Gavin Baker - Watts and Wafers — structured breakdown

TAKEAWAYS

My biggest takeaways from Gavin Baker, CIO of Atreides Management

01

If we don't get an AI bubble, TSMC will have single-handedly prevented it.

Gavin's Wafers thesis: 200 to 300 years of precedent (railroads, canals, South Sea, the internet) say a foundational technology should produce a bubble, and the only thing stopping this one is TSMC's wafer scarcity. If Taiwan Semi did what Jensen wanted, he thinks NVIDIA could sell $2 to 3 trillion of GPUs in 2026 or 2027, past the point where demand becomes an overbuild. He expects Intel or Samsung to break ramp discipline eventually and force everyone else to follow, so the one thing he watches for a bubble is TSMC's capacity decisions.

“So Taiwan Semi, if we don't get a bubble, we need to throw a party for them because they will have single-handedly prevented a bubble.”
Gavin Baker · 24:00
02

Every GPU is running at 100% utilization; in 2000, 99% of the fiber was dark.

Gavin says the reason this buildout is not the dot-com bubble is structural, not sentimental. The spending is overwhelmingly funded out of operating cash flow, and every GPU that ships is running at full utilization. The year-2000 fiber buildout had the opposite shape: almost none of the capacity was used, which is why he still expects a bubble on precedent but says this one looks nothing like the last.

“every GPU is running at 100% utilization when 99% of fiber was unutilized.”
Gavin Baker · 24:00
03

AI has net destroyed trillions of dollars of value at the application layer.

Gavin's contrarian ledger: profits are accruing to energy, data centers, chips and models, and even after counting Cursor and Cognition, the most successful AI natives, the application layer is net negative by trillions. The SaaS and cloud revolution created $5 to 10 trillion of value; this wave is eating it. The winners he sees are the companies with the highest ratio of utilized GPUs per human.

“AI has net destroyed, even if you count Cursor, Cognition, the most successful AI natives, trillions of dollars of value has been destroyed by AI at the application layer.”
Gavin Baker · 1:12:02
04

If you're a software company or an AI company of any kind, you have to be in the token path.

Gavin's survival test for founders: the idea has to be non-obvious, different, and hard to execute, and if it becomes obvious to the world before you reach scale, you're in trouble. Coding companies won by getting to scale before the labs turned on them. Outside the token path only a truly niche, defensible area works, and he is dubious even of vertical data moats, because that data comes from humans and the frontier labs can undercut it. His one hopeful caveat: if the returns to frontier tokens relative to other tokens come down, value creation explodes at the application layer.

“if you're a software company or an AI company of any kind, you have to be in the token path.”
Gavin Baker · 51:29
05

A violation of the bitter lesson is the biggest risk to the whole AI trade.

Gavin names one tail risk above all others: human algorithmic ingenuity beating raw compute, which would break the case for the entire buildout. He notes the people closest to the models are the most skeptical it will happen. He is a little less skeptical himself only because he thinks we are very close to ASI, and nobody knows whether the bitter lesson holds for a 400 IQ model that starts making itself more efficient.

“a violation of Richard Sutton's bitter lesson is for sure the biggest risk to this trade, to all of AI.”
Gavin Baker · 31:17
06

Usage-based pricing is why OpenAI and Anthropic will exceed $200 billion in ARR this year.

Gavin reads the move from all-you-can-eat plans to pay-by-the-drink as deeply bullish for the labs, the same way cellular and long-distance grew under metered pricing. Combine that with more compute coming online and he puts both labs well past $200 billion in ARR by the end of 2026. He also dismisses the TurboQuant scare about DRAM demand: he could not find a single AI engineer who believed it would matter.

“I think the shift to usage based pricing is probably why you will see OpenAI and Anthropic exceed well over $200 billion in ARR this year.”
Gavin Baker · 36:20
07

Anthropic at $50 billion in ARR is a compute-constrained company that could be doing $150 billion.

Gavin argues the rumored $900 billion mark on roughly $50 billion of ARR misreads the business, because Anthropic is selling every token it can produce. Unconstrained, he puts it well north of $100 billion, maybe $150 billion, which makes the mark closer to 5x revenue. The evidence is in the product: Opus is generating 70% fewer tokens for the same question, an effective rationing of intelligence. Anthropic also burned roughly 80% less capital than OpenAI to get to similar scale, which he calls a different animal on capital efficiency.

“if Anthropic could just wave a magic wand and get all the compute they wanted, they'd probably be doing well north of $100 billion today, maybe $150 billion.”
Gavin Baker · 10:41
08

Orbital compute is racks in space connected by lasers, not floating pentagon-sized data centers.

Gavin's Watts thesis has two clocks. Near term, capitalism solves the power shortage starting in 2027 to 2028, with zoning and approval now a bigger gate than energy or chips. Long term, a Blackwell rack is 3,000 pounds and eight feet by four by three, and a satellite that size with 500-foot solar wings and a radiator is what SpaceX, which already cools every Starlink and runs 98 to 99% of satellites in orbit, says it has solved. Inference goes to orbit; training stays on Earth, so he is not bearish on terrestrial data centers.

“I would like to redefine orbital compute has racks in space, not giant, floating pentagon-sized data centers in space. That's silly. What makes a data center is you're connecting these racks with lasers.”
Gavin Baker · 14:39
09

Disaggregated inference gives GPUs 10-to-15-year lives, and that may single-handedly save private credit.

Gavin's answer to the depreciation bears: once prefill and decode are split, an older GPU gets repurposed for prefill behind a newer decode system instead of being scrapped, so useful life is 10 to 15 years rather than the 1 to 4 skeptics assume. Lenders are underwriting to 3 or 4 years and CoreWeave's cheapest financing was in the low 7% range; longer lives could bring that to 5 to 6%, which changes the cost of the whole buildout. It also makes owners of installed bases, the sellers of shortage, the advantaged party in an agentic world.

“The disaggregation of inference means that I think these GPUs are going to have 10 or 15 year lives.”
Gavin Baker · 46:36
10

If you're trying to make a better GPU, good luck. Do something different, and hard.

Gavin's chip-startup rule borrows the tank designer's iron triangle: NVIDIA gets the best TSMC prices and will fast-follow any easy trade-off, so a challenger has to pick something different that is also hard to copy. Cerebras qualifies with wafer-scale computing, which took three chip generations to get right and lets it run any size model. Among GPU-shaped challengers he ranks Trainium ahead of AMD's MI450 because Trainium 3 has the switch scale-up network needed to inference MoE models economically. His venture math: 1% of AI chip share is worth about $100 billion.

“if you're trying to make a better GPU, good luck. If you are doing something different, it also needs to be hard to do.”
Gavin Baker · 40:21
11

SemiCap at 40x earnings and DRAM at mid-single digits cannot both be true.

Gavin's biggest worry is cross-sectional, not the index: semicap equipment companies trade at about 40 times next quarter's annualized earnings while DRAM companies trade at mid-single digits, a gap far wider than the 5-versus-12 seen at prior cycle peaks and not justified by the business models. Because we are in shortages, the lowest-quality, highest-cost commodity suppliers are doing best and getting bid up on X, which reverses in the downturn. And he cannot find anyone like him who is not bullish on DRAM, a diversity breakdown he says historically precedes a bubble.

“you have SemiCAP equipment companies trading at 40 times next quarter's annualized earnings and DRAM companies trading at mid-single digit.”
Gavin Baker · 1:00:06
12

To understand what frontier AI can do today, you need Claude Code or Codex on a usage-based plan.

Gavin's tactic for investors: the fixed-price consumer plans serve a rate-limited, lobotomized version of the models, and anyone judging AI from that will get the intuition wrong. Even for a non-coding use case, he says you have to be on an enterprise usage-based plan with Claude Code or Codex to see the real capability. He calls it a little dystopian that the best AI is now only available to people with a lot of money, and says that needs solving.

“To understand what Frontier AI is capable of today, even for a non-coding use case, you need to have Cloud Code or Codex.”
Gavin Baker · 36:20

post Copy as a post — the same 12 takeaways as plain text, timestamps included

My biggest takeaways from Gavin Baker, CIO of Atreides Management:

1. If we don't get an AI bubble, TSMC will have single-handedly prevented it. Gavin's Wafers thesis: 200 to 300 years of precedent (railroads, canals, South Sea, the internet) say a foundational technology should produce a bubble, and the only thing stopping this one is TSMC's wafer scarcity. If Taiwan Semi did what Jensen wanted, he thinks NVIDIA could sell $2 to 3 trillion of GPUs in 2026 or 2027, past the point where demand becomes an overbuild. He expects Intel or Samsung to break ramp discipline eventually and force everyone else to follow, so the one thing he watches for a bubble is TSMC's capacity decisions. [24:00]

2. Every GPU is running at 100% utilization; in 2000, 99% of the fiber was dark. Gavin says the reason this buildout is not the dot-com bubble is structural, not sentimental. The spending is overwhelmingly funded out of operating cash flow, and every GPU that ships is running at full utilization. The year-2000 fiber buildout had the opposite shape: almost none of the capacity was used, which is why he still expects a bubble on precedent but says this one looks nothing like the last. [24:00]

3. AI has net destroyed trillions of dollars of value at the application layer. Gavin's contrarian ledger: profits are accruing to energy, data centers, chips and models, and even after counting Cursor and Cognition, the most successful AI natives, the application layer is net negative by trillions. The SaaS and cloud revolution created $5 to 10 trillion of value; this wave is eating it. The winners he sees are the companies with the highest ratio of utilized GPUs per human. [1:12:02]

4. If you're a software company or an AI company of any kind, you have to be in the token path. Gavin's survival test for founders: the idea has to be non-obvious, different, and hard to execute, and if it becomes obvious to the world before you reach scale, you're in trouble. Coding companies won by getting to scale before the labs turned on them. Outside the token path only a truly niche, defensible area works, and he is dubious even of vertical data moats, because that data comes from humans and the frontier labs can undercut it. His one hopeful caveat: if the returns to frontier tokens relative to other tokens come down, value creation explodes at the application layer. [51:29]

5. A violation of the bitter lesson is the biggest risk to the whole AI trade. Gavin names one tail risk above all others: human algorithmic ingenuity beating raw compute, which would break the case for the entire buildout. He notes the people closest to the models are the most skeptical it will happen. He is a little less skeptical himself only because he thinks we are very close to ASI, and nobody knows whether the bitter lesson holds for a 400 IQ model that starts making itself more efficient. [31:17]

6. Usage-based pricing is why OpenAI and Anthropic will exceed $200 billion in ARR this year. Gavin reads the move from all-you-can-eat plans to pay-by-the-drink as deeply bullish for the labs, the same way cellular and long-distance grew under metered pricing. Combine that with more compute coming online and he puts both labs well past $200 billion in ARR by the end of 2026. He also dismisses the TurboQuant scare about DRAM demand: he could not find a single AI engineer who believed it would matter. [36:20]

7. Anthropic at $50 billion in ARR is a compute-constrained company that could be doing $150 billion. Gavin argues the rumored $900 billion mark on roughly $50 billion of ARR misreads the business, because Anthropic is selling every token it can produce. Unconstrained, he puts it well north of $100 billion, maybe $150 billion, which makes the mark closer to 5x revenue. The evidence is in the product: Opus is generating 70% fewer tokens for the same question, an effective rationing of intelligence. Anthropic also burned roughly 80% less capital than OpenAI to get to similar scale, which he calls a different animal on capital efficiency. [10:41]

8. Orbital compute is racks in space connected by lasers, not floating pentagon-sized data centers. Gavin's Watts thesis has two clocks. Near term, capitalism solves the power shortage starting in 2027 to 2028, with zoning and approval now a bigger gate than energy or chips. Long term, a Blackwell rack is 3,000 pounds and eight feet by four by three, and a satellite that size with 500-foot solar wings and a radiator is what SpaceX, which already cools every Starlink and runs 98 to 99% of satellites in orbit, says it has solved. Inference goes to orbit; training stays on Earth, so he is not bearish on terrestrial data centers. [14:39]

9. Disaggregated inference gives GPUs 10-to-15-year lives, and that may single-handedly save private credit. Gavin's answer to the depreciation bears: once prefill and decode are split, an older GPU gets repurposed for prefill behind a newer decode system instead of being scrapped, so useful life is 10 to 15 years rather than the 1 to 4 skeptics assume. Lenders are underwriting to 3 or 4 years and CoreWeave's cheapest financing was in the low 7% range; longer lives could bring that to 5 to 6%, which changes the cost of the whole buildout. It also makes owners of installed bases, the sellers of shortage, the advantaged party in an agentic world. [46:36]

10. If you're trying to make a better GPU, good luck. Do something different, and hard. Gavin's chip-startup rule borrows the tank designer's iron triangle: NVIDIA gets the best TSMC prices and will fast-follow any easy trade-off, so a challenger has to pick something different that is also hard to copy. Cerebras qualifies with wafer-scale computing, which took three chip generations to get right and lets it run any size model. Among GPU-shaped challengers he ranks Trainium ahead of AMD's MI450 because Trainium 3 has the switch scale-up network needed to inference MoE models economically. His venture math: 1% of AI chip share is worth about $100 billion. [40:21]

11. SemiCap at 40x earnings and DRAM at mid-single digits cannot both be true. Gavin's biggest worry is cross-sectional, not the index: semicap equipment companies trade at about 40 times next quarter's annualized earnings while DRAM companies trade at mid-single digits, a gap far wider than the 5-versus-12 seen at prior cycle peaks and not justified by the business models. Because we are in shortages, the lowest-quality, highest-cost commodity suppliers are doing best and getting bid up on X, which reverses in the downturn. And he cannot find anyone like him who is not bullish on DRAM, a diversity breakdown he says historically precedes a bubble. [1:00:06]

12. To understand what frontier AI can do today, you need Claude Code or Codex on a usage-based plan. Gavin's tactic for investors: the fixed-price consumer plans serve a rate-limited, lobotomized version of the models, and anyone judging AI from that will get the intuition wrong. Even for a non-coding use case, he says you have to be on an enterprise usage-based plan with Claude Code or Codex to see the real capability. He calls it a little dystopian that the best AI is now only available to people with a lot of money, and says that needs solving. [36:20]

— Invest Like the Best, "Gavin Baker - Watts and Wafers" (2026-05-20)

.txt .json