This piece says AI stocks crashed 40-60% in July 2026, but Gavin Baker thinks the market is too pessimistic because real demand (GPU prices, token usage) is accelerating. He likes NVIDIA (cheapest forward P/E in 10 years, market underestimates it), Meta (didn't cut spending, released a new model), and SpaceX (potential huge compute business not priced in). His biggest worry is regulation and public backlash against data centers.
Gavin Baker discussed on the Invest Like the Best podcast the disconnect between recent turbulence in the AI market and actual demand in Silicon Valley. Despite a sharp sell-off in the AI sector in July (described as "2022 compressed into one month"), Baker noted that spot prices for older GPUs actu
Gavin Baker is the Founding Partner and CIO of Atreides Management. This episode discusses the disconnect between the severe sell-off in the AI sector in July 2026 ("2022 compressed into one month") and the accelerating real-world demand in Silicon Valley. Baker's core thesis is that despite AI stocks plunging 40-60%, all quantifiable demand indicators (GPU pricing, DRAM spot prices, token growth) are accelerating, revealing a massive gap between market pricing and fundamentals.
Baker attributes the July crash to a series of events interpreted negatively by the market, but argues that most of these narratives do not hold up to scrutiny:
1. Meta Leasing Compute Capacity: The market interpreted this as Meta having idle capacity and planning to cut capital expenditure. Baker points out that Meta did not actually cut capex; instead, after seeing SpaceX sell compute clusters at a "massive premium," Meta wanted to first prove the IRR on a small scale before financing expansion. Meta subsequently released Llama 1.1 — "the best model in two years."
2. Panic over Open-Source Kimi and GLM 5.2: The leap in open-source model capabilities caused a "mix shift" in the Silicon Data token index — from high-margin frontier tokens to open-source tokens. The market viewed this as a negative signal, but Baker explains: "A token is a token. The flops, memory, and watts required to produce one token are exactly the same. Open source simply shifts the profit from the frontier model layer to the infrastructure layer."
3. China's DUV Lithography Breakthrough: The market heavily sold off semiconductor equipment stocks. Baker believes this was "likely an overreaction," but acknowledges it was indeed a "phase change" — China went from having nothing to having something, albeit with a technology lag of about 25 years.
4. Credit Market Deterioration: Rising real interest rates, widening CDS spreads, and a surge in NVIDIA's CDS. This is the only signal Baker considers "undeniable" as negative.
| Market Concern | Baker's Response |
|---|---|
| Massive debt financing is needed; deteriorating credit conditions will cause the capital cycle to collapse | Construction is primarily funded by operating cash flow, not credit |
| Hyperscaler operating cash flow is insufficient to support it | Assuming compute is priced at the Ampere level (two generations behind), it amounts to $1.3-1.4 trillion; if priced at a discount to current Blackwell, it reaches $2 trillion |
| A closed credit market will halt construction | If compute is repriced, credit metrics improve, making financing easier; even if credit is needed, "the compute shortage itself is the solution" |
> Key Data: NVIDIA's forward P/E ratio is at a 10-year low, having only been cheaper during the DeepSeek and Liberation Day events. Baker's judgment: "The market is 100% convinced that NVIDIA is massively over-earning."
Baker notes that the market generally expected GPU prices to decline modestly, but the reality is the exact opposite:
This means all hyperscalers are underearning – their contract prices are far below the current spot market.
Baker compares memory supply agreements (LTAs) to "Game of Thrones":
> "If you default on an LTA in 2027, and then two or three years later the leverage shifts back to memory manufacturers – you're out. Game over."
There are only four key players: Amazon (Trainium), Google (TPU), AMD, and NVIDIA (far larger than the other three combined). Baker believes that for the foreseeable future, market share will be determined by supply chain allocation, not price competition. Defaulting on an LTA means losing future allocation rights – "you could destroy your entire business and franchise."
Baker describes an NVIDIA strategy that the market has misunderstood:
> "They've rolled out a very clever business model – I call it credit guarantees plus revenue sharing. If GPU prices have a floor, this allows them to quickly build a massive cloud business through royalties."
NVIDIA does not directly provide loans but helps buyers finance through equity investments and revenue-sharing agreements. Baker points out that NVIDIA has invested in nearly every AI company (except memory manufacturers), because Jensen can see the progress in all the labs – "what he sees makes him bullish."
Baker provides a detailed breakdown of how the router mechanism is misunderstood by the market:
> "After a company sets up a router, AI spending may decline—but this has nothing to do with GPU compute hours. You are simply shifting tokens from a frontier model with a 90% gross margin to an open-source model with a 30% gross margin, yet the compute required to produce those tokens is exactly the same."
The actual effect is: cheaper tokens stimulate more usage. Baker cites data:
Baker compares open source to dark matter in the universe—difficult to measure in public markets, but demand is clearly accelerating. The capability leaps brought by GLM 5.2 and Kimi K3 are driving more inference demand. Inference cloud companies such as Fireworks, Together, Modal, and Base10 are "growing almost as fast as early frontier labs, but burning very little cash."
Baker acknowledges that if continuous learning and sample-efficient learning are solved, it could create a "temporary discontinuity" in training demand—models shifting from training on 300 trillion tokens to just 10 trillion. However, he emphasizes: "This is good for the world, and it is hard to believe this would negatively impact AI infrastructure demand."
Baker argues that the AI industry has done a "terrible job" with public relations:
> "New York has already imposed a moratorium on data centers. We live in a post-fact, post-logic political world. The narrative for the average American is: data centers will raise electricity prices, drain water resources, and steal jobs."
The reality is quite the opposite: Baker points out that due to "behind-the-meter deals," local electricity prices typically fall after a data center is built; developers now promise to build hospitals, schools, and police stations; and data centers represent "the best thing for blue-collar wages in my lifetime."
Baker calls on the industry to proactively communicate three core facts:
1. Data centers lower local electricity prices
2. They create sustained, high-paying blue-collar jobs
3. AI is saving lives and curing rare diseases
> "If you have a sick child, a sick parent, a sick loved one — AI significantly increases their chances of recovery. Everyone needs to tell these stories."
Baker points out that the market does not seem to fully appreciate SpaceX's AI computing business since its IPO:
Baker cautiously remarks: "I wouldn't bet against Elon, but 8 gigawatts in 18 months would be a truly incredible feat."
Benchmark has invested in Starcloud, an orbital computing company that uses SpaceX's Starlink laser technology. Baker considers this an important "sanity check" — "The people at Benchmark are quite smart, and they chose to invest in an orbital computing company that has no internal launch costs."
| Position | Analyst Stance | Key Data |
|---|---|---|
| NVIDIA | Bullish (forward PE at 10-year low; market sees it as over-earning but Baker believes it is not fully reflected) | Forward PE at 10-year low; B200 spot price up 50-60% in 7 months |
| Meta | Bullish (no CapEx cuts; Llama 1.1 is the best model in two years) | No CapEx cuts; Llama 1.1 released |
| Microsoft | Bullish (operating cash flow accelerating) | Large computing capacity launched in June not reflected in Q2 earnings |
| Amazon | Bullish (operating cash flow accelerating) | Operating cash flow accelerated from 28% to 32% (35% adjusted) |
| Anthropic | Neutral to cautious (third-party data suggests the curve may deviate from trajectory, but shareholder disputes exist) | Not specified |
| OpenAI | Bullish (accelerating growth, almost certain to generate substantial free cash flow) | Not specified |
| SpaceX | Bullish (an undervalued computing giant by the market) | Consensus 2027 revenue at $73 billion; possibly 8 gigawatts online in 18 months |
| Fireworks | Bullish (inference cloud, efficient growth) | Launched Nexus product; three lines of code to customize models |
| Together / Modal / Base10 | Bullish (inference cloud, minimal cash burn but rapid growth) | Not specified |
| Hynix | Neutral (recommends a credit guarantee strategy similar to NVIDIA's) | Not specified |
| Cognition | Bullish (dark horse) | Cognition Index shows companies with high AI spending grow faster |
| Starcloud | Bullish (orbital computing, Benchmark investment) | Uses Starlink laser technology |
1. "A token is a token" — Gavin Baker argues that the market misreads the impact of open-source models on compute demand. Open-source tokens have lower gross margins but identical compute costs; the actual effect is to stimulate more demand, shifting profits from frontier models to the infrastructure layer.
2. "If you default on an LTA in 2027, and then two or three years later leverage swings back to memory vendors—you're out. Game over." — Baker uses game theory to explain why memory supply agreements (LTAs) are nearly unbreakable in the current environment, which reinforces NVIDIA's moat.