Gavin Baker, managing partner and chief investment officer of Atreides Management, spent July and August asking people who run AI businesses for one quantitative data point that was getting worse, and says he could not find a single one. He and a16z general partner David George work through the payback math on a $50 billion gigawatt, how few people are actually driving the demand, the politics of data centers, orbital compute, and why Baker thinks Nvidia's financing terms matter more than its chips.
👤 Guest: Gavin Baker, managing partner and chief investment officer of Atreides Management, who ran Fidelity's OTC Portfolio from 2009 to 2017 before founding the firm in 2019
🎙️ Host: David George, general partner at a16z
📰 Published: 31 August 2026
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 1 hr 15 min | ✅ Time saved: TIMESAVED min
Key Takeaways
Two months of asking AI companies for a deteriorating number produced nothing
"And it's at least in July and August, I haven't been able to find a single person."
Public AI names sold off over the same two months, which he called a wide but shallow index move: "And I mean, you can drown crossing a river that's on average two feet deep."
The frontier labs can move reported revenue by a factor of four without losing a customer
Ten gigawatts, eight of them on inference: "They're allocating eight to inference, and let's just say they're monetizing that inference at whatever, $60 billion a year. So $480 billion a year in revenue."
Flip the allocation to training and "your revenue just went from 480 to 120"
A gigawatt costs $50 billion and the customer prepays half of it
Nebius discloses enough to get to a nine-to-10-month payback, on Baker's arithmetic
The rest is financed by Blackstone, KKR and Apollo, not by the chip vendor
The heavy-user base may be under ten million people against 1.5 billion knowledge workers
"And so, yeah, your 30 million is probably way overstated. It might be sub 10."
a16z's most AI-native portfolio companies spend "like 10% plus" of human compensation on tokens; old-economy companies about 1%
Data centers are the best thing that has happened to working-class America, and nobody in AI will say it
"Well, you know what? It's probably the best thing that has ever happened to working class Americans."
His model for how to say it is Sheryl Sandberg reading out named small businesses on Meta's early earnings calls
The risk is undersupply through 2028, so the price of intelligence goes up rather than down
"And by the way, like, there's no capacity available with all the forecast builds that will happen through 28, which are probably now going to be delayed, given the political dynamics they have."
Blocking the build produces "real compute inequality" in which only large companies and wealthy people can buy intelligence
Open-source tokens cost the same compute as frontier tokens, and Kimi charges for them
"But people have this idea that open source tokens are free. They're not."
The Kimi license stipulates "a 30% share of any revenue" — open weights, not open source
Orbital compute is an accounting argument, not a science-fiction one
Of a $50 billion gigawatt, about $35 billion is IT and travels; the $15 billion of power, cooling and labor is what you stop paying, and it is the inflationary part
With Starship reusability the launch cost "goes to under a billion" and "the economics just instantly flip"
Every 1% of accelerator share is worth roughly $100 billion, which is an argument for not fighting Nvidia
"My number one thing is if you're a semiconductor CEO, the only thing you should ever say is thank you, Jensen."
Nvidia has locked up "70, 80, somewhere in there" of the world's supply, fab through DRAM, NAND, lasers and capacitors
Nvidia's advantage is that its data centers are financeable, not that its chips are fastest
A $50 billion Nvidia data center needs a $15 billion equity check; the residual value guarantee sits below Nvidia's gross profit on the chips, and it takes a revenue share on top
Open source raises the number of Nvidia tokens consumed, so Jensen Huang funds it
A 90% margin on a token becomes a 40% margin, and more tokens get consumed in a supply-constrained world
The shape of a chip deal reveals what customers actually want, because the shortage hides it
Chip company invests in the customer, then residual value guarantees, then warrants tied to a fixed price per million tokens — a descending order of confidence
Nobody Could Name a Single Number Getting Worse
George opened by putting it to Baker that he had spent the summer on the West Coast hunting for someone to give him a bearish case and make his sentiment more negative, and asked whether he had found anybody.
He asks the same question of everyone and got no answer to it for two months. "And it's at least in July and August, I haven't been able to find a single person."
The one qualification he offered was that Anthropic is in a quiet period, so it may have slowed
His read is that everything else sped up. OpenAI has clearly accelerated, open source has accelerated more, and Grok has had "a pretty dramatic acceleration" since Grok Bot, he said
Public markets went the other way. AI stocks have "fallen out of bed" over the last two months, and while they bounced a little in August some of the names are still in significant drawdowns
"And I mean, you can drown crossing a river that's on average two feet deep." — there is not much action at the index level, which is what hides the individual drawdowns
Everyone Wins Is a Real Scenario, and the Zero-Sum Framing Is the Error
George brought up Benchmark's Eric Vishria, who had said on Patrick O'Shaughnessy's podcast that "maybe everyone wins" — Anthropic, OpenAI, SpaceX, Meta, Google by selling TPUs, open source, and the application companies. Baker agreed with the shape of it, with one carve-out: "Maybe not all applications. Applications that I think execute well and navigate this."
George said the framing is the mistake, and that his LP conversations all start in the wrong place. Every other conversation begins with how this is all going to go wrong, and his answer is "This is not an or thing. It's an and thing, right?"
"Frontier's going to work really well. And minus one models are going to work really well. Open source is going to work really well. There's going to be a bunch of application companies that work really well."
The clouds and the lab companies are probably going to do really well too, he said
On Anthropic, Baker offered a hypothesis rather than a claim. He thinks the company "probably trued up and cleaned up some accounting" and rebased so it is comparable to OpenAI, then did its testing-the-waters process
His guess is that the next disclosure is a reacceleration
The release timing is a game between the labs. Anthropic is "clearly waiting for OpenAI to release Astra", and then, he said, "it's like the next day, here's Fable 5.1" — magically available several hours later
Being shot at during a quiet period is the current cost. "And they're in a quiet period, so they can't really shoot back." He thinks having OpenAI and Anthropic public will help the market, because of how public investors currently model the industry: "And a lot of public investors, you hear, oh, Sarah Friar said this in an all hands meeting and it's on the cover of Wall Street Journal."
He is uneasy about one recruiting practice. "Anthropic is now in their culture interviews saying, how would you feel if the equity went to zero?"
His objection is practical: the firm wants missionaries, but it also needs people to make money, and at zero equity it cannot buy the compute the mission needs
"They're like the accidental enterprise company." George's version: "They're kind of like the accidental everything."
Reallocating Power Between Inference and Training Is a Revenue Dial the Public Markets Have Never Seen
The arithmetic he uses is deliberately conservative. Take a lab with 10 gigawatts of power: "They're allocating eight to inference, and let's just say they're monetizing that inference at whatever, $60 billion a year. So $480 billion a year in revenue."
George: on a revenue basis, that is roughly a one-year payback, "not a gross profit basis"
"People seem to think Anthropic and OpenAI are both monetizing it $100 billion a gigawatt today."
A research breakthrough could invert the split overnight. Move to eight gigawatts on training and two on inference and "your revenue just went from 480 to 120" — and Baker said he thinks a lab would actually make that decision
George's point is that this is a new kind of disclosure risk. "And this is just something that public markets are going to really have to get used to."
Baker expects the reality to be less dramatic once a lab is listed, because the stock affects morale, recruiting and retention — but much of the revenue is still under management's control, through which checkpoint gets released, where it is priced on the Pareto curve, and how power is split
The internet-era comparison does not hold. Baker said Meta and Google never faced this trade-off. "Well, there was no massive tradeoff they had to make in terms of the cost or infrastructure to serve revenue side." The two were totally separate
Free Cash Flow Is Not Coming, and Microsoft Already Blinked Once
Operating cash flow, not free cash flow, is the number to watch. "For sure, I don't think they will generate free cash flow anytime soon. I think they're going to generate a lot of operating cash flow", which then goes straight into buying chips
First-party products are being subsidized. Token consumption on the labs' own products is heavily subsidized, Baker said, alongside the research spend
The scaling-law belief is what rules out a profit pivot. Given what the labs all seem to believe about scaling laws, he does not think any of them will focus on free cash flow
He named the one retreat and its cost. Satya Nadella gave an interview at Davos about capital spending in which he said "I know I'm good for my 80 billion", and Baker said Microsoft blinked, slowed down, and regrets it
Dario Amodei's caution was deliberate and is now being tested. Baker relayed the argument: underspend and you lose share, overspend and you can go bankrupt. "And like, those are both bad things, but bankruptcy is worse than losing share, so I'd rather be conservative."
OpenAI was aggressive and is back in the game; SpaceX was aggressive; George said the returns on those decisions look clearly right on both a short and a long horizon
A Gigawatt Costs $50 Billion and Pays Back Inside a Year
The disclosed numbers get him to a sub-year payback. Nebius and CoreWeave both gave enough away, he said, to reach a nine-to-10-month payback for Nebius: a gigawatt costs $50 billion, and an upfront customer payment covers 50% to 60% of it
"So now, you're talking about $25 or $30 billion, and then you can monetize it if you put it into the spot market."
SpaceX is faster still, because it builds big clusters and brings them on quickly, and can monetize them at a higher rate
He has changed the unit he thinks in. "I have tried to shift, to think of pricing and per megawatt rather than per GPU."
The opportunity set is unusual in his career. "And I just, 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."
On the circularity complaint, he pointed at who is actually writing the cheques. "And it's like, well, I don't know. I know a lot of smart people who work at Blackstone and KKR and Apollo. And they're the ones that are financing it." — at a relatively low cost of capital, which George agreed is very low today
One reason the financing works is that useful lives keep getting extended as models improve and the return on token spend rises
The Demand Side Might Be Under Ten Million People
George put the demand question on the table: the labs are doing, "call it 180 billion of revenue", on the back of maybe 30 million heavy paying users getting real value — mostly developers.
Baker thinks even that is too generous. "I might take the under on 30 million, man." and then "And so, yeah, your 30 million is probably way overstated. It might be sub 10."
Inside a16z's portfolio the spending is a power law twice over. George said old banks are probably spending 1% and very tech-forward companies high single digits, and that within those companies "The highest spending engineers are spending 10 or sometimes 100 X more than the median engineer."
The denominator is the whole point. "There's a one and a half billion knowledge workers. Like it feels like we're nowhere on the demand side and we're massively supply constraint."
Baker asked George for the ratio he uses. Token spend per month against human compensation: "Oh, high single digits, some at 10%, like some of the very AI-native ones, like 10% plus.", against about 1% at old-economy companies doing a good job
George's conclusion: diffusion into the real economy could disappoint, but over a 10-year stretch "we're nowhere"
Baker's Own Token Bill Went Up a Hundredfold in Five Months
His firm is the case study. Internal token consumption at Atreides rose 100-fold from March through August, and after two people got access to Grok Bot Enterprise he expects it to rise another 10x or 20x in a month
"Like we have some heavy Grok Bot users here and like it is very productive use. Like this is not like wasteful tokens"
He is 50 and says he will never be native at this. He compared himself getting his parents onto an iPhone and an iPad, then contrasted the generation below him: "And you see these like 23-year-old kids and just the way they use AI. They're just fluent and native in it." George is 42
The build time is what changed, not the idea. He had built a podcast summarizer, a Substack summarizer, an X summarizer and a sentiment tracker for topics and stocks
"And like that, all of those would have taken me, I don't know, hours working with Claude Code. And they each took seven to 12 seconds with Grok Bot."
"So to me, Grok Bot does feel like another, at least for me, like kind of ChatGPT moment"
George drew the line between assistance and work. The most sophisticated engineers went from roughly 20% of their code written with AI to more than 90%, but George called what Baker described building still reactive: "It's all like knowledge enhancing, which is part of your job. but it's not actually doing the work for you."
The next step is the agent proposing the action. "Yeah, and now you have a Grok Bot that says, what are the recommended actions?", and Baker said he is now racing Grok Bot against Codex on the same jobs
George said he uses Town, an a16z portfolio company, the same way
Every Real Technology Gets a Bubble, and This One Is Still Funded Out of Cash Flow
He put the pattern on the record before making the bull case. Going back to the South Sea bubble, every genuinely new technology — the automobile, TV, radio, the internet, the PC, railroads, steel mills — produced one
"You get a bubble because the markets get really excited. and they get ahead of themselves. Things get overvalued. That overvaluation leads to an overbuild."
He corrected his own earlier account of the South Sea episode on a previous podcast: he had thought it was connected to the invention of longitude, and "Turns out it was not."
What makes this cycle different is who is paying. "And even today, a majority of this is still being funded out of operating cash flow, which I think is really helpful."
"You know, debt funded buildouts, they demand immediate ROI, not an ROI in two years."
The physical constraints are real, and he thinks they are a feature. The build is big enough to hit the raw productive capacity of whole industries — every copper investor now has an AI thesis, he said
"You know, we're in this acute shortage with, I don't know, several million people are driving a crazy global compute shortage. What happens when that's 500 million?" and "And how many copper mines do we need to build to, like, support this?"
"And so, like, these fundamental constraints, I think, are slowing us down. And I think that's good. I actually think that's good for society."
He added two more brakes. Real rates: "So it makes sense that real rates are going up." given how much is being invested. And regulation, where he said "I'm kind of shocked at what's happening in America. We're in a really bad place."
The Industry's Best Argument Is the One It Refuses to Make
Baker's complaint is that the only group able to tell the AI industry's truth is the AI industry, and it is not doing it. He cited an exchange in which Dario Amodei pointed out that he had written two essays, one positive and one negative, and Baker's reply was that being 50% negative lands differently when the negative case is existential.
He wants the pitch changed from curing cancer to curing cancer. "I thought one of the best things Dario said was, like, what we need to do is stop talking about curing cancer and actually cure cancer."
The line he keeps coming back to is scriptural. "But just somebody, like my favorite line in the Bible is the truth shall set you free."
On data centers, he thinks the industry is conceding an argument it should be winning. "Well, you know what? It's probably the best thing that has ever happened to working class Americans."
The trade alternative to college: you can learn to be "an electrician, a plumber, an HVAC tech" and "make ungodly amounts of money"
With behind-the-meter power generation, he said, a data center transforms a town's tax revenue — it does not merely double — and is revitalizing dying small towns
"The water consumption thing is totally debunked. The water is nothing.", and the gas generally used is a relatively clean fuel
George pushed back on where the burden of proof now sits. "And I think the problem with it now, it's like the burden of proof is on not curing cancer, but actually delivering some real tangible everyday American benefits beyond using chat, or grok to like answer your questions or substitute for a search engine."
The China Argument Is True and Useless; Loudoun County Is the One That Works
Baker said the stay-ahead-of-China case is correct and ineffective. He is a patriot and believes it, but "it's way too abstract" for the average American: "I'm pretty sure the Pacific Ocean is really big."
The concrete version is a county. Loudoun County, Virginia has the highest per capita income in the country and the highest density of data centers, and makes a great deal of tax revenue from them
Told by a data center opponent that they would like to see the things built in the richest zip code and the richest county, his answer was that they already are: "So we've done that. And it worked out really well."
"Yeah, but hey, don't bother me with the details. I'm on to my next talking points."
He made a strong and unproven claim about the opposition. "Like, there is an organized CCP-funded campaign. I think against data centers here in America, like, I think a lot of it gets laundered through TikTok."
The other half of the story is the input cost. With the Strait of Hormuz closed, "You know, natural gas here is two or three bucks. It's now 25 bucks in Europe and Asia, or 20 bucks or whatever it is."
Gas feeds electricity, and electricity feeds nearly every manufacturing process, so the US now has a large cost advantage on a basic input
"We are reindustrializing America. And it's awesome." — the thing both parties have said they wanted, arriving in the towns the steel mills left
"I mean, I tried to do it on every podcast, but like I'm just a dude." George's answer was that he is preaching to the tech audience that already agrees
Meta Told This Story Once Before, One Small Business at a Time
George suggested Meta is doing the best job of telling it now, and Baker said the habit is old and wired into the company.
The template is a roll call of named businesses. Early in Meta's life as a public company, "Sheryl would run through, 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."
His example of the form: a cake bakery in Des Moines, "And it was two women who were single mothers working by themselves, and now they have 15 locations.", employing 50 people
He wants every AI company doing it. SpaceX, Anthropic, OpenAI, Google, Meta, Nvidia, AMD and Broadcom, each naming real businesses and real people, anonymized where permission is not available
"Just run through specifics because the truth shall set you free, but only if you tell it."
The Risk Is Undersupply Through 2028, Which Means the Price of Intelligence Rises
Baker's forecast is a shortage, not a glut. "And by the way, like, there's no capacity available with all the forecast builds that will happen through 28, which are probably now going to be delayed, given the political dynamics they have."
George: everybody is worried about oversupply. Baker: "Massively undersupplied."
That points at higher prices for access to intelligence, which George noted is the opposite of the consensus direction
Baker credited the framing to Dwarkesh Patel — "Yeah, well, Dwarkesh had a wild point." — and George supplied the version he remembered: "Like the cost of a token could go up 10x or something like that."
The reason it can happen is consumer surplus. People choose frontier tokens over cheaper ones for many reasons, but the biggest, he said, is that there is "a tremendous amount of surplus" even when using the frontier tokens
Blocking the Build Buys Compute Inequality
He named the consequence and who would own it. The result of what he called the data center degrowthers would be "real compute inequality", with big companies and wealthy people able to afford compute. "You know, that happened because you wouldn't let us build data centers."
"You know, that happened because you wouldn't let us build data centers."
George's mechanism is that cheap consumer access takes time to build. The route to a low-cost mass-market product is advertising, and building an advertising business is slow — as it was for the consumer internet companies a16z has backed
Baker called it a bad outcome for everyone. "Like a compute inequality, like future. That's not a good, that's not a good future for anyone, which is another reason open source is so important."
Open-Source Tokens Are Not Free, and Kimi Takes 30% of the Revenue
The cost of the token does not change with the license. "But people have this idea that open source tokens are free. They're not. And it's like, it takes the exact same amount of compute."
The only variable is the margin charged on top
The Kimi license is the detail he thinks is underappreciated. "And even then, the Kimi license, something that I don't think a lot of people appreciate, is the Kimi license stipulates a 30% share of any revenue."
"So, like, Kimi has taken a 30% cut of all the revenue generated on its, and this is because it's open weights, not open source."
It is also expensive to run. George said the model is extremely token hungry; Baker said that on a task basis rather than a token basis it is "far more inefficient"
The Age of Elon and Jensen
He put the two of them in the same sentence as a historical period. "I think this will be like the age of Elon and Jensen." — the Victorian age being his comparison — because they are altering the fabric of human society and civilization
The Starlink example is the one he says nobody talks about. "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 was not there either
Any incremental internet capacity now comes from space rather than from traditional builds, he said, which he called a huge unlock
An Orbital Data Center Is the Size of an Airplane
Baker's first job is deflating the mental picture. "It's not like big buildings in space." George's correction: "Yeah, it's like a big rack."
The rack itself is about "five of us standing together", and the airplane comparison is the solar wings
The orbit keeps a radiator permanently in the rack's shadow, which is how it is cooled
The physics objections have a pattern. He described people on X announcing that they hold a physics PhD and that this is impossible, including an investor friend who does hold one and argued the point repeatedly
"And then he goes to the SpaceX day and he talks to the SpaceX engineers he's like, well, I was wrong."
"Have you thought about this for an hour? Have you thought about it for 10 hours?" against "10,000 of the world's smartest engineers at SpaceX who've thought about this each for hundreds, if not thousands of hours"
George granted the physics and moved to cost. He said the history of the Elon Musk companies is that the cost curve gets dramatically better — Starlink was not commercially available when a16z first invested, and the same questions were asked about launch and about the Model 3
"It feels clear to me at a minimum it will be swing capacity. Yeah. And in the fullness of time, maybe it will be larger."
The Orbital Case Is an Accounting Argument About the $15 Billion You Stop Paying
The question to ask is Starship reusability, not whether compute works in space, Baker said
The split he uses: of a $50 billion gigawatt, about $35 billion is IT and travels with you, growing a little because it is going into space
"The rest is power, cooling, labor, all sorts of things that you don't need in space because you have the solar panel and the big radiator."
That $15 billion is the inflationary part — electrician compensation, materials, copper, optics
So the comparison is launch cost against the $15 billion. "And so what you have to compare it to is the cost of launch. And with Starship reusability, that goes to under a billion. So the economics just instantly flip."
Earth keeps training. There are latency and speed-of-light constraints, and advantages to having GPUs next to each other, so "So data centers on Earth, they're not going anywhere."
"But an increasing fraction of the world's compute is going to be in orbit."
There is a launch date on the record. "And Elon said that he and Jensen have co-designed a Rubin rack. And it's going to launch in the fourth quarter of 27." Baker's point was that even a slip of a couple of quarters leaves it close
He credited Brad Gerstner with the observation that nobody is really paying attention to this: "And it's like kind of happening in plain sight."
Starlink Mobile, X Ads, and Why George Calls It Heads-You-Win
Baker's answer to orbital-compute skeptics is that they do not need to agree. Starlink Mobile has a credible plan on its own: "And that wireless is, call it another $900— $900 billion of revenue that they address."
"So, yeah, your mobile plus your broadband, whatever, it's called it like close to $2 trillion of a market."
On top of that sit the terrestrial businesses already working — Cursor, Grok and Grok Bot — and X advertising, where he said "we have telemetry" showing growth
He expects the bundle. A Starlink, Grok Bot and X-advertising package, on the logic that Google bundled its way into paid products: "maybe you're bundling the ads with AI" — "but why not do that?"
George's framing of the position: "And so they've made the very aggressive compute investments to enable that first party work, and that's the kind of heads you win."
The tails case is that they overbuild for their own inference and training and sell the excess into a shortage, with a "sub six month payback on the compute side"
It also answers the bear case that a world of two dominant labs designing their own chips leaves nobody else any room: "Well, like, I don't think they're going to have a reusable starship and multiple spaceports anytime soon."
Baker: "Yeah, then they're a massive infrastructure business."
Starbase Louisiana and Thousands of Launches a Year
George raised it as the thing he is most excited about. "Yeah, I'm so fired up about the Starbase, Louisiana." and "I was reading about it last night, and yeah, it's sort of like, they now have the infrastructure for thousands of launches a year."
Baker expects more of them, in more countries. Multiple coasts, somewhere in the Middle East, whichever European country is least bureaucratic at the time, and possibly Japan or South Korea
The throughput number is what makes it feel unreal. "Yeah, the capability to do, call it whatever, 5,000 launches a year.", at "two a day, two a day per pad" — and Baker said he thinks the pads are being engineered for more than two a day
He drew a line between catching a rocket and reusing one. China did catch a rocket, using what he called a jury-rigged system of wires that had been suggested on the SpaceX subreddit before the first Falcon landing
"And like China's clearly paying close attention to SpaceX subreddit."
"But that's very different catching that thing from what they're trying to do with Starship", where the booster is caught, moved, stacked, fueled and sent straight back up
Asteroid Psyche, and Bezos's Line About Earth Being Zoned Residential
George asked for the most futuristic thing Baker thinks about with SpaceX, recalling a conference debate about the first $10 trillion company where Baker had said "I have no idea, but I know which one's going to be the first $20 trillion company."
The answer was asteroid mining. "Look, I mean, 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, every precious metal in it that exists in the Earth's crust."
The mechanics as he described them: capture with Starship, possibly needing a lunar base, park it in a stable geosynchronous orbit over an American-owned atoll with no humans within about 50 miles, work it with Optimus robots, and let delivery to Earth be free
The framing he borrowed is Jeff Bezos's, from about 15 years ago. "He said, I think in the future, Earth is going to be zoned residential." and "He's like, all heavy industry will take place in outer space."
That answers the pollution objection, he said, and the worry about still being able to see the stars underestimates how big space is
The nearer item is Mars. "Let's just say at the outside, this is eight years away." for a fleet of starships landing there
"You're going to have Optimus robots holding American flags, like walk down, and then they're going to pull out a bunch of solar panels and batteries and racks of compute", dropping Starlink capacity as well
4K video from Optimus robots all over Mars, then humans: "And that's going to be an amazing moment for America."
George's response was that there is not a lot of chatter about that one, and Baker's was that he thinks it is highly likely to happen
Microsoft Failed at the Frontier and the World Got Friendlier Anyway
The failure is not in dispute. "You know, they clearly tried to make a frontier model. They failed." and "You know, Satya said, we're going to have our own models that are very competitive. Like, I think he said that 18 months ago, they don't have their own models that are competitive."
The friendlier world is one of many models rather than two. "But what you're seeing with, I think the future is an ensemble of models. You know, there's a Pareto curve." and no single model is best at everything
He expects the best open model to be an Nvidia model in the near future, and said the chip companies could fund the training in a world where open source wins: "It's trivial to do a $50 to $100 billion training run, for Jensen."
"And maybe soon, I do wonder if this is kind of Google's, like, super long-term play." — monetize compute at high rates, sell TPUs externally, and let the cash flow eventually decide who funds the biggest training runs
"But I do think you're going to see American Open Source, led by Nvidia, get really close to the frontier." The Poolside acquisition, he said, was made for a reason: "Poolside actually had a lot of really good American open source talent."
Nemotron is the shape of what an enterprise wants. It has not had much post-training, which leaves a good pre-trained base to do what you want with
The alternative — handing your enterprise context to a frontier lab — is where he thinks the risk sits: that context is "the context embedded in all of your data", and sharing it "may be hazardous for your financial health". George pointed at the change in zero-data-retention policy as the specific trigger
So the play is a capable open model with heavy reinforcement learning and supervised fine-tuning on your own data, so "you want to own and control your intelligence"
The production shape already exists. On Grok Bot: "I think it's Gemini 3.7, Flash, Grok 4.6, and some Opus." behind a router, though he expects Musk to want it all running on his own models
"It'll be kind of transparent to the most frontier for planning and then have execution run by everything else that's lower cost."
Meta got back in and Google fell out of the conversation. Baker said Meta deserves a lot of credit for coming back from being out of the game; George said that a year ago, when Gemini was ascendant, nobody would have predicted "The Gemini wouldn't even be in the conversation."
Who Becomes the Abstraction Layer, and Why Running a Retail Chain Explains the Difficulty
George named the prize. "It's, yeah, who's the arbiter of intelligence for global enterprises and probably consumers?" — which he called the most contested position he can think of in the history of business
Baker's analogy is that easy-sounding businesses are not. "I start an American retailer in any category because America is so big that's worth over $50 billion."
"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, consumer preferences, you need to have them stocked with the right products at the right time for that region, at the right prices."
"They need to be staffed by friendly and knowledgeable employees who don't steal from you.", at turnover of at least 100% a year, in stores that are clean and well lit
In the history of American business, he said, the number of people who have done it is more than one hand but not many
Cursor is his example of the opposite instinct. While the labs talked about digital deities, AGI and ASI, the Cursor founders said "we want to make great product", and of everyone at the frontier they were probably the most product-focused
"Yeah, I'd say, and now they're part of SpaceX, but that suits Elon and his mindset really, really well." — make it an engineering problem, build the model factory, ship a good product
George's version is that Cursor had a similar end-state vision and a different path: meet the customer and the technology where they are, then leg up into autonomy
Coding is the exception that flatters Microsoft. "Coding is unique compared to everything else in knowledge work." It is verifiable and perfectly documented. "And like nothing else in Enterprise is verifiable and perfectly documented."
George listed what makes the abstraction layer hard to actually build: link everything, train on the customer's data, convince them it will not be shared, put it behind a seamless router, and continuously upgrade the underlying open model — "It's not just some middleware."
And the competition is not only the labs: Databricks, Palantir, the inference providers and the application companies, with George saying Harvey "has done an incredible job of this"
Baker: legal is unusual because it is very documented and somewhat verifiable, and tax may follow — but the broad, appealing pie is "very messy to go get"
Kirkland and Ellis Is Spending $500 Million to Build Its Own, Which Is the Bull Case
Baker read the law firm's decision as validation of the category, not a threat to it. "Because Kirkland and Ellis said, we're going to spend 500 million bucks to build this ourselves. Like, first of all like, good luck."
"But that actually tells you that the pie is really big."
The reason he doubts it is that the work never ends: "That model has to be continuously updated, switching out the base model.", all of it transparent to the user
He expects a collision rather than a segmentation. Fireworks Nexus-style products, legal agents, coding agents, Microsoft, Databricks, Snowflake, Salesforce and Workday all going after the same position: "It's just going to come down to who executes the best."
George: "And this is just who has the lowest costs?" Baker agreed, and said an unvertically-integrated company has to be exceptional to win it
Which is why he now values the hyperscalers off their assets. "Because net PP&E is compute and that is just what the market thinks you're going to monetize your fleet of compute at.", and he said there are some obvious inefficiencies in what that implies
George: "Yeah, kind of an AI version of price to book."
Every 1% of Accelerator Share Is Worth $100 Billion, So Do Not Fight Jensen Huang
The strategy Baker credits is vertical integration with an open perimeter. "So I think he's in a very, very good position and his strategy of being vertically integrated, but horizontally open."
His advice to chip founders is close to a script. "My number one thing is if you're a semiconductor CEO, the only thing you should ever say is thank you, Jensen. Thank you for creating this opportunity. Thank you. How can we work with you? We want to enable you."
"But my rule of thumb for accelerators, every 1% share today is probably worth $100 billion."
"So there's no need to go head on with Nvidia. Yeah. Just pick a niche. Get your 1%."
Nvidia already sells nine chips — multiple accelerators, CPUs, Ethernet switches, two kinds of DPUs — across scale up, scale out, scale across and now scale in, and its biggest customers already compete with several of them
The basketball version of the same advice. "So I just make sure your semiconductor guys, do not talk trash about Michael Jordan." and "Be nice to MJ."
"And then it's like sometimes it's like you tug on Superman's cape and you get confident."
Nvidia's Real Moat Is That Its Data Centers Are Financeable
The equity cheque is the number. "Like, let's say it's $50 billion. For an Nvidia data center, you need a $15 billion equity check." and "You can finance the other $35 billion."
He rejects the circular-financing description directly. "And it's not circular financing. I have a lot of respect for the people I have met from Blackstone and KKR and Apollo.", who underwrite each of them
The residual value guarantee is structured so Nvidia cannot lose. "And then there's a residual value guarantee, which 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 like essentially it's super NPV positive with very little risk for him."
"And then he gets a revenue share."
Competing chips are financeable on worse terms. He said TPUs are probably the second most financeable and still take at least double the equity cheque, with higher rates on the rest — "And so cost of capital is a huge advantage, and that's why you just want to be part of his ecosystem."
The guarantees are also a competitive weapon against the labs. "I think one reason he's doing these RVGs is if he doesn't do them, it's kind of an Anthropic and OpenAI dominated world because they can pay the most for compute. He can effectively help other people compete with Anthropic and OpenAI."
George: the same way he stood up the neoclouds in the first place. Baker: "It's just democratizing compute, which is good for the world."
He thinks the incentives happen to point the right way. "Which is really good because he's like a, he is a ruthless competitor, and 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."
Open Source Is Good for Nvidia, Not a Threat to It
The margin trade is the whole argument. "Because it means that instead of having a 90% margin on top of a token made with an Nvidia GPU, maybe it's a 40% margin, so more of those tokens are going to be consumed, which means you need more compute." — in a supply-constrained world
The supply is already spoken for. "And let's just, what percentage of the world's supply has he locked up? 70, 80, somewhere in there." Asked by George whether he meant fabs, Baker's answer was "All of it. All of it. You know, it's just because he saw this coming before everybody else."
DRAM, NAND, laser and capacitor capacity, plus what is needed to build the racks
The scale of his bets changed by two orders of magnitude. Fifteen years ago the line was "I'm making a two or three billion dollar bet every two years and I'm moving really, really fast." Now the bets run into the hundreds of billions, with the supply chain and the financing brought along
He standardized the financing so that Blackstone, KKR, Apollo, Goldman Sachs, Morgan Stanley and JPMorgan can underwrite it: "And, like, that is hard to compete with."
Silicon Is Hard: The Chip Comes Back and Sometimes It Does Not Work
Baker's own scar tissue is the caveat on all of it. Atreides has a sizeable private portfolio in semiconductors, and "You know, Elon said a lot of people are going to learn a hard lesson in hardware."
"Like you can bet on the best team and you tape the chip out. You feel great." Emulation and simulation are getting good, tape-outs are faster than ever
"The chip comes back from the lab. Everybody you get a FaceTime from the CEO. They plug it in." — and sometimes it does not work
The failure case is expensive and slow. Back to the drawing board, another few hundred million to a billion dollars, and a two-year wait, assuming financing is available at all
George raised Cerebras as the known example. "Yeah, this famously happened with Cerebras twice, right?"
Baker's correction was substantive: "I think the chip, I think each Cerebras chip worked, it just struggled to find product market fit." for the first two generations
What makes Nvidia's execution unusual is everything that has to move with it. The entire supply chain has to come, and the financing has to come, at that scale and that speed
Jalapeno Is the First Good ASIC Outside TPU and Trainium
He gave credit for the newest one. "Jalapeno is the, I would say, the first good ASIC other than TPU or Trainium I have seen.", done in what looks like a short amount of time by a good team
A lab that has the model and can see the direction of research has a real advantage in designing its own chip, he said
But one chip is not a system. "But again, it's just competitive with one of his eight or nine chips.", and to compete at the system level you need the other eight
The Chinese open models argue against standardization. "And if you look at the three big Chinese open source models, DeepSeek, Kimi, Qwen, they're kind of all evolving in very different ways." — they can all run on a general-purpose GPU, and specialization needs one at a minimum
Which is where the central-bank metaphor comes in. Baker credited Dylan Patel at SemiAnalysis: "He's like, he's the central bank of AI. He's the Federal Reserve of AI."
On Musk choosing to work with Nvidia rather than against it, Baker had earlier called it "a very high ELO move"
The Shape of a Chip Deal Tells You What Customers Actually Want
The shortage hides the preference. "In a world that is so supply chain constrained, it's actually really hard to tell what true customer preferences are." Customers take anything, and "if you have a TSM allocation" you are going to be sold out, particularly with DRAM to pair with it
So he reads the deal structures instead, in descending order of what they imply:
"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." — which he said was to their immense advantage, because a chip needs to be used and there is a cold-start problem
Then the residual value guarantee, financed by Blackstone, Apollo, KKR or Goldman Sachs: "And as long as that RVG is actually less than your gross profit. You can't lose money.", with upside from a revenue share
"Then there are deals where you give warrants away, but they're tied to like a fixed price per million tokens." — fine as long as chip performance outruns stock performance, and potentially negative NPV if you hand warrants over unconditionally, because the better the stock does the more value the counterparty captures
"And so you can kind of look at that hierarchy of deals and infer something about true customer preferences."
On Nvidia's own terms: "Yeah, I mean, there's a reason that people I consider smart are investing in their deals."
Baker's bottom line is that the AI trade is supply-constrained rather than demand-constrained, that the constraint is now political and physical rather than technical, and that the companies which own their compute and can finance it — with Nvidia at the center of both — are the ones the shortage pays.
Products, Companies & Tools Mentioned
Atreides Management (Baker's firm; internal token consumption up 100-fold from March through August, and a sizeable private portfolio in semiconductors)
Nvidia (The center of the conversation: nine chips, 70% to 80% of the world's supply locked up, and the only data centers Baker says are truly financeable)
Anthropic and OpenAI (In a quiet period and heading for listings; Baker thinks public disclosure will improve how the industry is modeled)
SpaceX and Starlink (Aggressive on compute, fast to bring clusters on, and the vehicle for both orbital data centers and the mobile-plus-broadband market George and Baker size at close to $2 trillion)
Grok and Grok Bot (Baker's own "ChatGPT moment" — summarizers and trackers built in seconds rather than hours, and a router mixing Gemini, Grok and Opus)
Claude Code and Codex (The coding agents he built with before and races against now)
Cursor (His example of a company that chose product over digital deities, and which he says is now part of SpaceX)
Meta (Back in the frontier game, and the model for how to tell the industry's story, via Sandberg's small-business roll call)
Google (Selling TPUs externally and monetizing compute, which Baker wonders may be a very long-term play)
Microsoft (Tried a frontier model and failed; the ensemble-of-models future is friendlier to it, if it can build the abstraction layer)
Nebius and CoreWeave (Their disclosures are what Baker uses to get to a nine-to-10-month payback on a gigawatt)
Blackstone, KKR and Apollo (The underwriters of the data center build, and his answer to the circular-financing complaint)
Goldman Sachs, Morgan Stanley and JPMorgan (Named alongside them as the financing Nvidia standardized its way into)
Kimi, DeepSeek and Qwen (The three big Chinese open-source models, evolving in different directions; the Kimi license takes a 30% share of revenue)
Nemotron and Poolside (Nvidia's lightly post-trained open base model, and the acquisition Baker says was made for its American open-source talent)
Fireworks (Its Nexus product is George's example of the multi-model abstraction layer actually shipping)
Harvey and Legora (The legal-AI companies George says have done an incredible job, in a category that is documented and somewhat verifiable)
Databricks, Palantir, Snowflake, Salesforce and Workday (The other contenders for the abstraction layer, all going after the same position)
Kirkland & Ellis (Spending "500 million bucks to build this ourselves", which Baker reads as proof the pie is big rather than as a threat)
Cerebras (George's example of silicon going wrong twice; Baker says each chip worked and the problem was product-market fit)
Tesla and Optimus (The cars he says drive everywhere, and the robots he expects to work an asteroid and walk down a Starship ramp on Mars)
TSMC (An allocation there is enough to be sold out, which is part of why customer preference is unreadable right now)
AMD and Broadcom (Named in his list of AI companies that should be telling specific customer stories)
X and TikTok (Where the advertising telemetry is growing, and where he claims the anti-data-center campaign is laundered)
Town (The a16z portfolio company George uses to push his own work toward automation)
Fidelity (Where Baker ran money before founding Atreides, and one of the enterprises he sees owning its own intelligence)
Books & Resources Mentioned
If Anyone Builds It, Everyone Dies – Eliezer Yudkowsky (Baker's shorthand for the existential end of the AI argument, which he thinks the industry answers badly)
Invest Like the Best with Patrick O'Shaughnessy (Where Benchmark's Eric Vishria made the "maybe everyone wins" case George opens with, and where Baker described his own AI workflow five months ago)
SemiAnalysis – Dylan Patel (Source of the central-bank-of-AI framing for Nvidia)
The Last Dance (The Michael Jordan documentary behind his advice to chip executives about talking trash)
The SpaceX subreddit (Where the wire-catch method China used was suggested years before the first Falcon landing, by his account)
Get the latest market chatter as it happens:
If this saved you the listen, send it to one person who would rather have the time back.