AI infrastructure stocks endured a steep July selloff even as GPU rental prices, cloud cash-flow margins and long-term memory commitments pointed to continued pressure on available computing capacity, according to Gavin Baker, founder and chief investment officer of Atreides Management.
Speaking with Patrick O’Shaughnessy on Aug. 4, Baker described the market move as “2022 compressed into one month,” saying many AI-related stocks had fallen 40% to 60% from their highs. His central argument was that equity valuations fell far faster than underlying operating signals deteriorated, creating a gap between public-market sentiment and the cost of securing advanced computing hardware.
Baker said market checks showed spot GPU pricing had risen 50% to 60% over the preceding six months, rather than declining as would be expected if supply were rapidly catching up with demand. He also said operating cash-flow margins at hyperscale cloud companies had increased from roughly 28% to 35%, after adjusting for one-time items.
Those figures do not settle the question of whether AI infrastructure spending will ultimately generate sufficient returns for every company involved. They do challenge the simple reading of July’s selloff: that the market had suddenly found a broad surplus of high-end GPUs or a sharp collapse in enterprise demand.
Gpu renewals point to tighter capacity economics
Baker highlighted contract renewals as a more direct measure of the market than daily stock-price moves. In one case, he said, a company had rented a cluster containing several thousand NVIDIA Blackwell GPUs at a price around the mid-$2 range per GPU hour. Seven months later, the same buyer sought to renew B200-class capacity for less than $4 per GPU hour.
He also cited an inference-cloud operator that had publicly indicated it expected to pay twice as much for Blackwell capacity when a contract came up for renewal. Inference refers to the computing work performed when an AI model generates an answer, image, video or other output for a user, rather than the initial process of training the model.
The distinction matters for infrastructure demand because training can be episodic, concentrated around major model releases, and subject to efficiency improvements. Inference is tied more directly to ongoing product use. Baker said inference, rather than training, is likely to account for the larger share of sustained compute demand.
That view comes with a meaningful technical risk. Baker said advances in continuous learning or sample-efficient learning could reduce the number of tokens required to train frontier models. He offered a hypothetical example in which training requirements fell from 300 trillion tokens to 10 trillion before models learned more efficiently from real-world deployment. Such a shift could temporarily reduce training-related hardware demand even if inference use continued to expand.
Meta’s compute plans became a flashpoint
Baker linked part of the selloff to disclosures that Meta planned to rent out computing capacity. Markets interpreted that prospect partly as evidence that large platforms had overbuilt infrastructure, he said. Baker took the opposite view, arguing that renting installed capacity at higher prices would represent monetization of scarce assets rather than an admission of excess supply.
He added that the telemetry he followed around Meta’s capital spending did not indicate a retrenchment. The debate reflects a wider uncertainty around the economics of hyperscale AI: companies are spending heavily before the long-term revenue mix from cloud rentals, enterprise AI services and consumer products is fully established.
Credit markets added pressure to that uncertainty. Baker said real yields rose during the period and credit conditions tightened, while credit default swap spreads widened across companies. He pointed to Meta pricing a bond sale on weaker-than-expected terms as one example of more demanding financing conditions.
Higher borrowing costs can change the market’s view of capital-intensive infrastructure projects quickly. The largest cloud operators may have substantial operating cash flow, but their AI buildouts require spending on chips, networking, data centers, electricity and memory years before every asset reaches mature utilization.
The financing debate depends on GPU pricing
Baker argued that consensus models may be undervaluing new GPU fleets by applying older Ampere-era monetization rates to newer Blackwell and future Rubin systems. Under his framing, hyperscaler operating cash flow is estimated at $1.3 trillion to $1.4 trillion when capacity is valued at legacy run-rates.
Using a higher monetization assumption—below then-current Blackwell pricing but above Ampere pricing—Baker said operating cash flow could approach $2 trillion. That would reduce the debt required to finance AI infrastructure by about $700 billion in his sensitivity case.
The calculation depends heavily on the durability of premium GPU pricing. If customers accept higher rates at renewal, cloud providers would have more room to fund expansion internally. If rental prices retreat as new capacity comes online, the financing burden would remain closer to the lower case.
Memory supply is another constraint in that equation. Baker said large hardware buyers were increasingly pursuing long-term agreements, known as LTAs, that include prepayments and pricing boundaries. He named Amazon’s Trainium program, Google’s TPU efforts, AMD and NVIDIA as the four scaled buyers shaping that market.
Such arrangements can secure allocation during shortages, but they also carry risk when negotiating leverage changes. Buyers that sign aggressive supply agreements could face less favorable terms if memory manufacturers regain pricing power.
Crypto-linked infrastructure faces a separate valuation test
The July equity decline may influence sentiment across digital assets associated with decentralized computing, data-center capacity or energy generation. It does not establish a direct valuation case for tokens tied to those networks.
GPU rental rates and hyperscaler capital spending measure demand for centralized, high-performance infrastructure. A decentralized computing network must separately demonstrate that it can attract workloads, deliver reliable hardware access, maintain competitive pricing, and generate revenue that reaches token holders or network participants under its stated rules.
Open-source AI models could increase the need for compute by making powerful systems accessible to more developers, as Baker argued. Yet lower barriers to model access do not automatically direct demand toward token-based infrastructure. Many customers will choose providers based on latency, security, compliance, uptime and existing cloud relationships, not solely on the availability of raw compute.
Baker also cited policy and supply-chain risks that could reshape the sector. He identified data-center restrictions, including a New York ban, as a major regulatory concern. He discussed reports that China had gained access to DUV lithography tools, which could affect expectations for ASML’s future order book over a longer horizon, even if the manufacturing implications take years to emerge.
For crypto traders assessing infrastructure-linked tokens, July’s dislocation offers a reason to examine operating evidence more closely rather than treating falling AI-equity prices as either a definitive warning or a blanket buying signal. Contract pricing, utilization, power availability, financing costs and verified network revenue will provide a clearer test of whether compute demand reaches decentralized systems in a durable economic form.
For deeper context on AI-driven market moves, explore Toobit’s academy piece today AI moves shake markets now.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
