All podcasts / Invest Like the Best / Summary

Gavin Baker - AI Market Jitters

2026-08-04 - 66 min - source - Read full transcript

Patrick O'Shaughnessy (host)Gavin Baker

Key insights

July's 40-60% AI-stock drawdown was not accompanied by a single negative quantitative demand metric; every metric (GPU rental pricing, DRAM spot prices, token growth) accelerated.

Baker says he spent the week actively hunting for one negative data point in Silicon Valley and could not find one. GPU rental pricing, DRAM spot prices, and token growth all accelerated through the sell-off, and the one company he cites (a startup renting a Blackwell cluster) expects to pay roughly 50-60% more per GPU-hour seven months from now than today, the opposite of the gentle price decline most investors expected.

ai-compute-economics

The gap between legacy contracted GPU pricing and the current spot market is unusually wide, and as those contracts roll off, the reprice should show up directly in hyperscaler operating cash flow.

Because many neoclouds signed long-term off-take agreements in 2024-25 to finance GPU purchases at prices well below today's spot rates, the installed base of compute is trading at a discount to its current market value. Baker points to Microsoft, Meta, and Amazon's operating cash flow already accelerating from roughly 28% to 32-35% growth this quarter, before new Rubin capacity or contract repricing has even hit the numbers.

ai-compute-economics

Rising open-source model quality (GLM 5.2, Kimi K3) is not bearish for AI infrastructure demand; it shifts token mix from high-margin frontier tokens toward lower-margin open-source tokens without changing the compute required per token.

A token costs the same flops, memory, and watts regardless of which model produced it. The market misread a dip in a frontier-token index as demand weakness when it was really a mix shift toward cheaper open-source tokens, which Baker argues pulls margin dollars out of the model layer and pushes more of them into the infrastructure layer, net expanding compute demand via elasticity.

open-source-model-economics

Widening CDS spreads and rising real yields on AI-related debt are the one legitimate bearish signal Baker found, because a debt-financed buildout is structurally fragile.

Nvidia's CDS and hyperscaler credit spreads blew out, and a recent Meta bond priced worse than expected. Baker treats this as real, drawing an analogy to the dot-com-era telecom buildout: debt-financed capacity expansion demands immediate repayment, so if supply and demand go even slightly out of balance the unwind can be very fast.

credit-financing-risk

If installed compute reprices toward current spot rates, Baker estimates hyperscalers could generate roughly $2 trillion in incremental operating cash flow versus a status-quo baseline, removing an estimated $700 billion of projected credit demand.

Consensus estimates currently model new Blackwell and Rubin capacity to monetize at roughly the rate of the two-generations-old Ampere chip. Baker thinks that is too conservative; if capacity instead monetizes near a discount to current Blackwell spot pricing, the resulting cash flow largely displaces the credit financing the buildout would otherwise need.

credit-financing-risk

Long-term supply agreements (LTAs) have become far stickier than in past hardware cycles because market share is now largely determined by pre-purchased allocation, not by chasing the lowest spot price.

With only a handful of scaled buyers (Amazon, Google, AMD, and a dominant Nvidia) and memory/chip suppliers who remember who honored contracts during the last shortage, breaking an LTA to save money risks permanent loss of allocation priority the next time capacity is scarce. Baker contrasts this with Apple's historical unilateral leverage over suppliers, which does not apply in AI compute's more balanced, multi-buyer market.

ai-compute-economics

Nvidia's new 'credit wrapper' financing model, backing GPU buyers' financing in exchange for a revenue share once prices clear a floor, functions as disguised vendor financing that smooths hyperscalers' negative free cash flow and gives Nvidia a fast-growing royalty stream.

Baker frames this as a logical extension of LTAs: buyers trade upside for durability, and Nvidia (or a financing partner) captures a cut of ongoing revenue on top of the original hardware sale, effectively increasing Nvidia's revenue per gigawatt while also strengthening its competitive moat, since competing chipmakers cannot offer comparably cheap or reliable financing.

ai-compute-economics

Regulatory backlash is the biggest identified risk to the AI buildout, driven largely by the industry's poor public storytelling rather than by the underlying economics.

Baker points to New York's data-center moratorium as an early warning sign and criticizes the industry for letting false narratives (like a book's since-corrected 10,000x overstatement of data-center water usage) spread unchallenged, comparing it to the debunked myth that spinach is unusually high in iron due to a misplaced decimal point ('the Popeye effect').

ai-regulatory-risk

Public-market reactions to AI news have become unusually correlated because most investors now interpret breaking news through the same handful of AI models.

Baker likens this to a probabilistic 'Walter Cronkite' for the stock market: when a headline gets fed into Claude (or a similar model) and a large share of market participants trade on that single AI-mediated interpretation, the normal diversity of investor opinion that dampens overreaction breaks down, which he connects to Michael Mauboussin's theory that diversity breakdown drives bubbles and crashes.

ai-compute-economics

SpaceX is systematically underestimated by public markets as a compute company, having brought on GPU capacity faster and cheaper than any player besides the hyperscalers.

Baker cites SpaceX, CoreWeave, and Crusoe as the only non-hyperscalers to bring on more than 500 megawatts of power in a year, with SpaceX doing it fastest and cheapest. A third-party estimate he cites suggests SpaceX could add roughly 8 gigawatts of compute over 18 months; even a fraction of that, at cited monetization rates of $50-73 billion per gigawatt, would dwarf what he believes is priced into the stock.

spacex-compute-buildout

Solving continual and sample-efficient learning, an active research focus at multiple labs including SSI, could eventually shrink training's share of AI compute demand toward a small asymptote, but Baker does not believe this is net negative for infrastructure.

If models could learn efficiently from far fewer tokens rather than needing hundreds of trillions of training tokens, training demand for compute could structurally decline over time. Baker still expects overall infrastructure demand to grow because inference demand, driven by expanding real-world usage, is a much larger and still-early opportunity.

ai-compute-economics

Companies at the frontier of AI adoption increasingly route queries to fine-tuned open-source models behind a router, cutting per-token cost without necessarily reducing total GPU compute consumed.

Baker describes 'AI-pilled' companies spending 20-50% of total compensation-equivalent budget on tokens, some (like Dylan Patel's company) as high as 30%. Cheaper per-token pricing tends to increase the volume of tokens used, so total GPU-hours consumed can rise even as reported AI spend flattens or declines, because the router shifts volume toward lower-margin open-source tokens that require the same underlying compute per token.

open-source-model-economics

Media referenced

Substack report estimating SpaceX could bring on ~8 gigawatts of compute over 18 months - article - Baker cites this third-party estimate (author not clearly identified in the transcript) to argue the market has not priced SpaceX's compute buildout potential.

Companies

Nvidia - Central to the episode: cheapest forward P/E in a decade despite dominant position; rolled out a 'credit wrapper' financing-plus-revenue-share model for GPU buyers.
Meta - Sold excess compute capacity at a premium (misread by the market as bearish); reported accelerating capex and priced a bond that spooked credit markets.
Microsoft - Brought a large slug of new compute online in June that had not yet shown up in reported operating cash flow, evidence of accelerating, not decelerating, buildout.
Amazon - One of the hyperscalers whose operating cash flow accelerated from the high-20s to low-30s percent this quarter; makes Trainium chips.
Google - Discussed as one of four scaled compute buyers (alongside Amazon, AMD, Nvidia) whose TPU strategy and LTA behavior matters for the chip game theory Baker lays out.
Anthropic - Frontier lab in 'pole position'; third-party data suggesting its growth curve softened slightly was the only negative data point Baker could find all month.
OpenAI - Described as having 're-accelerated' and back in the frontier race alongside Anthropic and xAI/Grok.
xAI / Grok - Grok 4.5 called one of the best models released in two years, contributing to July's frontier-model acceleration.
AMD - One of the four scaled compute players (with Amazon, Google, Nvidia) whose long-term supply agreements now determine market share allocation.
SK Hynix - Memory maker discussed in the LTA game-theory example; breaking a supply agreement with it risks losing allocation in the next up-cycle.
Fireworks AI - Inference cloud growing efficiently; its 'Nexus' product lets companies fine-tune and route to open-source models in three lines of code.
Modal - Cited as one of the open-source inference clouds growing fast with unusually low cash burn.
Together AI - Cited alongside Modal and Fireworks as an inference cloud whose growth the public market underweights because it is private.
Cognition - Its published index showing AI-heavy-spend companies growing faster is cited as bull-case evidence for AI ROI.
Cursor - Cited as accelerating meaningfully after its recent acquisition and as an example of routing tasks between frontier and cheaper models.
Harvey - Legal AI company cited as leaning into fine-tuned open-source routing to become more defensible than a 'wrapper'.
Legora - Legal AI company grouped with Harvey as adopting open-source fine-tuning plus routing.
SpaceX - Major topic: argued to be underestimated by public markets as a compute company after standing up massive compute capacity faster and cheaper than most rivals.
Starlink - SpaceX's satellite business, discussed alongside its data-center ambitions and orbital-compute plans.
StarCloud - Orbital-compute startup funded by Benchmark and partnering with SpaceX's laser tech, cited as validation that orbital compute is becoming real.
Benchmark - VC firm hosting the recording; funded StarCloud, which Baker treats as an independent sanity check on his SpaceX compute thesis.
CoreWeave - One of the few non-hyperscaler companies that has brought on more than 500 megawatts of power in a year.
Crusoe - Cited alongside CoreWeave and SpaceX as a rare non-hyperscaler capable of large-scale power buildout.
ASML - Referenced via the DUV/EUV lithography analogy used to assess China's reported domestic DUV breakthrough.
Etched - AI chip startup Baker is an investor in, used as an example of a company vulnerable if a hyperscaler breaks a supply agreement.

Techniques and frameworks

Long-term agreements (LTAs) - Supply-chain contracts between hyperscalers and chip/memory makers that trade near-term price upside for guaranteed volume; now function like game-theoretic loyalty contracts because breaking one risks losing allocation priority in the next capacity crunch.
Capital cycle theory - Baker frames the credit-risk debate through the classic capital-cycle pattern: debt-fueled buildouts unwind fast if supply and demand go out of balance, as in the dot-com telecom bust.
Mauboussin's diversity-breakdown theory - Referenced to explain unusually correlated market reactions: when most investors run the same news through the same AI model (Claude), the diversity of interpretation that normally dampens overreaction breaks down.
Router-based model orchestration - Companies increasingly route queries between an expensive frontier model and cheaper fine-tuned open-source models, cutting per-token cost without necessarily cutting total compute consumed.
Supervised fine-tuning and reinforcement learning on open-source models - Inference clouds like Fireworks let companies customize open-source models on proprietary data, then serve them behind a router alongside frontier models.
Pre-fill/decode workload disaggregation - Baker's technical thesis that splitting inference into pre-fill, attention, and feed-forward stages, and routing the feed-forward stage to SRAM-based accelerators, meaningfully improves ROI on AI infrastructure.

Summary

Recorded live at Benchmark's offices roughly two months after their prior conversation, Gavin Baker and Patrick O'Shaughnessy dig into what Baker calls the most humbling month he has had pressure-testing his AI infrastructure thesis: a July in which many AI-related stocks fell 40-60% in a straight line while, in Baker's telling, every quantitative demand metric he could find (GPU rental pricing, DRAM spot pricing, token growth) actually accelerated. He walks through the sequence of catalysts, Meta's compute-rental announcement misread as bearish, an open-source model wave (GLM 5.2, Kimi K3) that the market mistook for demand weakness rather than a margin-neutral mix shift, China's reported DUV lithography breakthrough, and finally widening credit spreads, and argues that only the credit signal holds up as a genuinely legitimate concern.

The core of the episode is Baker's reframing of the AI-infrastructure debate around the gap between legacy contract pricing and current spot pricing for compute. Because many neoclouds signed multi-year off-take agreements in 2024-25 at prices well below today's market, he argues the installed base of compute is trading at a steep discount to its true value, and that reported hyperscaler operating cash flow (already accelerating at Microsoft, Meta, and Amazon) will keep climbing as those contracts roll off and reprice. He estimates this repricing could add roughly $2 trillion in incremental cash flow versus a conservative baseline, which would remove hundreds of billions of dollars of projected credit demand, defusing the debt-financed-bubble scenario that most concerns him.

A recurring thread is how supply-chain dynamics have changed the game theory of the chip and memory business. Long-term agreements (LTAs) between hyperscalers and suppliers like Nvidia, AMD, and SK Hynix are now much stickier than in past cycles because market share is increasingly determined by pre-purchased allocation rather than price shopping; breaking an LTA risks permanent loss of priority the next time capacity tightens. Baker connects this to Nvidia's newer "credit wrapper" business, in which Nvidia or a financing partner backs a GPU buyer's financing in exchange for a revenue share once prices clear a floor, a model he expects to expand into a large, high-margin royalty stream layered on top of hardware sales.

Baker and O'Shaughnessy also spend meaningful time on SpaceX, which Baker argues public markets still treat primarily as a rocket and satellite company rather than as a compute company that has demonstrated the ability to stand up massive GPU clusters faster and cheaper than almost anyone besides the hyperscalers. Citing a third-party estimate of roughly 8 gigawatts of new compute over 18 months and per-gigawatt monetization figures in the tens of billions of dollars, he argues the stock reflects little of this potential, a view he stress-tests by noting that Benchmark's own investment in the orbital-compute startup StarCloud, made independently of the SpaceX ecosystem, corroborates the broader orbital-compute thesis.

The conversation closes on risk and narrative. Baker identifies regulation, illustrated by New York's data-center moratorium, as the single biggest threat to the buildout, arguing it stems less from real economic harm and more from the AI industry's failure to counter misinformation (like a book's later-corrected 10,000x overstatement of data-center water usage) before it spreads. He also flags a subtler market-structure risk: because most investors now interpret breaking AI news through the same handful of large language models, market reactions have become unusually correlated, a dynamic he ties to Michael Mauboussin's theory that a breakdown in the diversity of investor opinion is a precursor to bubbles and crashes.

Notable Quotes

"I mean, however you cut it, whether you cut GPU availability, whether you cut GPU rental pricing, whether you cut the spot price of DRAM this month, token growth, everything is actually accelerated." - Gavin Baker

"A token is a token, and you need the exact same amount of compute to make a token, all else equal." - Gavin Baker

"It's kind of Walter Cronkite for the stock market, and everybody just believes whatever it says." - Gavin Baker

"If you do not speak your own truth, no one else will." - Patrick O'Shaughnessy

"I've never seen more companies go from being founded to like $50 million a year in revenue and generating cash flow in like whatever it is, nine months." - Gavin Baker