Rental prices for AI chips climb 60%, deepening the scramble for capacity

Anyone who already owns chips, data centres or electricity now holds the upper hand in pricing, which gives hyperscalers, neoclouds and chipmakers an advantage over startups that must rent computing capacity from them. More expensive rentals and stricter terms suggest compute vendors will see stronger near-term revenue, but they also raise concentration risk, because only the largest and best-funded labs can secure capacity in advance. Using overbooking as a hedge creates the potential for an abrupt reversal if demand falls short, saddling firms with heavy lease commitments and a surplus of capacity being resold. That fragility echoes recent warnings that financing, power and data centre delays, rather than demand, are now the main limits on the AI build-out.

---

Earlier:

• Broadcom’s chief executive compares the AI boom to a casino while aiming for $230 billion in chip sales
• Goldman Sachs identifies power and compute as two bottlenecks of the AI boom

---

AI’s real constraint is not ideas or money but raw computing capacity, and the battle to secure it is turning competitors into collaborators, executives into deal brokers and startups into casualties.

Summary:

• The Wall Street Journal (gated) reports that a shortfall of computing power has set off a hyper-competitive, sector-wide rush in Silicon Valley, complete with internal power struggles and surprising deals between competitors.
• According to the research firm SemiAnalysis, the price of renting Nvidia H100 chips on a one-year contract has climbed 60% over the past year.
• Cloud providers are now requiring multiyear agreements with upfront payments often as high as 30%, sometimes backed by guarantors, which leaves many smaller companies unable to participate.
• Top executives are personally brokering deals, such as Anthropic’s agreement to lease more than 300 megawatts from SpaceX, a pact reported to be worth as much as $45 billion.
• Much announced capacity never actually appears, as projects face delays, cancellations or reductions, and some agreements collapse when data centres slip.
• Venture firms are now holding GPU clusters in reserve for the companies they back, since access to compute has become as critical for startups as access to capital.

The race to secure computing power for AI has become a desperate, industry-wide scramble that is overturning Silicon Valley, pushing bitter rivals into alliances, drawing chief executives into deal talks and leaving a generation of startups struggling to keep up, according to the Wall Street Journal.

Compute — the industry’s term for computing power — has become the most coveted resource in the AI boom, and too little exists. Its scarcity shows up in costs. SemiAnalysis, a research firm, says the hourly rate for renting Nvidia H100 chips on a one-year contract is up 60% from a year earlier. Alphabet expects capital spending of as much as $200 billion this year, with much of it poured into AI infrastructure.

Because of the shortage, the mechanics of dealmaking are changing. People familiar with the matter say Anthropic co-founder Tom Brown visited Elon Musk at xAI in late March to negotiate renting compute, after demand for Anthropic’s Claude Code caused a capacity crunch that contributed to frequent outages. In May, Anthropic said it would lease more than 300 megawatts of computing power from Musk’s SpaceX, a deal later reported to be worth up to $45 billion. The Journal reported that Anthropic chief executive Dario Amodei also talked to Meta about capacity, but Meta decided against supplying chips for now.

The conflict extends inside the largest companies. Google paid $2.7 billion in 2024 to bring AI researcher Noam Shazeer back, but his team’s compute allocation was later reduced in favour of another project, and he then left for OpenAI. Google co-founder Sergey Brin occasionally overrides the formal allocation process, according to people familiar with the matter.

Smaller players are facing tougher terms. Where startups could once rent chips flexibly, they are now offered multiyear contracts that often demand up to 30% upfront and sometimes require guarantors. Some are desperate enough to turn to providers with weaker security or infrastructure. Much of the promised capacity is not yet in place, with projects delayed, cancelled or scaled back. One open-source startup had a capacity deal fall through when its partner’s data centre was delayed.

There are big risks on either side. Companies that buy too much capacity may be crushed by lease commitments, while those that buy too little may put their growth at risk. Some venture capital firms are now reserving GPU clusters themselves or guaranteeing contracts for the startups they back.

Because regulation and political backlash are slowing data centre construction, the shortage is unlikely to ease soon. The risk is that the next stage of AI will be determined less by who has the best ideas and more by who can secure the hardware to run them.

No, that is not so.