Compute shortage is redefining Silicon Valley’s competitive landscape, making compute capacity the crucial variable for AI firms.

Firms across the valley are in a fierce race to secure GPUs and data-center resources to fuel model development and profitability.

Anthropic and SpaceX agreed to deliver over 300 megawatts of compute capacity, roughly 220,000 Nvidia GPUs, with potential contract value in the tens of billions over the coming years.

Executives from Anthropic, SpaceX, Meta and others are actively trading access to compute, with outages common due to crunches and high-stakes negotiations.

Long build-out timelines for data centers, regulatory hurdles, and uncertainty about future capacity complicate planning for startups and incumbents alike.

For startups, access is tightening as cloud providers push multi-year contracts and upfront payments, prompting some to pursue a global computing power gold rush to source underutilized servers.

H100 GPU rental prices have surged on multi-year deals, as cloud providers require upfront payments and guarantors, squeezing smaller players.

Anthropic’s Dario Amodei sought more compute from Meta earlier, was declined, and later secured a deal with SpaceX; Google faced internal allocation tensions after backing Noam Shazeer’s return, who then moved to OpenAI due to compute limits.

Google reportedly paid Noam Shazeer $2.7 billion to return, yet he still faced allocation issues contributing to a move to OpenAI.

SpaceX’s February acquisition of xAI consolidated AI assets and compute capacity, influencing negotiations with Anthropic and others.

SF Compute and other brokers highlight a global, labor-intensive market for securing data-center capacity, with capacity sourced from diverse sites.

The trend favors incumbents with self-built infra, while startups struggle to secure timely compute, accelerating consolidation among large compute holders.