Gavin Baker — Why the Markets Are Pricing AI Wrong Investment Thesis

Source: Why the Markets Are Pricing AI Wrong | Gavin Baker, Invest Like the Best (Patrick O'Shaughnessy), August 4, 2026.

The Framework: Ground Truth vs. Narrative

Baker describes July 2026 as "2022 in a month" — a violent sell-off in which AI stocks dropped 40–60% from highs in a straight line. His organizing lens: pressure-test every bear thesis with on-the-ground quantitative data before selling. He found none. The single thread holding his bull case together is the Repricing Flywheel — installed compute was contracted at 2024 discount prices, and as those contracts roll off, hyperscaler operating cash flows will jump sharply enough to self-fund the entire CapEx buildout without debt, eliminating the credit risk bear case from the inside.

Bear Narrative	Ground Reality (Baker's Finding)	Investment Implication
Open source is crushing frontier revenue	Token = token; same flops/watts/memory regardless of model source; margin shifts to infra layer	Infrastructure demand structurally intact
AI CapEx needs debt; CDS widening	Contracted compute reprices to spot → OCF flywheel funds buildout	Credit risk self-resolves as OCF accelerates
NVDA is significantly over-earning	NVDA at 10-year forward PE low; every quantitative metric accelerating on the ground	Contrarian buy during a narrative panic
SpaceX is just a launch company	xAI/Cursor = $10B ARR; $50B/GW compute; only non-hyperscaler bringing >500 MW/year	Consensus severely underestimates compute revenue
Open source takeshare = less compute needed	Same flops required; market share shift just moves margin from model to infra layer	Open source is structurally bullish for NVDA, CRWV

Investment Thesis #1: The Repricing Flywheel — Contracted Compute Is Severely Mispriced

Argument In 2024–25, every hyperscaler and neocloud locked in long-term supply agreements at prices anchored to a world where GPU costs were expected to fall. Instead, spot prices went vertical. The contracted base of installed compute is now generating revenue at roughly Ampere-era pricing (two generations behind current Blackwell/Rubin hardware), while spot rates are dramatically higher. A company Baker spoke with was paying approximately $2/GPU-hour for B200s and is now bidding just under $4 for the same cluster seven months later. As contracts roll off, hyperscaler operating cash flows will jump sharply.

"Everyone in '24 and '25 thought that GPU prices, if you were really bullish, you thought they would decline slowly. I think anyone in '24 or '25 thought that the prices of old GPUs would still be going vertical in 2026. And so essentially you have the contracted base of installed compute trading at a massive discount to the current spot market."

The contrarian element Consensus models Blackwell and Rubin monetizing at Ampere rates — two full generations below current spot. Baker estimates that closing this wedge pushes the $1.3–1.4T in hyperscale OCF toward ~$2T, eliminating roughly $700B of the credit demand driving the bear case. Microsoft, Meta, and Amazon already reported OCF accelerating 28–32% in the most recent quarter. Microsoft brought a major slug of capacity online in June that didn't appear in Q2 results at all.

Trigger Quarterly OCF acceleration prints from MSFT, META, AMZN, GOOG confirming the repricing in action.

Names NVIDIA (NVDA), Microsoft (MSFT), Meta (META), Amazon (AMZN), Google (GOOG), CoreWeave (CRWV).

Investment Thesis #2: NVIDIA at 10-Year Forward PE Trough

Argument As Baker records, NVIDIA is trading at its lowest forward PE in 10 years. The only prior times semiconductors were cheaper were Liberation Day and DeepSeek — both V-shaped bottoms. The market's implicit assumption is significant over-earning. Baker's on-the-ground pressure test found the opposite: GPU availability, GPU retail pricing, spot DRAM prices, and token growth are all accelerating. He could not identify a single negative quantitative metric despite spending weeks in Silicon Valley specifically looking for one.

"The underlying fundamentals are improving. And NVIDIA's actually, as we record this, at its lowest forward PE of the last 10 years. The market 100% thinks they're significantly overvalued."

The contrarian element NVDA is executing a novel credit wrapper strategy — providing equity financing and revenue-share arrangements for compute purchasers. Baker views every equity stake Nvidia has not taken as a mistake in hindsight. Jensen, who sees inside every frontier lab and all the continual-learning startups, is acting as a structural bull with his own capital. The wrapper converts lumpy CapEx demand into a recurring revenue relationship that the market is not pricing.

Trigger NVDA equity stake or revenue-share disclosures; hyperscaler OCF acceleration confirms the repricing thesis.

Names NVIDIA (NVDA).

Investment Thesis #3: SpaceX — The Compute Business Is Not in Consensus

Argument Baker says the market does not understand SpaceX as a company. Since the June 2026 IPO, fundamentals have only improved: Grok 4.5 launched, Cursor was acquired, and SpaceX has again demonstrated it can bring more compute to market faster and cheaper than anyone. It is one of only five entities — alongside the four major hyperscalers, CoreWeave, and Crusoe — to have brought more than 500 megawatts of power online in a single year, and SpaceX did it fastest at the lowest cost. Compute rates have gone up since SpaceX signed its last contracts, not down.

"Grok 4.5 and Cursor — the sum of that probably hits a $10 billion ARR pretty quickly. Forget Starlink V3, forget direct-to-cell. If they bring out anywhere near that compute, consensus is 73 billion. And that's 8 gigs at $50 billion a gig."

The contrarian element A major New York hedge fund short case rests on spot compute prices falling 90%. Baker's counter: SpaceX brought a vast amount of compute to market overnight and the market absorbed it entirely without a blip. His view is that very little of the potential compute build is in estimates or investor thinking. Orbital compute (Star Cloud, Benchmark-funded, using Starlink laser technology) is a further call option the market is ignoring.

Trigger SpaceX compute utilization/revenue disclosure; Grok 4.5 + Cursor ARR vs. $10B target; capacity additions toward 500 MW threshold.

Names SpaceX (IPO'd June 2026), CoreWeave (CRWV).

Investment Thesis #4: Token Is a Token — Open Source Is Bullish for Infrastructure

Argument The July sell-off was partly triggered by open-source models (GLM 5.2, Kimi K3) taking significant share from frontier labs. Baker argues the market misread this entirely. Producing a token — regardless of model — requires the same flops, watts, cooling, and memory. Open source taking share only shifts gross margin from the model layer (Anthropic, OpenAI, ~90% margins) to the infrastructure layer. It does not destroy compute demand; through lower token prices and resulting demand elasticity, it may increase total compute demand.

"All open source taking share does is take margin dollars out of the frontier model layer and effectively drive token demand. You need more demand for compute. You're literally just shifting tokens from really expensive tokens with 90% gross margins to tokens with maybe 30% gross margin. And that's where the savings are coming from — but the tokens cost the same amount of compute to produce."

The contrarian element Jensen Huang is the world's largest champion of open source. Baker asks whether it makes sense that Jensen would make this a signature issue if it were bad for Nvidia. Inference clouds (Fireworks, Together, Modal, Base10) are growing as fast as frontier labs did in their early days and burning very little cash — Rule of 40 numbers Baker calls "crazy."

Trigger Inference cloud revenue disclosures; NVDA/CRWV utilization remaining high despite open-source shift; token volume data from Fireworks or Together.

Names NVIDIA (NVDA), CoreWeave (CRWV), Microsoft (MSFT) (Azure inference), GE Vernova (GEV) (power demand unbothered by model mix shift).

Investment Thesis #5: SRAM Accelerators — The Compute Multiplier Nobody Is Discussing

Argument Baker identifies an architectural shift that is absent from most investor conversations. SRAM-based accelerators — often made on older nodes that don't compete with leading-edge GPUs — can dramatically improve inference ROI by disaggregating the decode phase of inference. Decode has two components: attention (requires HBM DRAM) and feed-forward network (where SRAM is unbeatable). Disaggregating these three workloads — prefill, attention, feed-forward — across specialized chips improves ROI of the existing install base without requiring new leading-edge silicon.

"The ROI on adding these SRAM accelerators to the existing install base of compute — what we're seeing is you do better. You just can't beat SRAM in particular for that feed-forward network. And you just can't, no matter how much you try to get the ratio of compute to HBM DRAM to SRAM on a single chip correct. Like the workloads are always changing."

Trigger SRAM accelerator company product announcements; inference providers reporting cost-per-token reductions from disaggregated architectures.

Names Beneficiaries include existing compute operators (NVIDIA (NVDA), CoreWeave (CRWV)) whose install bases get higher ROI from add-on SRAM disaggregation.

The Ecosystem Map

Direct positions Baker holds or describes with high conviction

NVIDIA — 10-year forward PE trough; credit wrapper misunderstood; Jensen's equity deployment is structural
SpaceX — Compute business not in consensus; Grok 4.5 + Cursor $10B ARR; orbital compute option
Microsoft, Meta, Amazon, Google — Direct beneficiaries of contracted compute repricing; OCF flywheel

Named companies Baker calls important

Anthropic — Almost certainly FCF positive and growing strongly; private, but heavily referenced
OpenAI — Massively accelerated in July; growing faster than Anthropic third-party data suggests
CoreWeave (CRWV), Crusoe — Only non-hyperscaler compute providers at >500 MW scale
Fireworks AI, Together AI, Modal, Base10 — Inference clouds with frontier-lab growth rates, minimal cash burn
Star Cloud (Benchmark-funded) — Orbital compute company partnering with SpaceX Starlink laser tech; Benchmark's independent bet is Baker's sanity check on orbital compute

Baker's view on value chain Infrastructure captures incrementally more value as open source grows; frontier model margin is diluted but total token volume expands. Memory (HBM for attention, SRAM for feed-forward) is the single most important lever for increasing token output per unit of compute.

Key Risks

Regulation — #1 risk per Baker. NY data center moratorium is "the first of many." Political narrative (data centers raise power prices, waste water, take jobs) is factually wrong but spreading. Industry PR is "terrible."
Operating cash flows don't accelerate — If contracted compute repricing thesis is wrong and OCF stays flat, credit risk becomes real and the CapEx cycle requires debt
Credit contagion — Real yields up, spreads widening, Meta CDS blowing out, Meta bond priced poorly; debt-funded overbuild would replay the Netscape-era unwind
Sustained GPU price decline — Any prolonged visible decline in Blackwell/Rubin spot prices would be an instant negative signal
Continual/sample-efficient learning breakthrough — If models trained on 10T tokens (vs. 300T today) can learn efficiently in the world, training demand could fall structurally; Baker calls this the most interesting technical question from his Silicon Valley trip
China DUV capability — China reportedly has a DUV machine; a phase transition from nothing to a 25-year-old tool, not immediately threatening but shouldn't be dismissed; market probably overreacted in July

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
Tier 1	NVIDIA (NVDA)	10-year forward PE low; market pricing over-earning that ground data contradicts; credit wrapper adds undiscounted accretion	Hyperscaler OCF acceleration; NVDA equity stake disclosures
Tier 1	Microsoft (MSFT)	OCF accelerating 28–32%; June capacity slug not in Q2; direct repricing flywheel beneficiary	Q3 OCF print
Tier 1	Meta (META)	OCF accelerating 28–32%; open-source Llama strategy means every token still runs on the same compute	Q3 OCF print; token volume data
Tier 1	Amazon (AMZN)	OCF accelerating 28–32%; AWS + Anthropic stake (FCF positive, growing strongly)	Q3 OCF print; Anthropic IPO catalyst
Tier 2	SpaceX (IPO'd June 2026)	Compute business not in consensus; Grok 4.5 + Cursor $10B ARR; only non-hyperscaler at >500 MW pace	Compute revenue disclosure; xAI ARR milestones
Tier 2	CoreWeave (CRWV)	One of handful of entities bringing >500 MW/year; direct beneficiary of spot repricing above contracted rates	Capacity additions; OCF acceleration
Tier 2	Google (GOOG)	OCF flywheel beneficiary; Anthropic investor; TPU supply chain leverage	Q3 OCF print; Anthropic FCF disclosure
Tier 3	GE Vernova (GEV)	Turbine manufacturers are a supply bottleneck for watts; capitalism reconditioning jet turbines for data center power; LTA signings rising	Data center LTA disclosures; turbine order book expansion

Monitoring Checklist

MSFT, META, AMZN, GOOG Q3 OCF prints — The single most important data point for validating the repricing flywheel; Baker's core thesis stands or falls here
B200 / Blackwell spot pricing — Baker cites ~$4/GPU-hour bid vs. prior ~$2; sustained increase validates repricing; any sharp sustained decline invalidates thesis
NVDA forward PE trajectory — Currently at 10-year low; a re-rating higher signals market accepting the repricing thesis
SpaceX compute revenue / utilization disclosure — Grok 4.5 + Cursor ARR vs. $10B target; gigawatt capacity vs. 500 MW threshold
Inference cloud growth disclosures (Fireworks, Together, Modal, Base10) — Confirms token-is-a-token thesis; watch for Rule of 40 numbers
New York data center moratorium expansion to other states — Baker's #1 risk; any replication in red or swing states is a structural negative
Hyperscaler debt issuance surge — OCF-funded buildout is the bull case; any pivot to large-scale debt financing shifts risk profile dramatically
China DUV commercial yield and volume — Not an immediate threat; watch for production ramp and orders over 12–18 months

Bottom Line

NVIDIA at a 10-year forward PE low is the most asymmetric entry in the cycle. Baker pressure-tested every bear thesis in Silicon Valley for weeks. The only negative he found — third-party data suggesting an Anthropic plateau — was hotly contested and offset by OpenAI and open-source accelerating sharply.
The repricing flywheel is the thesis. Contracted compute was locked in at Ampere prices. As it rolls off at Blackwell spot rates, hyperscaler OCF will jump by hundreds of billions — enough to fund the entire CapEx buildout and destroy the credit risk bear case simultaneously.
Open source taking compute market share is not a bear thesis for infrastructure. It is a bull thesis. Every token — frontier or open source — demands the same compute. Open source only compresses model-layer margins; infrastructure demand is unaffected or grows via demand elasticity.
SpaceX is being priced as a launch company. The xAI/Cursor $10B ARR and the compute build are not in consensus. Baker thinks very little of the potential compute gigawatt upside is in the stock price.
Watch for continual/sample-efficient learning. It is the one legitimate technical risk to training demand. Multiple well-funded labs are reportedly close. If solved, the implications for the shape of compute demand (not the level) are significant.