Gavin Baker on Orbital Compute, TSMC, and Frontier Models

ELI5 / TLDR

Gavin Baker thinks the AI boom is the wildest thing he’s ever seen — Anthropic added more revenue in a single month than Snowflake, Palantir and Databricks built over a decade combined. He argues the bottleneck shifts from chips to electricity to permits to physical land, and ends up in orbit, with SpaceX flying refrigerator-sized racks of GPUs that cool themselves into the vacuum. He worries that everyone now agrees with him — and when everyone agrees, the price stops being your friend.

The Full Story

The most extraordinary moment in the history of capitalism

Baker doesn’t reach for that phrase lightly. The frame he keeps coming back to: in roughly one month, Anthropic added more annualized revenue than Palantir, Snowflake and Databricks built across a decade of grinding effort. The whole software-as-a-service revolution created somewhere between five and ten trillion dollars of value. Anthropic is currently doing fifty billion in annualized revenue, growing at rates that don’t really have a precedent.

“Anthropic added their combined businesses in one month. Nothing like that has ever happened in the history of capitalism.”

That’s the setup. The puzzle is that the market kept selling AI off through March and April even as this was happening. Baker thinks people were spooked by the Strait of Hormuz — but he came around to the view that closing it was actually a relative gift to America. US natural gas prices fell twenty percent. Asian and European gas prices doubled or tripled. Electricity is the input to AI, and America just got cheaper electricity while its competitors got more expensive electricity overnight.

So you had the most bullish AI fundamentals in history, paired with one of the cheapest relative valuations for tech in a decade. Baker bought.

Anthropic vs OpenAI as businesses

He thinks of them as different animals. Anthropic has burned roughly eighty percent less capital than OpenAI to reach a similar revenue scale. They are probably already cash-generative on inference. OpenAI is bigger but more expensive to run, with Sarah Friar working on the unit economics.

Anthropic at nine hundred billion for fifty billion of ARR sounds rich until you account for the fact that they are demonstrably compute-constrained. They have been quietly degrading Claude — even Opus now produces seventy percent fewer tokens per question than before. And in language models, token count is roughly proxy for quality of answer. If they had unconstrained compute, Baker thinks they would be doing well past a hundred billion, maybe two hundred. At which point the multiple is more like five times revenue.

“Why don’t they just raise a hundred billion at three trillion?”

Because Elon, of all people, is the case study Baker reaches for. SpaceX compounded at low thirty percent per year for a decade because Elon kept the price reasonable. By making investors money for twenty years straight, Elon earned what Baker calls a superpower — the ability to raise as much capital as he wants whenever he wants. Anthropic appears to be playing the same game. Don’t squeeze every last dollar of valuation now; preserve the option to raise forever.

Watts and wafers

This is Baker’s standing framework for the AI build-out. You need watts (electricity) and wafers (chips). Both are bottlenecks. Both are constraining the trade.

On watts, he’s relatively optimistic. Capitalism will solve it. The bigger worry is regulatory and political — zoning, approval, midterm-election politics. Turbine capacity is being expanded; jet engines are being repurposed (Boom Aerospace is doing this); the West is relearning how to cast turbine blades it stopped making eighty years ago. He expects the watts shortage to start easing in 2027-2028.

And then, he thinks, orbital compute really cracks it open.

Racks in space, not pentagons in space

Here Baker insists on a reframe. When people hear “data centers in space,” they picture some Borg cube the size of a stadium floating in low earth orbit. That isn’t what’s coming. What’s coming is racks in space — individual GPU racks, about the size of a Blackwell rack on Earth (eight feet high, four feet deep, three feet wide, three thousand pounds), each with five-hundred-foot solar wings on each side and a similarly long radiator extending behind to dump heat into the vacuum.

The orbit is the trick. Keep the satellite in a sun-synchronous orbit — meaning the sun is always on the panels and the radiator is always pointed at deep space. Solar in front, cooling behind, compute in the middle. Link the racks together using lasers travelling through vacuum, which is what every Starlink satellite already does.

“What makes a data center is you’re connecting racks with lasers. So it’ll be racks in space connected with lasers into a virtual data center.”

Why does this work now and not before? Because SpaceX. Starship makes the economics work. Starlink V3 satellites will already operate at twenty kilowatts; a Blackwell rack runs at a hundred. The gap isn’t that big. And SpaceX already operates the world’s largest satellite fleet (somewhere around ninety-eight to ninety-nine percent of all satellites in orbit) and the world’s largest data center on the ground. They have both the hardware engineers and the launch capacity, which no one else does.

Baker is direct about the skeptics. He’s spent a lot of time at SpaceX’s Starbase facility, talked to a lot of their engineers, thinks they’re the most talented hardware group on the planet, and they’re confident they’ve solved the radiator and repair problems. The thing they’re still working on is what happens when one rack fails — for now the answer is roughly “you don’t repair it, until you have floating robots that can.” But the launches are cheap enough that you treat the rack as a consumable.

Crucially, Baker says training will stay on Earth for a long time. Inference is what migrates to orbit. So this isn’t bearish for terrestrial data centers — humanity is going to consume every watt of compute it can produce. But if you’re an investor in turbine makers or cooling companies whose entire thesis is the watts shortage lasting forever, you should think about what happens around 2028.

The TSMC question

If watts will eventually get solved, wafers are the harder one. Wafers come down to “this group of flinty older humans in Taiwan” who think very differently from Western executives. They view themselves as inheritors of Morris Chang’s sacred legacy. Twenty years ago, when Baker asked Taiwan Semi if they could catch Intel, they said it was a beautiful dream for their grandchildren. They did it.

Jensen Huang has no contract with TSMC. They do business on handshakes — “we’ll be fair to each other over time.” This is wild given the dollar amounts.

Baker thinks TSMC may have single-handedly prevented an AI bubble. If they expanded capacity at the pace Nvidia actually wants, Nvidia could probably sell two to three trillion dollars of GPUs in 2026-27, which would absolutely cause a supply overhang and a crash. Instead, TSMC’s discipline keeps wafer supply scarce enough that demand stays ahead.

The risk: Intel and Samsung are eight to fifteen months behind on leading-edge nodes. One of them eventually breaks ranks on capacity discipline, which forces everyone else to break. Bubbles tend to follow.

Terafab, the Elon move

Then there’s a new piece of the chessboard: Terafab, a joint venture led by SpaceX (and possibly Tesla) to build the world’s largest fab in America, in partnership with Intel.

Baker thinks this works for unusually specific reasons. ASML, KLA-Tencor, Lam Research and Applied Materials — the equipment vendors — never liked having a single monopsony customer in TSMC. They want a second source. Their A-teams will go to wherever the next serious challenger is. Because of Elon’s reputation in hardware engineering, those A-teams will show up at Terafab. And because Elon is a “living deity” (Baker’s words) in Taiwan, South Korea and Japan, he can do something Intel and Samsung can’t — recruit the best engineers from each country and bring them to Texas, building a Taiwan Town and a Korea Town next door so they have their actual favorite restaurants.

“Elon tends to do things differently. Everybody else has taken three years to build a data center. He built one in 122 days.”

Frontier tokens still win, which surprises Baker

Here is one of the more counterintuitive points of the conversation. The deflationary case — “open-source Chinese models will be 95% as good at 1% of the cost” — has not happened in any way that matters to revenue.

Frontier tokens, meaning the very best model at the moment, are capturing an overwhelming share of the economic value generated at the model layer. Gemini 3.1 Pro was mind-blowing when it came out, and is now, in Baker’s words, “intolerable” — not because it got worse, but because the frontier moved.

Within frontier models, there’s a Pareto frontier of intelligence-versus-cost. Nine months ago, Google dominated every point on it. Today the frontier is Anthropic, OpenAI, and Grok 4.3 (a strong low-cost option). Gemini 3.1 is “hanging on.” This is the Nvidia effect made flesh — Google made conservative design choices on TPU v8 to take share away from Nvidia and Broadcom, and Nvidia kept making aggressive choices, and Google lost its cost-per-token leadership.

If frontier tokens stop commanding a premium, value will explode at the application layer. If they keep commanding it, the application layer keeps struggling.

Richard Sutton’s bitter lesson

The “bitter lesson,” for anyone outside the AI world, is Richard Sutton’s observation that throwing more compute at a problem reliably beats human algorithmic cleverness. Every time someone thinks they’ve found a trick to make AI more efficient, brute force wins anyway.

A violation of the bitter lesson — meaning some clever trick that actually does let you do more with less compute — would be the biggest risk to the entire trade. Baker is a believer that we are very close to artificial superintelligence, and ASI might be the thing that finally violates the bitter lesson by making itself more efficient. The first thing a 400-IQ model would want is more resources, and the way to get more resources is to use less per task.

The new prisoner’s dilemma

Baker frames an interesting game-theoretic point. The old AI prisoner’s dilemma was about spending — if you don’t spend, your competitors will and they’ll win. The new one is about whether to release your frontier model through an API at all.

If all the frontier labs agreed not to release their best models via API, Chinese open-source would fall further behind quickly. But if one lab defects and releases — they capture all the revenue, get more cash, and pull ahead. So someone always defects. The game is the same game TSMC, Samsung and Intel play with capacity discipline. Eventually someone breaks.

Baker thinks Jensen will keep open-source roughly a generation behind the frontier — close enough to keep the application layer alive, far enough to protect frontier economics.

Different and hard

This is Baker’s mental model for chip startups and for venture investing in general. Two questions: Is what you’re doing different? And is it hard? If it’s different but not hard, Nvidia will copy it. If it’s hard but not different, you’re trying to be a better GPU and you will lose. You need both.

Cerebras did “wafer-scale computing” — putting an entire model on a single wafer-sized chip — which is both genuinely different and genuinely hard. It took them three chip generations to get right.

For application-layer startups, Baker uses Jamin Ball’s “token path” framing. If you’re a software company in the AI era, you have to be in the token path — generating, processing, or routing AI tokens. If you’re not, life will be hard.

The disaggregation of prefill (loading the model’s context) and decode (generating new tokens) opens new design space. Prefill is memory-capacity-bound; decode is memory-bandwidth-bound. Andrew Fox’s analogy: imagine an 18th-century British naval cannon. Prefill is loading the cannon. Decode is firing it. You can build different chips for each step.

One quiet consequence: GPUs might have ten- or fifteen-year useful lives instead of three or four, because you can keep an old Hopper running prefill workloads even after it’s too slow for state-of-the-art decode. Baker thinks this could “single-handedly save private credit” — the funds that financed GPU leases on three-year amortizations are going to look much better.

The hyperscalers, one by one

Google has the most installed compute of anyone, which matters more in a world of shortages. Their TPU advantage has been lost (the Nvidia effect again). But Google’s never going to be in a bad position given the data they have, the YouTube footage which is genuinely valuable for robotics training, and the search business.

Meta gets immense credit from Baker. Zuckerberg made Meta an AI-first company internally — the only one of the original internet giants to actually do that. The billion-dollar talent contracts looked silly at the time. Llama’s successor “Muse” arrived close to the Pareto frontier.

Amazon is in a strong position because of Trainium chips and because robotics is about to flow into the retail P&L over the next eighteen months.

Microsoft is the courageous-decision story. Satya Nadella flinched briefly in early 2025 on capex — and in this business, flinching means losing allocation that’s very hard to get back. He recovered by making a hard call: use Microsoft’s compute internally to improve products like Copilot rather than just selling it to OpenAI. Microsoft would probably be an eight-hundred-dollar stock if they just sold all their GPUs to OpenAI, but Nadella is choosing the long game.

Engagement with startups is the tell. Amazon and Nvidia engage deeply. Google somewhat. Broadcom in its own way — getting Broadcom for your first chip is, in Baker’s phrase, “manna from heaven.” AMD, Microsoft and Meta engage with startups essentially not at all. Baker thinks this will hurt them.

The Last Samurai problem

Baker has a metaphor for human investors in the AI era. He recently rewatched the Tom Cruise film The Last Samurai. The samurai are master warriors who fight with sword and bow. At the end, an elite samurai is killed by an untrained peasant with a machine gun.

“The machine gun is here. If we do not all become masters of the machine gun, we’re going to get mastered.”

But — and this is the optimistic note — a fifty-year-old veteran samurai who learns to use the machine gun will be more dangerous than the peasant. So the leverage for an experienced investor learning to use AI tools well is enormous, for at least a window of time. Baker’s most useful agent today is one that reads podcasts for him and surfaces only the parts that match his idiosyncratic investing interests — what management is incentivized to do, how PSU structures have changed, that kind of thing.

The diversity-breakdown worry

Baker is starting to get nervous. Not because the fundamentals are weak — he thinks they’re extraordinary — but because nobody disagrees with him anymore. Carlota Perez’s framework: foundational technologies always go through a bubble. The market correctly recognizes the importance, everyone becomes bullish, and Michael Mauboussin’s “diversity breakdown” sets in. That breakdown is what funds the buildout but ends in a crash.

Right now Baker can’t find anyone who is bearish on DRAM. Nuclear stocks and quantum stocks already feel like little bubbles. Low-quality high-cost players in industrial supply chains are getting bid up because shortages make the marginal supplier look good — exactly what happens at the top of a commodity cycle.

One countervailing piece of evidence: actual high-quality AI names like Nvidia were, at the April lows, as cheap relative to the market as they have been in ten or twelve years. That’s not bubble behavior.

But the cross-sectional valuations make no sense together. Semicap-equipment companies trade at forty times next-quarter annualized earnings while DRAM companies trade at mid-single-digit multiples. Both can’t be right.

The dark coda

Baker ends on something he says is “a little dark.” He thinks AI is going to make the world higher-variance and higher-beta. He worries about personal safety for people associated publicly with AI — citing the Molotov cocktails thrown at Sam Altman’s house — and recommends that every family and every company have a “safe word.” Not a digital one. One you decide on in person, at the ocean, with phones left behind, in case someone deepfakes a family member asking you to wire a million dollars.

Geopolitically, Ukraine is starting to win not because they have better drones but because they have the best battlefield AI outside America and Israel. China is processing what that means. The West has had a huge AI lead, which is destabilizing for the rest of the world.

But Baker is still an AI maximalist. He shared a story about a friend whose daughter was diagnosed with a very rare genetic mutation — the friend spun up enormous amounts of compute and AI agents, found a drug already on the market that could affect her disease, and started a company to develop it. Her life has changed already, because of AI.

Key Takeaways

Anthropic added more ARR in one month than Palantir + Snowflake + Databricks built in a decade. The exponential is the steepest Baker has ever seen.
Anthropic has burned roughly 80% less capital than OpenAI for similar revenue scale — fundamentally different unit economics.
Anthropic is compute-constrained. Claude Opus now produces 70% fewer tokens per question than before; unconstrained, Anthropic would likely be at $100-200B ARR.
Elon’s “superpower” is the ability to raise unlimited capital because he made investors money for 20 straight years. SpaceX compounded at low-30%/year. Anthropic seems to be playing the same long game on valuation discipline.
America got a relative gift from the Strait of Hormuz scare: US natural gas dropped 20% while Asian/European gas doubled or tripled. Electricity is AI’s input — America’s manufacturing competitiveness improved overnight.
“Watts” (electricity) will be solved by capitalism over 2027-2028. The constraint is shifting to zoning and political approval.
“Orbital compute” is racks in space, not stadium-sized data centers. Refrigerator-sized GPU racks with 500-foot solar wings and radiators, in sun-synchronous orbit, linked by lasers through vacuum.
Inference will move to orbit; training stays on Earth for a long time. Terrestrial data centers remain valuable, but turbine/cooling investors should think about 2028.
TSMC’s capacity discipline may have single-handedly prevented an AI bubble. If they expanded as fast as Nvidia wants, Nvidia could sell $2-3T of GPUs in 2026-27.
Jensen has no contract with TSMC. They operate on handshakes — “we’ll be fair to each other over time.”
Terafab (SpaceX + Intel + maybe Tesla) is the new American fab play. The thesis: Elon can attract the wafer-equipment vendors’ A-teams and recruit talent from Taiwan/Korea/Japan in a way Intel and Samsung can’t.
Frontier tokens still capture overwhelming economic value — the “open-source models at 1% the cost” thesis hasn’t played out where it counts.
The Pareto frontier of intelligence-vs-cost is now Anthropic, OpenAI, and Grok 4.3. Google dominated it nine months ago and has fallen off.
The “Nvidia effect”: Google made conservative TPU design choices to claw back share, Nvidia kept making aggressive ones, and Google lost cost-per-token leadership.
Richard Sutton’s bitter lesson — compute beats cleverness — is the biggest risk to the trade. ASI might be what eventually violates it by making itself more efficient.
The new prisoner’s dilemma: do frontier labs release their best models via API? If everyone holds back, Chinese open-source falls behind. Someone always defects.
Chip startup rule of thumb: be different AND hard. 1% market share is worth ~$100B. Don’t try to be a better GPU.
Disaggregation of prefill (memory-capacity-bound) and decode (memory-bandwidth-bound) opens new chip design space. Prefill = loading the cannon. Decode = firing it.
GPU useful lives are extending to 10-15 years because old Hoppers can run prefill workloads. This may “single-handedly save private credit.”
AI shifted from all-you-can-eat to usage-based pricing. The $200/month plans are now “lobotomized” — to access the real frontier you need enterprise plans.
Continual learning + memory are the next frontier. Humans are wildly more sample-efficient than AI. If continual learning lands, you get a fast takeoff.
Hyperscaler engagement with startups (in order): Amazon and Nvidia deepest, then Google, then Broadcom in its own way. AMD, Microsoft and Meta engage essentially zero — Baker thinks this will hurt them.
Microsoft chose to use compute internally rather than sell it to OpenAI. Probably foregoes $800 stock price short-term, but positions for a post-API-access frontier world.
“Diversity breakdown” — when everyone agrees, the trade gets dangerous. Baker can’t find a DRAM bear anywhere, which worries him.
Cross-sectional valuations don’t make sense together. Semicap equipment at 40x and DRAM at 5x can’t both be right.
“The Last Samurai” frame: AI is the machine gun. An experienced investor who masters AI tools will dominate for a window of time.
Baker recommends every family/company have a non-digital “safe word” to defend against deepfake social engineering. Decide it in person, at the ocean, phones off.
Ukraine is winning not because of drones but because of battlefield AI. America/Israel/Ukraine have the lead. China is processing what this means.
The token path: if you’re a software company in the AI era and you’re not in the token path, you’re in trouble.

Claude’s Take

This is one of the densest hours of investor talk you can find on what is actually happening underneath the AI trade. Baker is not a hype merchant — he’s spent his career running tech-focused funds at Fidelity and now Atreides, and his framework here is mostly the same boring one he’s always used: cost curves, market structure, capacity discipline, who controls the bottleneck.

The signal-to-noise ratio is unusually high. The reframe of orbital compute (racks not pentagons) is genuinely clarifying — I had the same mental picture of a Borg cube he’s pushing against. The Carlota Perez / Mauboussin diversity-breakdown framing is exactly the right pair of glasses for where the trade is now. The bitter-lesson violation as the biggest tail risk is the right intellectual hook to hang your skepticism on, even if you don’t think it’s likely.

A few places I’d push back. Baker is closer to the Elon orbit than is comfortable for a totally neutral read — the “living deity” line about Elon, the “doing more for America than any other American” — these are tells that he is talking his book on SpaceX and Terafab. The orbital compute thesis specifically is conditional on Starship working, which it is, mostly, but the economics are still emergent. His Frontier-tokens-still-win observation is also somewhat shaped by where Atreides has its money — if the model layer is where the value continues to acrue, his fund does well.

The personal-safety / safe-word bit at the end is the thing I keep returning to. It’s the most actionable advice in the whole episode and it’s free. Doesn’t matter whether you believe in orbital compute. Pick a safe word with your family. Don’t put it in a text.

Score: 9. A point off because the Elon material is uncritical and because Baker himself flags he’s becoming consensus, which is exactly the problem with treating his framework as Truth. But the texture of how someone with a billion-dollar portfolio is currently thinking about each layer of the AI stack is rare and useful, and there is real, specific, name-dropping detail throughout — Trainium 3’s switch scale-up network, Cerebras’ shoreline-IO ratio, the Pareto frontier composition shift, the Microsoft compute reallocation. This is the version of the AI conversation that takes the buildout seriously as physical infrastructure rather than vibes.