster than ever. You feel great about it, and we're getting really good with the emulation and the simulations, so you feel great about it. Then the chip comes back from the lab. Everybody, you get a FaceTime from the CEO. They plug it in.
Yeah.
And then sometimes it doesn't work.
Yeah.
You know? It's just like—
Yeah, yeah, yeah.
The real—
This famously happened with Cerebras twice, right? And then they've powered through and done great jobs.
Yeah.
Yeah.
Well, I think each Cerebras chip worked. It just struggled to find product-market fit for the first 2 generations.
Yeah, yeah, yeah. Fair.
The chip worked.
Yeah. Yeah, yeah. Fair.
It just did not have product-market fit.
And they've done great with it. Yes.
But there's a different thing: you plug it in—
Plug it in, and it doesn't work at all. Yeah.
—and it doesn't work at all.
Yeah, yeah, exactly.
If it doesn't work at all, you might be back to the drawing board: "Hey, we need another hundreds of millions of dollars, or a billion dollars, and we've learned our lesson. It's going to work the next time, 2 years from now."
Yeah.
—
Yeah.
Assuming you can get financing. Semiconductors are hard. The real world is hard. Hardware is hard. What he's doing at the scale and speed he's doing it, bringing all of this alongside him—the land and the power have to come—
Yeah.
The entire supply chain has to come. The financing has to come.
Yeah.
And so, given that he's 70%, 80%, whatever we want to say, you just want to plug into that ecosystem.
Yeah. Yeah, that's why Elon made—
Be nice to Michael Jordan.
—that decision. Part of why Elon made the decision he made, right? Yeah.
Yeah, which I also think was a very high-Elon move.
Yeah, totally.
So you've had everybody else try to build their own ASIC. They've gotten up on stage. Sometimes they say negative things about Jensen or NVIDIA or take shots.
Yeah.
I did think what the Halapeno team did last night was pretty smart, and we should give credit where credit is due. Halapeno is, I would say, the first good ASIC, other than TPU or Trainium, I've seen from an internal team—
Yeah.
—in a pretty short amount of time.
In what seems to be a pretty short amount of time. It is very impressive.
In a pretty short amount of time. It's impressive. We should give credit where credit is due, and—
They do have a good team working on it. Yeah.
They have a good team, yeah.
Yeah.
They had a really good team. I think they had a lot of advantages, and I do think if you are a lab, have the model, and see the direction of research, that's a big advantage for designing your own chip.
Yeah.
But then you go back to NVIDIA, and they work with everyone.
Yes.
Everybody keeps thinking it's going to really standardize, and if you look at the 3 big Chinese open-source models—DeepSeek, Kimi, and Qin—they're all evolving in very different ways.
Yeah. Yeah.
They can all run on a more general-purpose chip, a GPU, but if you want to specialize—
Yeah, you're going to need general-purpose chips at a minimum for the types of evolution you see from that. Yeah.
Yeah.
Yeah.
I'm very happy his incentives as a CEO are perfectly aligned with what's good for America.
Yes.
Make sure your semiconductor guys—
Be nice to MJ.
—do not talk trash about Michael Jordan—
Be nice to MJ.
—ever.
Be nice to MJ.
Be nice to MJ.
Yeah, exactly.
Yeah, and then sometimes you tug on Superman's cape, and you get confident. You get confident, and you start to talk a little bit of trash. Well, Superman sometimes just flies away.
Yeah. Yeah.
That's what happened to the TPU team.
Yeah. Yeah.
And Halapeno, they're tugging on Superman's cape a little bit.
Yeah, we'll see. Yeah, yeah, yeah.
We'll see.
Yeah.
It is kind of amazing that Halapeno did something that none of the big—
This is as competitive a chip as I have seen.
Yeah.
But again, it's just competitive with one—
Yeah, one of his 9—
—of his 8 or 9 chips.
Yeah, one of his 9.
Right?
Yeah, of course.
And it's—
They'll continue to work closely together, yes.
Yeah, they'll continue to work closely together. It's like, hey, that's great. You did the one thing. To actually be competitive with him at the system level, you need another 8 chips.
Yeah, yeah, yeah. Exactly.
Dylan Patel at SemiAnalysis talks about how he's the bank of AI. He's the central bank of AI.
Yeah. Of course.
He's the Federal Reserve of AI.
Yeah.
And so I actually think it was really smart for Elon, instead of competing with somebody who is—
Yeah.
Fully aligned.
Mm-hmm. Yeah.
And I think history is going to judge that to be a wise decision. In a world that is so supply-chain-constrained, it's actually really hard to tell what true customer preferences are.
Right.
Because you come out with a—
Yeah, they'll take anything. That's—
You—
This is how you know that—
Yeah.
The very old generation—whatever. The resale price of H100 is very high.
Yeah, yeah. And if you have a TSM allocation, you're going to be sold out.
Yes.
Particularly if you can get the DRAM to pair with it.
Yeah.
You're going to be sold out.
Yes.
So it's actually kind of hard to infer true customer preferences, and I actually think one of the best ways you can see true customer preferences is the kind of deals they cut with chip companies. So broadly speaking, the first deal is where the chip company invests in a customer, and you saw TPU and Trainium—Amazon and Google—do that with Anthropic.
Yep.
And that was to their immense advantage because it really helped their businesses. I think it helped those chips really level up, because you need to use a chip.
Yeah, yeah.
There's a cold-start problem. And in that scenario, as long as the dollars you invest are less than the gross profit, you can't lose money.
Then there's the scenario where you do the RVG. Blackstone finances it, or whoever—Blackstone, Apollo—
Yeah.
KKR and Goldman Sachs finance it.
As long as that RVG is actually less than your gross profit—
You're fine.
You can't lose money, and you have upside, probably through a revenue share on top of it.
Then there are deals where you give warrants away, but they're tied to a fixed price per 1 million tokens.
Yep.
And as long as the performance of your chip outruns the performance of your stock, you're going to do—
Yeah, that's valuable.
Good in that situation.
Yeah.
If you just give warrants away, it could be negative NPV—
Yeah, for sure.
Because the better the stock does—
Yeah. The more value—
The worse the deal is.
That's captured by the person, yeah.
Yeah. And so you can look at that hierarchy of deals and infer something about true customer preferences.
Yeah, that's interesting.
Yeah.
So Nvidia does pretty good deals.
Yeah. I mean, there's a reason that people I consider smart are investing in their deals.
Yeah, I see it.
Gavin, thank you.
Fun. Always fun to hang out with you. Thank you.
Thanks, Gavin. This was great, man.