StockResearch
All posts

Gavin Baker's 'No Negative Quantitative Metric': the Checklist Version

During the worst week of the July 2026 AI selloff, Gavin Baker said he could not find a single negative quantitative metric. That is a checkable claim. Here is the checklist: every metric he cited, what it signals, and an honest accounting of which ones a retail investor can actually verify.

Gavin BakerAI fundamentals checklistGPU pricesDRAM spot pricehyperscaler capextoken growthAI investing metricsInvest Like the Best
By Jake, founder of StockResearch.app

In the week the AI trade was blowing up, the same week a leveraged fund was force-selling a $16–20B book to Citadel, Gavin Baker sat down with Patrick O'Shaughnessy and made a claim that is unusual for a podcast: a falsifiable one.

After spending a week in Silicon Valley explicitly hunting for negatives, Baker said he could not find a single negative quantitative metric in the AI buildout. Not one. GPU availability, GPU pricing, DRAM spot, token growth, hyperscaler cash flows: all accelerating, while the stocks fell 40–60%. The episode is worth the full 79 minutes: "Why the Markets Are Pricing AI Wrong", Invest Like the Best, August 4, 2026.

What makes the claim useful is not that Baker is bullish; a fund manager talking his book is not news. It is that "no negative quantitative metric" implies a list of metrics, each of which can move against him. That list is a monitoring dashboard for the entire AI thesis, usable by bulls and bears alike. So here is the checklist version: every metric he cited, what direction means what, and which ones you can honestly verify without a hedge fund's rolodex.

The bull column: what Baker says is accelerating

MetricBullish readingBearish readingCan retail verify?
GPU rental/spot pricingRising, old GPUs repricing upSustained sharp declinesPartially
DRAM spot pricesRising (memory supply war)Rolling overMostly yes
GPU availability / lead timesScarce, allocation-constrainedGPUs easy to getPartially
Hyperscaler operating cash flow growthAcceleratingDeceleratingYes
Token growth (inference demand)AcceleratingPlateauingPartially
1. GPU rental and spot pricing. Baker's central non-consensus claim. In 2024–25, everyone modeled GPU rental prices declining on a depreciation curve; instead, old GPUs repriced upward. His anecdotes: a startup renting B200s at mid-$2 per GPU-hour expecting just under $4 at renewal seven months later; an inference cloud budgeting +100% at renewal. His structural point: the contracted installed base sits at a large discount to spot, so compute revenue repricing continues as contracts roll even if spot softens. Verification: partial. Public GPU marketplaces (Vast.ai, RunPod, Lambda) publish hourly rates you can track in a spreadsheet, and cloud providers post on-demand pricing. What you cannot see is the contract-renewal market where Baker's most dramatic numbers live: those are private negotiations, and you are trusting his anecdotes. 2. DRAM spot prices. Memory-per-flop is, in Baker's telling, the dominant axis of token output, and memory is in a supply war with long-term agreements (LTAs) locking capacity for the four buyers who matter (Amazon, Google, AMD, Nvidia; together bigger than everyone else combined). Spot DRAM pricing is the cleanest public thermometer. Verification: mostly yes. TrendForce and DRAMeXchange publish spot and contract price data, and the financial press covers inflections. A retail investor can genuinely track this weekly. 3. GPU availability. "GPUs becoming easy to get" is on Baker's own list of things that would scare him. Verification: partial. Cloud instance availability, marketplace listing depth, and management commentary on supply constraints are observable; true lead times are industry gossip. 4. Hyperscaler operating cash flow growth. Baker's anchor statistic: aggregate hyperscaler operating cash flow growth went 28% to 32% as reported this quarter, roughly 28% to 35% adjusting one-timers: an acceleration, before the next GPU generation (Rubin) ships at premium pricing. His credit rebuttal runs through the same line item: consensus models new gigawatts monetizing at roughly Ampere-era rates (~$1.3–1.4T of opex cash flow); at even a discount to current Blackwell rates it is closer to $2T, which would remove roughly $700B of prospective credit demand. Verification: yes, fully. Operating cash flow is a GAAP line in every 10-Q. Four companies, four filings a quarter. This is the single most retail-verifiable metric on the list, and arguably the most important: it is where "AI demand is real" either shows up in audited numbers or does not. 5. Token growth. Inference volume growth is the demand signal beneath everything. Verification: partial. OpenRouter publishes public token-volume rankings, and Google and Microsoft disclose token-processing figures at events. Directionally useful, selectively disclosed, not audited.

The bear column: what Baker says would worry him

A checklist that only tracks confirmations is a mood board. Baker names credit as the one bear case he respects, so the bear column gets equal standing:

MetricBearish readingCan retail verify?
Hyperscaler CDS spreadsBlowing outBarely
Corporate credit spreads / real yieldsWidening / risingYes
AI-related bond dealsPricing poorly (e.g., Meta's bond)Mostly yes
Data-center regulationSpreading moratoriaYes
Real yields up, spreads wider, hyperscaler CDS deteriorating, Meta's bond pricing poorly: debt-financed buildouts unwind violently when monetization disappoints; that is the dot-com pattern. Credit spreads and real yields are free on FRED; bond-deal reception gets covered in the financial press; CDS levels are effectively institutional data, so retail sees them only when journalists do. Regulation, Baker's biggest structural fear (with New York's data-center moratorium as "the first of many"), is trackable through ordinary news flow. One July detail cuts both ways: the credit stress that actually bit hit the investor layer (prime-broker margin calls), not the hyperscaler balance sheets. The bear case was right about leverage, wrong about the address.

Baker's own falsifiers

To his credit, Baker states what would change his mind: operating cash flow growth failing to keep accelerating; a sustained sharp drop in GPU spot pricing; GPUs becoming easy to get; total lab revenue plateauing for reasons other than open-source share shift. Add his regulatory fear, and you have five explicit tripwires. When a bull hands you his falsifiers, the productive response is not applause. It is monitoring.

Honest limitations of the checklist

The strongest claims rest on private data. The GPU-renewal anecdotes that anchor the repricing thesis are unverifiable by design. You are choosing how much weight to give secondhand quotes from a manager who is long the thesis. A checklist is not position sizing. The cautionary tale is standing right there: Situational Awareness ran essentially Baker's exact thesis (memory, neoclouds, power, GPU repricing, name for name) and was arguably right on every checklist item while losing 67% in a month, because it expressed the view at 4x leverage with 76% of the book in five names. Baker runs the same conviction unlevered and preaches "I don't know" position sizing. The checklist tells you about the thesis. It tells you nothing about how much of it anyone should own. That is a risk question, and July graded it separately from the fundamentals question. Metrics lag at turns. Every item above described an accelerating buildout in July 2026. A cycle top would, by definition, start with these same metrics still glowing green. The checklist detects deterioration; it does not predict it.

The meta-point: this should be a screen, not a podcast

Step back and Baker's whole argument has a shape: the multiple compressed while the fundamentals accelerated, the pattern we unpacked in the Nvidia forward P/E post. His evidence for the fundamentals half is anecdotal by design; the quantitative half (where a stock's multiple sits against its own history, and whether the underlying trajectory is actually decelerating) should not require a Benchmark office visit to check. That two-axis view is what the valuation history tools on StockResearch are built for, and making "compression without deceleration" a self-serve screen across every AI-infrastructure name is exactly where we are taking the product.

Until then, the checklist above is the manual version. Ten metrics, five falsifiers, four 10-Qs a quarter. The question Baker's episode leaves open is not "is he right?" It is: when one of his tripwires fires, will you notice before the market does?


Jake is the founder of StockResearch.app, where he writes data-first research on valuation dislocations. This article summarizes claims made by Gavin Baker on the cited podcast; those claims are his, not independently verified. It is for informational purposes only and does not constitute financial advice. Nothing here is a recommendation to buy or sell any security. Always do your own research before making investment decisions.
Gavin Baker's 'No Negative Quantitative Metric': the Checklist Version | StockResearch