No Evidence of Bearish Fundamentals in AI Stocks! Was the July Crash a Misfire? Prominent Investor: 'Underlying Business Is Fine'

Following a sharp decline in artificial intelligence (AI) stocks in July, markets began questioning whether the AI investment boom was cooling down. However, renowned tech investor Gavin Baker, after touring Silicon Valley, believes there is no fundamental evidence to justify the bearish sentiment—except for tightening credit conditions and policy risks.

Baker notes the sell-off began when Meta (META-US) announced it would rent out some of its idle computing power. The market interpreted this as a warning sign that the company might be scaling back capital expenditures, causing its stock price to drop.

However, he considers this a misunderstanding. Meta’s capital expenditure plans have not shrunk—in fact, they’ve become more aggressive. The subsequent release of Meta’s high-performing Llama model further confirms the company hasn’t slowed down.

Then came the emergence of open-source models like Kimi K3, along with structural changes in the Silicon Data Token Index, sparking market fears that open-source players could capture market share and weaken overall demand for AI infrastructure.

Baker counters this argument: regardless of whether models are open- or closed-source, the computational resources, memory, and power required to generate each token are identical. Open-source models gaining market share merely shifts profits from cutting-edge model developers to the infrastructure layer, without materially impacting hardware demand.

Later rumors about DUV lithography equipment briefly hammered semiconductor equipment stocks, but Baker believes the market overreacted—though not entirely without merit.

The only real concern: Credit market cooling

Among all triggers, Baker admits the only variable he takes seriously is the shift in credit markets.

He points out that rising real interest rates and widening credit spreads are undeniable. Recently, Meta’s bond pricing fell short of expectations, and CDS spreads for major tech giants have generally widened.

The key question is how much of the massive future funding needs for AI infrastructure must rely on debt.

Baker analyzes that Microsoft (MSFT-US), Meta, and Amazon (AMZN-US) have accelerated their operating cash flow growth from 28% last quarter to 32%. Excluding one-time items like EU fines, the underlying growth rate reaches 35%—remarkable given their massive revenue base.

He further estimates that if hyperscale cloud providers reflect their compute monetization ability using current Blackwell chip pricing—rather than older, cheaper generations—overall operating cash flow could jump from the currently expected $1.3–1.4 trillion to approximately $2 trillion. This would eliminate around $700 billion in potential credit demand from the market.

In other words, if pricing holds, future build-out expenses may not require significant borrowing and could be funded internally.

Baker cites what he believes is the clearest indicator of market reality:

A startup rented a cluster of Blackwell GPUs seven months ago at about $2 per GPU per hour. Now, upon renewal, the quoted price is close to $4—an increase of 50% to 60%, completely contrary to earlier expectations of gradual price declines.

He also mentions that an inference cloud provider has publicly stated it will pay double the original Blackwell rental fee upon contract renewal, suggesting major cloud providers may have underestimated their latent profitability.

Baker emphasizes his trip aimed to uncover negative data, yet aside from debates over whether Anthropic’s growth is slowing, he found almost no credible bearish indicators.

The memory market game: Breaking long-term contracts is self-sabotage

Baker devotes significant attention to the role of Long-Term Agreements (LTAs) in the memory market.

He highlights that High Bandwidth Memory (HBM) is currently the biggest bottleneck in AI computation. The more memory paired with each unit of compute, the more tokens can be generated—making it the most critical variable for efficiency.

Under this context, he applies game theory to analyze the high cost of breaching LTAs. Suppose a cloud provider tries to renegotiate or unilaterally cancel its contract in 2027 or 2028, citing oversupply. Memory suppliers could simply redirect the allocated quota to competitors, instantly stripping the defaulter of market share.

Given the industry’s cyclical nature—oversupply often followed by undersupply—the defaulter may struggle to regain priority allocation in the next upswing.

Baker bluntly states that in today’s environment, breaking LTAs could directly destroy a company’s market position—a consequence far more severe than in the past.

NVIDIA’s new business model is being underestimated

On NVIDIA, Baker believes its current valuation is clearly too low.

He describes a new “financing facilitation plus revenue sharing” model: NVIDIA assists buyers in financing GPU purchases and participates in revenue sharing when resale prices exceed a certain threshold. This could allow NVIDIA to rapidly build a massive cloud business primarily driven by royalty-based income.

Baker argues the market fundamentally misunderstands this model and urges NVIDIA to improve external communication. His rationale: there are almost no assets in the market easier to finance than NVIDIA GPUs.

Combined with NVIDIA’s capabilities in land and power integration, and its equity stakes in numerous AI labs, the company is systematically strengthening its competitive moat using vast free cash flow.

He notes NVIDIA’s forward P/E ratio is at a near two-year low, indicating the market broadly assumes its profitability is overstated—yet he found no data in Silicon Valley to support such pessimism.

SpaceX’s compute ambitions are similarly underestimated

Citing Substack author Will Funda AI, Baker notes SpaceX plans to build 8 gigawatts of compute capacity within 18 months. While he doesn’t claim Elon Musk will definitely succeed, he acknowledges it would be an astonishing engineering feat.

Based on SpaceX’s current monetization efficiency of about $50 billion per gigawatt, and the consensus forecast of $73 billion in revenue next year, the annualized recurring revenue from core businesses could soon surpass $10 billion—potential not yet fully reflected in valuations.

He also mentions venture firm Benchmark has invested in orbital computing startup Star Cloud, which is collaborating with SpaceX on laser technology applications for Starlink—suggesting this path, while bold, isn’t baseless.

The biggest risk isn’t technical—it’s policy

Finally, Baker stresses that what truly unsettles him isn’t technology or finances, but policy risk—specifically New York City halting data center development.

He criticizes the AI industry’s poor public communication, leading the general public to widely believe data centers raise electricity prices, deplete water resources, and take away jobs.

He refutes each point: most data center site agreements actually lower electricity prices for nearby residents; a widely cited academic paper estimating data center water usage was found to contain errors up to four orders of magnitude, with the author repeatedly admitting fault; and blue-collar job creation from data centers is long-term, not temporary.