Compute Becomes Hard Currency in Silicon Valley's AI Race

QbitAI cited an August earnings disclosure showing quarterly AI business revenue of $2.561 billion, up 247% from a year earlier and more than tripled from the previous quarter. The report described SpaceX as a middleman for Nvidia GPUs, selling compute to Anthropic while the AI company sought more capacity for its models.

The scramble is also visible at Google. QbitAI cited foreign media reports that Noam Shazeer, a Transformer co-author, returned to Google in 2024 in a deal valued at $2.7 billion and became a co-leader of Gemini. His team nonetheless faced reallocation of compute resources, and he moved to OpenAI. Demis Hassabis had long complained about insufficient compute for his projects, a factor that influenced his decision to step down as DeepMind CEO, according to the report. Google's internal resource allocation has been contentious, with Sergey Brin sometimes intervening to redirect resources to projects he considers important. The article said the underlying problem is that compute remains insufficient.

Anthropic's compute needs have grown so large that it has turned to outside suppliers. In May, an agreement with SpaceX was to deliver more than 300 megawatts of new capacity that month, involving 220,000 Nvidia GPUs, and raise usage limits for Claude Code and API, according to the report. The goal was to use more backend compute to serve front-end users: more compute, stronger models and better user retention. Competition is now running on two tracks: product-level competition for users and developers, and infrastructure-level cooperation with whoever can supply resources in time.

Smaller AI startups face a sharper squeeze. They cannot build their own data centers and have relied on flexible, on-demand GPU rentals from cloud platforms such as AWS. As leading AI labs lock up capacity, some cloud providers are demanding multi-year contracts, prepayments of up to 30% of contract value and even payment guarantors. Startups must decide whether to buy too much compute and risk wasting cash if user growth disappoints, or buy too little and miss growth opportunities. That has created a business of global compute prospecting: finding underused AI server resources and renting them out. San Francisco Compute, a representative company, has collected servers from old chicken coops in rural America, and Musk has discussed further compute cooperation with TSMC, according to QbitAI. Top Silicon Valley venture capital firms are also stockpiling GPUs. Y Combinator launched a dedicated GPU cluster for its startups in July, and Radical Ventures helps portfolio companies coordinate supplier contracts.

The shortage has complex causes. OpenAI, Meta and Google are all building new data centers and promising vast GPU resources in the future, but model demand cannot wait for that capacity to come online. The way users interact with AI is adding pressure. Since this year, agents have begun replacing single-turn question-and-answer interactions as a core usage pattern, forcing servers to handle more parallel computing for long-running tasks. Rental prices have moved accordingly. SemiAnalysis data published in April showed the one-year H100 rental contract price rose from a low of $1.70 per GPU per hour in October 2025 to $2.35 in March 2026, an increase of nearly 40%.

The result is a compute hierarchy. According to QbitAI, the narrative of startups challenging industry giants will fade because even a good roadmap is useless without enough computing resources. Compute is becoming an entry threshold for the AI era and reinforcing the advantages of large companies that own infrastructure. AI product companies face more variables: they must worry about whether anyone will use their products and whether user growth will outpace their compute expansion. The middle layer selling shovels is benefiting, the report said.

Editor's Summary Compute access is now central to Silicon Valley's AI competition, with Anthropic turning to Meta and SpaceX for capacity and Google struggling to retain star researchers. The shortage is raising costs for startups, changing rental prices and pushing venture capital firms to secure GPUs. The dynamic favors large companies with infrastructure while creating opportunities for intermediaries that supply compute.