SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?

TAKEAWAYS

My biggest takeaways from Gavin Baker of Atreides Management

1. 01

   The AI ROI debate is over: $100B of ARR at 80% inference margins means the returns are already there.

   Gavin puts OpenAI and Anthropic at roughly $100 billion of ARR today with 80%-ish gross margins on inference, and notes Anthropic was even positive for the Wall Street Journal in the most recent quarter. Across LLMs broadly, excluding Google and the recommender and ad use cases, he sees $200 to $400 billion of ARR at high margins by the end of this year. At those numbers the returns question answers itself.

   “if OpenAI and Anthropic call it $100 billion of ARR now, with 80%-ish gross margins on inference, the returns are there”

2. 02

   The AI market is cross-sectionally inefficient: Nvidia, memory, and the power names can't all be priced right.

   Gavin's frame for the whole chip trade is that the multiples on memory, on Nvidia, and on the power, cooling, and optical names are mutually inconsistent, so one group is mispriced. If the power, cooling, and optical multiples are correct, Nvidia and memory go up a lot; if the Nvidia and memory multiples are correct, everything else probably underperforms. Either way, today's prices can't all hold.

   “The AI market is cross-sectionally inefficient right now”

3. 03

   The ASIC share-loss story is wrong: Nvidia is growing faster than hyperscaler capex.

   Gavin says that on a true apples-to-apples basis, Western and ex-China, Nvidia is indisputably growing faster than hyperscaler CapEx, which means it is gaining share rather than losing it to custom silicon, even with Broadcom guiding 143% year-over-year AI semiconductor growth for the comparable quarter. His tell on the rival ASICs: they aren't submitted to clean benchmarks like MLPerf and InferenceMax because they would lose, and you can't fight shadows.

   “it is indisputably true that he is growing faster than hyperscaler CapEx”

4. 04

   Older GPUs have a 10-to-15-year useful life, which kills the Burry amortization bear case.

   Gavin's answer to Michael Burry's short-amortization argument: put disaggregated accelerators in front of older GPUs and those chips stay useful for 10 or 15 years. That long life is what makes GPUs financeable, and he calls an asset-backed financing for GPUs at a lower rate than other chips a profound competitive advantage, pointing to CoreWeave's lowest rate at around 6 percent.

   “those older GPUs, they have a useful life for 10 or 15 years. And this means that you can finance GPUs”

5. 05

   Cursor going Pareto-dominant after three weeks on Colossus proves proprietary data decides the coding race.

   Gavin calls Composer 2.5 extraordinary: the same base model as Composer 2 (Kimi K25), then three or four weeks of reinforcement learning on Colossus 2 with Cursor's data, and it lands well outside the Pareto frontier. He reads it as extremely significant for both XAI and Cursor, and as evidence that in an agentic world the harness is essentially as important as the model, so the two need to be developed together.

   “Composer 2.5 is literally well outside the Pareto frontier. And that's after three weeks”

6. 06

   SpaceX's real AI edge is that it stamps out data centers faster than anyone: 122 days, then 91, then 66.

   Gavin's underwriting case for the compute business starts with build speed: SpaceX's first data center took 122 days, the second 91, the third about 66, at lower cost than anyone else. With a clear offtake partner like Anthropic already renting Colossus, he expects SpaceX to stamp out data centers fast and grow the AI and compute business far faster than anyone contemplated.

   “They build data centers dramatically faster than anyone else at a lower cost”

7. 07

   The only people skeptical of orbital compute are the ones not involved in data centers or space.

   Gavin says Google, Anthropic, Amazon and Nvidia are all convinced orbital compute will be a reality, and notes there is already a working Nvidia H100 in space that Andrej Karpathy both trained a model on and ran inference with. His point prediction for orbital compute becoming real is the second half of 2028 to the first half of 2030, and he thinks SpaceX selling compute to Google is the easy first step.

   “The only people who are skeptical of orbital compute are those people who are not involved in data centers or space.”

8. 08

   Rapid reusability is Starship's whole game, and it's a year or two out, maybe sooner.

   Gavin defines the hard problem as flying and landing the same rocket multiple times per day, not merely reusing it. Even if rivals solve reusability, he expects their mass-to-orbit to asymptote to a small number. His guess for SpaceX getting there: a year or two, maybe sooner.

   “Rapid reusability means that you can fly the same, fly and land the same rocket multiple times per day”

9. 09

   There's no dollar crisis coming: America is the best house in a bad neighborhood.

   Gavin's rebuttal to Friedberg's 'Dr. Doom' monologue: high debt is a global problem, and the US is energy and food self-sufficient, now the world's largest oil and gas producer and exporter, and leads in AI, with no better-fundamentaled reserve currency on offer. Rising rates concern him, but the tech bubble happened with the 10-year and 30-year much higher than today, and Nvidia trades at a low-teens multiple of real earnings versus Cisco's 100x.

   “I don't think a dollar crisis is around the world”

10. 10

    Once you give a power to the government it's almost never taken back, so don't hand it AI yet.

    Gavin thinks it's too early for the US to regulate or test frontier models, and he especially dislikes the United States doing it while no one else, China included, does. America already has self-regulation and the courts as recourse. His deeper objection is structural: government power is a one-way ratchet that needs to grow.

    “once you give something, give a power to the government, it's almost never taken back. It needs to grow”

11. 11

    Crime is now a choice.

    With Flock Safety cameras, gunshot detection and Las Vegas-style drone-and-detection policing available, Gavin argues policing failures are no longer a capability gap but a policy decision. Different municipalities will choose to be pro-crime or anti-crime, and voters will sort it out.

    “Crime is now a choice.”

12. 12

    Sell China the deprecated Nvidia GPUs; it's the best way to keep America ahead.

    Gavin's counterintuitive read on chip export policy: selling older Nvidia GPUs into China lowers the odds it builds its own, more power-hungry chip ecosystem, which keeps America ahead and in control of AI. He'd pair it with energy leverage: cut off Venezuelan, American and Brazilian oil, keep the Strait closed, and China is left with only Russian oil, in no position to fight a war against the US-led bloc.

    “selling in deprecated Nvidia GPUs to China, lowers the odds of them developing their own alternative ecosystem”

My biggest takeaways from Gavin Baker of Atreides Management:

1. The AI ROI debate is over: $100B of ARR at 80% inference margins means the returns are already there. Gavin puts OpenAI and Anthropic at roughly $100 billion of ARR today with 80%-ish gross margins on inference, and notes Anthropic was even positive for the Wall Street Journal in the most recent quarter. Across LLMs broadly, excluding Google and the recommender and ad use cases, he sees $200 to $400 billion of ARR at high margins by the end of this year. At those numbers the returns question answers itself. [4:34]

2. The AI market is cross-sectionally inefficient: Nvidia, memory, and the power names can't all be priced right. Gavin's frame for the whole chip trade is that the multiples on memory, on Nvidia, and on the power, cooling, and optical names are mutually inconsistent, so one group is mispriced. If the power, cooling, and optical multiples are correct, Nvidia and memory go up a lot; if the Nvidia and memory multiples are correct, everything else probably underperforms. Either way, today's prices can't all hold. [1:14:22]

3. The ASIC share-loss story is wrong: Nvidia is growing faster than hyperscaler capex. Gavin says that on a true apples-to-apples basis, Western and ex-China, Nvidia is indisputably growing faster than hyperscaler CapEx, which means it is gaining share rather than losing it to custom silicon, even with Broadcom guiding 143% year-over-year AI semiconductor growth for the comparable quarter. His tell on the rival ASICs: they aren't submitted to clean benchmarks like MLPerf and InferenceMax because they would lose, and you can't fight shadows. [1:14:22]

4. Older GPUs have a 10-to-15-year useful life, which kills the Burry amortization bear case. Gavin's answer to Michael Burry's short-amortization argument: put disaggregated accelerators in front of older GPUs and those chips stay useful for 10 or 15 years. That long life is what makes GPUs financeable, and he calls an asset-backed financing for GPUs at a lower rate than other chips a profound competitive advantage, pointing to CoreWeave's lowest rate at around 6 percent. [1:20:48]

5. Cursor going Pareto-dominant after three weeks on Colossus proves proprietary data decides the coding race. Gavin calls Composer 2.5 extraordinary: the same base model as Composer 2 (Kimi K25), then three or four weeks of reinforcement learning on Colossus 2 with Cursor's data, and it lands well outside the Pareto frontier. He reads it as extremely significant for both XAI and Cursor, and as evidence that in an agentic world the harness is essentially as important as the model, so the two need to be developed together. [49:34]

6. SpaceX's real AI edge is that it stamps out data centers faster than anyone: 122 days, then 91, then 66. Gavin's underwriting case for the compute business starts with build speed: SpaceX's first data center took 122 days, the second 91, the third about 66, at lower cost than anyone else. With a clear offtake partner like Anthropic already renting Colossus, he expects SpaceX to stamp out data centers fast and grow the AI and compute business far faster than anyone contemplated. [48:03]

7. The only people skeptical of orbital compute are the ones not involved in data centers or space. Gavin says Google, Anthropic, Amazon and Nvidia are all convinced orbital compute will be a reality, and notes there is already a working Nvidia H100 in space that Andrej Karpathy both trained a model on and ran inference with. His point prediction for orbital compute becoming real is the second half of 2028 to the first half of 2030, and he thinks SpaceX selling compute to Google is the easy first step. [52:18]

8. Rapid reusability is Starship's whole game, and it's a year or two out, maybe sooner. Gavin defines the hard problem as flying and landing the same rocket multiple times per day, not merely reusing it. Even if rivals solve reusability, he expects their mass-to-orbit to asymptote to a small number. His guess for SpaceX getting there: a year or two, maybe sooner. [1:06:49]

9. There's no dollar crisis coming: America is the best house in a bad neighborhood. Gavin's rebuttal to Friedberg's 'Dr. Doom' monologue: high debt is a global problem, and the US is energy and food self-sufficient, now the world's largest oil and gas producer and exporter, and leads in AI, with no better-fundamentaled reserve currency on offer. Rising rates concern him, but the tech bubble happened with the 10-year and 30-year much higher than today, and Nvidia trades at a low-teens multiple of real earnings versus Cisco's 100x. [1:29:54]

10. Once you give a power to the government it's almost never taken back, so don't hand it AI yet. Gavin thinks it's too early for the US to regulate or test frontier models, and he especially dislikes the United States doing it while no one else, China included, does. America already has self-regulation and the courts as recourse. His deeper objection is structural: government power is a one-way ratchet that needs to grow. [31:36]

11. Crime is now a choice. With Flock Safety cameras, gunshot detection and Las Vegas-style drone-and-detection policing available, Gavin argues policing failures are no longer a capability gap but a policy decision. Different municipalities will choose to be pro-crime or anti-crime, and voters will sort it out. [34:30]

12. Sell China the deprecated Nvidia GPUs; it's the best way to keep America ahead. Gavin's counterintuitive read on chip export policy: selling older Nvidia GPUs into China lowers the odds it builds its own, more power-hungry chip ecosystem, which keeps America ahead and in control of AI. He'd pair it with energy leverage: cut off Venezuelan, American and Brazilian oil, keep the Strait closed, and China is left with only Russian oil, in no position to fight a war against the US-led bloc. [1:36:16]

— All-In Podcast, "SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?" (2026-05-22)