All podcasts / Invest Like the Best / Summary

Gavin Baker - Watts and Wafers

2026-05-20 - 78 min - source - Read full transcript

Patrick O'Shaughnessy (host)Gavin Baker

Key insights

Capitalism will likely resolve the near-term power shortage by 2027-28, with the binding constraint already shifting from energy and chip supply to zoning and political approval.

A head of data-center infrastructure investing at a major PE firm told Baker that zoning and approval, not energy or chips, are now the biggest gating factors. Turbine makers are announcing capacity expansions despite decades-old manufacturing know-how gaps, and Baker expects the shortage to meaningfully ease as new energy sources and orbital compute come online.

watts-wafers-buildout

TSMC's own capacity discipline, not underlying demand, is the single most important variable determining whether AI infrastructure turns into a bubble.

Baker estimates that if TSMC matched Nvidia's full latent demand, Nvidia could sell $2-3 trillion of GPUs in 2026-27, which he thinks would be an overbuild. TSMC staying deliberately supply-constrained relative to demand is, in his view, the reason a debt-fueled bubble hasn't formed yet; watching how much lead TSMC keeps over Intel and Samsung is his preferred bubble indicator.

watts-wafers-buildout

Every prior foundational technology (railroads, canals, the dot-com internet) has produced a bubble, and AI should be expected to as well, but this buildout differs in being funded largely from operating cash flow rather than debt, and in running near 100% GPU utilization versus ~99% dark fiber left unused after 2000.

Baker cites Carlota Perez's framework on technological revolutions to argue markets correctly identify a paradigm shift early, over-invest, and then crash when supply outruns demand. He calls out the real risk case: a debt-fueled buildout (like telecom in 1999-2000) crashes far harder than a cash-flow-funded one, and points to widening CDS spreads and rising real yields as the one legitimate warning sign he's found.

watts-wafers-buildout

In a genuine shortage, the highest-cost and lowest-quality suppliers rally the hardest, which is producing real cross-sectional mispricing in AI-adjacent stocks right now.

Baker draws the analogy to commodity bull markets, where marginal high-cost producers go from near-bankruptcy to gushing cash and get bid up disproportionately by retail traders on X, while some higher-quality names lag. He cites Astera Labs as a name miscategorized into 'copper loser' baskets despite its switch product benefiting from both copper and optical connectivity.

watts-wafers-buildout

Orbital compute should be understood as networked 'racks in space,' not sci-fi Pentagon-sized structures in orbit.

A unit is roughly the size of a Blackwell rack (about 3,000 lbs) with solar wings roughly 500 feet long per side, kept in a sun-synchronous orbit so panels stay lit continuously; racks link to each other via lasers already deployed on every Starlink satellite, forming a virtual data center rather than one giant structure.

orbital-compute

SpaceX's existing satellite and cooling infrastructure gives it a compounding head start in orbital compute that skeptics are underweighting.

SpaceX already operates roughly 98-99% of all active satellites and cools about 20kW per Starlink V3 satellite versus the ~100kW a Blackwell rack draws; Baker expects that gap to close as SpaceX scales toward 100-120kW satellites, leveraging reusable-rocket economics that no other company has matched a decade after SpaceX first demonstrated it.

orbital-compute

Economic returns in AI have concentrated almost entirely at the frontier model layer, an outcome Baker calls surprising, and whether that premium persists is one of the most important open questions for AI investors.

Despite the availability of cheaper open-source and mid-tier alternatives, an overwhelming share of AI's economic value has accrued to frontier models. Baker notes some erosion at the prototyping stage (companies experimenting with Gemini or open source before production), but says frontier tokens still capture most realized economic value today, with the durability of that premium unresolved.

frontier-token-economics

The AI industry's shift to usage-based pricing (mirroring 1990s-2000s cellular and long-distance telecom) is structurally bullish for revenue growth rather than a sign of demand weakness.

Flat-fee plans effectively rate-limit and 'lobotomize' heavy users, similar to how flat-fee cellular plans capped telecom growth once adopted broadly. Usage-based pricing lets power users (including those running 100 concurrent agents) consume far more tokens, which Baker expects to push Anthropic and OpenAI's combined ARR well above $200 billion this year.

frontier-token-economics

A violation of Richard Sutton's 'bitter lesson' -- more compute beating human algorithmic ingenuity -- is the single biggest risk Baker sees to the AI trade, though he's less worried than most because approaching ASI could produce a temporary counter-violation.

He cites the TurboQuant episode, a Google DRAM memory-optimization paper that went viral on X and briefly spooked memory demand expectations, as an example of the market overreacting to an algorithmic-efficiency story; he could not find a single AI engineer who thought it would meaningfully dent DRAM demand. His counterargument is that a model near ASI might temporarily prioritize making itself more efficient, since intelligence and resource efficiency compound together at that point.

frontier-token-economics

Chip startups almost always lose if they try to build 'a better GPU'; the only durable path is doing something both different from Nvidia's design tradeoffs and hard to replicate.

Baker's 'iron triangle' analogy: every chip design lives within physics-imposed tradeoffs, and Nvidia sees every foundry's process roadmap before a 200-person startup does, so it can fast-follow any approach that reaches even 1-3% market share (which Baker estimates is worth roughly $100B). Cerebras is his example of doing something genuinely different and hard: wafer-scale computing, which took three chip generations to get right.

chip-architecture-strategy

Disaggregating inference into prefill and decode stages extends the useful economic life of older GPUs to 10-15 years and materially de-risks GPU-backed private credit.

Prefill (processing the prompt/context) is memory-capacity-bound while decode (generating output tokens) is memory-bandwidth-bound; pairing specialized accelerators (like Cerebras or Groq-style chips) for one stage lets older Hopper- or Ampere-generation GPUs keep running the other stage well past prior assumptions of a 1-4 year useful life, which Baker argues can lower AI infrastructure financing costs from the low-single-digit-percent range toward 5-6%.

chip-architecture-strategy

A durable AI-native venture idea has to be both non-obvious and hard to execute before it can reach scale; being merely non-obvious is not enough once an incumbent with scale advantages notices.

Baker's Amazon analogy: retail incumbents dismissed Amazon's e-commerce model as obvious and easy to copy, but Amazon's operational execution was hard enough that incumbents that tried to compete on margin still failed. He applies the same filter to today's AI-native founders and says many are struggling to find niches the frontier labs either won't or structurally can't absorb.

ai-application-layer-strategy

Books referenced

Technological Revolutions and Financial Capital - Carlota Perez - Baker cites her framework for why every foundational new technology (railroads, canals, the internet) produces a bubble as capital chases a correctly-identified paradigm shift, ahead of demand catching up to supply.

Media referenced

The Last Samurai - movie - Baker rewatched it and had his firm watch it too, using the film's 'master the machine gun' arc as his framework for how an experienced investor stays relevant by integrating AI into a decades-honed process rather than being replaced by it.

Companies

Atreides Management - Baker's firm; this is his sixth conversation on the show. He describes positioning it for the AI buildout and for a possible 'Mythos 3/4' world by overinvesting in cybersecurity and instituting family/company safe words against AI voice-cloning scams.
Positive Sum - Patrick O'Shaughnessy's firm, disclosed at the top of the episode as a standard compliance note.
Anthropic - Central example of the AI buildout's speed: added roughly $11B of ARR in a single month, more than the combined decade-plus build of Palantir, Snowflake, and Databricks; Baker estimates it burns ~80% less cash than OpenAI per dollar of revenue and would be running well north of $100-150B ARR if compute weren't the binding constraint.
OpenAI - Compared unfavorably to Anthropic on capital efficiency and cost-per-token, though Baker expects both to clear $200B+ combined ARR this year as usage-based pricing rolls out.
Databricks - Cited alongside Snowflake and Palantir as one of the three highest-profile SaaS companies of the last decade, whose combined decade of value creation Anthropic matched in one month.
Snowflake - Same SaaS-era comparison point as Databricks and Palantir.
Palantir - Same SaaS-era comparison point as Databricks and Snowflake.
SpaceX - Baker's orbital-compute thesis centers on it: operates ~98-99% of all active satellites, already cools 20kW per Starlink satellite, and is building the TerraFab semiconductor plant with Tesla and Intel.
Tesla - Named as a co-venturer with SpaceX on TerraFab, the planned largest fab in America, benefiting from Elon Musk's ability to recruit top hardware engineers.
Intel - TerraFab's foundry partner, providing institutional fabrication knowledge roughly one node (9-15 months) behind the leading edge; also a foundry competitor to TSMC whose capacity discipline Baker is watching for signs of a break.
Samsung - The other foundry competitor to TSMC; Baker argues that if either Samsung or Intel breaks pricing/capacity discipline, it forces the others to follow, which is his key risk case for an eventual AI hardware bubble.
TSMC - Described as the single most important gating factor preventing an AI bubble: its wafer capacity discipline keeps Nvidia's addressable GPU sales well below the $2-3T Baker estimates would occur under unconstrained supply.
Nvidia - Discussed as constrained by TSMC allocation, positioned to fast-follow any chip startup that reaches meaningful market share, and trading at what Baker calls historically cheap relative valuations as of early April.
Google - Lost its per-token cost leadership after conservative TPU design choices; still holds the largest installed compute base and richest data (including YouTube), with Gemini's ability to hold the Pareto Frontier seen as an open question heading into Google I/O.
xAI / Grok - Grok 4.3 called the best low-cost, ~500B-parameter model on the Pareto Frontier at the time of the conversation.
Cerebras - Atreides is a venture investor; held up as the model for how a chip startup should compete with Nvidia, wafer-scale computing being different and hard rather than an incremental 'better GPU.'
Groq - Referenced as a specialized inference accelerator (LPU) example in the disaggregated prefill/decode pipeline discussion.
AMD - Grouped with Trainium as GPU-alternative efforts; Baker says less is known about AMD's upcoming MI-450 than about Trainium 3, calling AMD's competitive position the harder one to read.
Broadcom - Positioned as the industry's default ASIC design partner; landing a first-generation chip deal with Broadcom is described as 'manna from heaven' for a chip startup.
Microsoft - Satya Nadella praised for a 'courageous' decision to redirect GPU capacity from selling to OpenAI/Anthropic toward Microsoft's own products, even though the market has punished the stock for it; Baker estimates Microsoft would be an ~$800 stock if it had kept selling capacity externally instead.
Amazon - Seen as well positioned via Trainium and expected P&L efficiencies from robotics in its retail business; Baker says its internal Nova models are better than they get credit for, and it is (with Nvidia) the most engaged of the big tech companies with AI startups.
Meta - Credited as the only 'internet giant' to fully become AI-first internally; its first MSL-trained model, Muse, was described as a surprisingly strong near-Pareto-Frontier result.
Astera Labs - Example of a cross-sectionally mispriced name: grouped by the market into 'copper loser' baskets even though its core switch product connects accelerators using both copper and optics, meaning it benefits regardless of which connectivity technology wins.
Cursor - Cited, with Cognition and early Anthropic, as an example of an application-layer company that correctly focused narrowly on coding roughly 18 months before this conversation, when OpenAI was still spreading itself across everything.
Cognition - Cited alongside Cursor as a coding-focused winner; Baker says it is now attempting something 'really, really different' from a pure coding-agent product.
Replit - Baker cites founder Amjad Masad's observation that coding may be the shortest path to ASI, since a sufficiently capable coding model can write the tools it needs to do anything else.

Techniques and frameworks

Iron triangle of chip design - Baker's tank-design analogy (attack/defense/mobility tradeoffs) applied to chips: every design lives within physics- and process-imposed tradeoffs, so a startup chasing 'a better GPU' inside the same tradeoff space will always lose to Nvidia's scale and foundry access.
Prefill/decode disaggregation - Splitting inference into prefill (memory-capacity-bound, 'loading the cannon') and decode (memory-bandwidth-bound, 'firing') lets chip designers make different, non-Nvidia tradeoffs for each stage, and lets older GPUs (Hopper, Ampere) be paired with newer accelerators to extend their useful economic life to 10-15 years.
Pareto Frontier of intelligence vs. cost - Baker's primary lens for ranking AI labs: plotting model intelligence against per-token cost; leadership on this frontier shifted from Google (nine months prior) to Anthropic, OpenAI, and Grok 4.3 by the time of this conversation.
Richard Sutton's 'bitter lesson' - The empirical finding that more compute and data reliably beats human algorithmic cleverness; Baker calls a violation of this the single biggest risk to the AI trade, tempered by the possibility that an ASI-capable model could temporarily out-optimize the compute-scaling curve by improving its own efficiency.
Diversity breakdown (Mauboussin) - The market-structure idea that when nearly all investors become bullish on the same paradigm-shifting technology at once, price-correcting disagreement disappears, which Baker says is exactly what feeds a bubble; he says he is starting to see early signs of it in AI sentiment.
'Different and hard' venture framework - Baker's filter for durable venture ideas: an idea that is merely non-obvious isn't enough, because an incumbent with scale will copy it once it's proven; it must also be hard to execute, the way Amazon undercut e-commerce rivals or Cerebras built wafer-scale compute.
The 'token path' (Jamin Ball) - Concept borrowed from Altimeter's Jamin Ball: software and AI companies need to be positioned so that growth in AI token usage flows through their business (e.g., Databricks); companies outside the token path, and outside a narrow enough defensible niche, face structurally harder economics.

Summary

In his sixth conversation with Patrick O'Shaughnessy, Atreides Management's Gavin Baker frames the next phase of AI around two physical constraints: watts and wafers. He opens by revisiting the extraordinary March-April stretch, when Anthropic added roughly $11 billion of ARR in a single month, more value creation than Palantir, Snowflake, and Databricks built combined over a decade, even as AI stocks sold off on unrelated macro noise (the Strait of Hormuz closure, which Baker argues was actually net-positive for US relative manufacturing competitiveness via cheaper domestic natural gas). He treats that drawdown as a buying opportunity rather than a warning sign, echoing the market's earlier overreaction to the original DeepSeek sell-off.

On watts, Baker expects capitalism to resolve the near-term power shortage by 2027-28 as new energy sources and, eventually, orbital compute come online; the binding constraint has already shifted from energy and chip supply toward zoning and political approval. He spends significant time reframing SpaceX's orbital-compute ambitions as networked "racks in space," roughly Blackwell-rack-sized units with long solar wings in sun-synchronous orbit, linked by the same laser technology already deployed across the Starlink fleet, rather than the sci-fi mega-structures skeptics imagine. He argues SpaceX's existing satellite fleet, cooling capacity, and unmatched reusable-launch economics give it a compounding head start that most public-market observers are underweighting.

On wafers, Baker's central thesis is that TSMC's own capacity discipline, not underlying demand, is the single most important variable preventing an AI hardware bubble. He estimates Nvidia could sell $2-3 trillion of GPUs in 2026-27 under unconstrained supply, an outcome he'd consider an overbuild; TSMC staying deliberately supply-constrained relative to Intel and Samsung is, in his telling, quietly doing the market a favor. He layers in Carlota Perez's framework on technological revolutions to argue every foundational technology produces a bubble, but that this cycle differs from 2000 by being funded largely out of operating cash flow rather than debt, running near-full GPU utilization rather than mostly-dark fiber, and by SpaceX and Tesla's TerraFab joint venture (with an Intel manufacturing partnership) adding a credible third source of leading-edge capacity.

The conversation moves through frontier-model economics: why economic value has concentrated so heavily at the frontier layer, the shift to usage-based pricing (which Baker compares to the shift from flat-fee to metered cellular and long-distance plans in the 1990s-2000s, expecting it to push Anthropic and OpenAI's combined ARR well past $200 billion this year), and why a violation of Richard Sutton's "bitter lesson" (that more compute beats algorithmic cleverness) is his top tail risk to the whole AI trade, tempered by a case that an ASI-capable model could temporarily out-optimize that curve by improving its own efficiency. On chip architecture, he lays out an "iron triangle" analogy for why most chip startups fail when they simply try to build "a better GPU," since Nvidia can fast-follow any approach that gains meaningful share; the durable path, exemplified by Atreides portfolio company Cerebras, is doing something both different from Nvidia's design tradeoffs and genuinely hard to replicate. He connects this to a practical financial point: disaggregating inference into prefill and decode stages extends older GPUs' useful economic life to 10-15 years, which he argues meaningfully de-risks GPU-backed private credit.

Baker closes with a rapid tour of the major hyperscalers (Google's compute and data advantages despite losing TPU cost leadership, Meta's genuine internal AI-first transformation under Zuckerberg, Amazon's Trainium and robotics position, and Microsoft's Satya Nadella making what Baker calls a courageous, market-punished decision to redirect GPU capacity toward Microsoft's own products), his venture framework for judging AI-native startups (ideas must be both non-obvious and hard to execute before an incumbent can copy them at scale), and a set of more speculative closing thoughts on personal safety risk for high-profile AI figures, the geopolitical stakes of America's AI lead, and cautious optimism that AI's benefits (illustrated by a personal story about a rare pediatric disease) will outweigh its disruptions if navigated with humility.

Notable Quotes

"There's just no precedent for this... we tech investors, we hear a lot of discussions about S curves and investing in exponentials. I've just never seen an exponential like this." - Gavin Baker

"If we don't get a bubble, we need to throw a party for [TSMC], because they will have single-handedly prevented a bubble." - Gavin Baker

"I would like to redefine orbital compute [as] racks in space. Not giant, glowing, Pentagon-sized data centers in space. That's silly." - Gavin Baker

"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

"The machine gun is here. If we do not all become masters of the machine gun... I am trying to become a master of the machine gun." - Gavin Baker