AI is becoming infrastructure
Big AI moves this week show the fight is now over power, chips and data‑centre deals rather than model demos — Meta just expanded its CoreWeave pact by $21 billion through 2032 and industry coverage describes OpenAI pivoting to a $122 billion infrastructure strategy. At the same time OpenAI paused its Stargate UK data‑centre project citing industrial electricity costs and regulatory uncertainty, and Amazon says its custom chips already make more than $20 billion a year while hinting the business could scale to $50 billion — all signals that compute, energy and contracts are the new battlegrounds. (thenextweb.com) (markets.financialcontent.com) (thenextweb.com) (thenextweb.com)
Meta just agreed to spend about $21 billion with CoreWeave through December 2032, and the deal is for cloud capacity, not a splashy new chatbot. CoreWeave said the contract will spread dedicated artificial intelligence computing across multiple sites and include early deployments of Nvidia’s Vera Rubin systems. (coreweave.com) That tells you where this market is moving: the scarce thing is no longer the demo on stage, but the warehouse full of chips behind it. CoreWeave describes itself as an artificial intelligence cloud, which means Meta is effectively reserving factory space for future model training and responses years in advance. (coreweave.com 1) (coreweave.com 2) OpenAI made the same turn in a different way when it closed a $122 billion funding round on March 31, 2026, at an $852 billion post-money valuation. OpenAI said the money will fund “the next phase” of building artificial intelligence infrastructure, which is a much more capital-heavy sentence than “we trained a better model.” (openai.com) Then the physical limits showed up immediately. OpenAI paused its Stargate United Kingdom data-centre project on April 9, 2026, and reporting said the company pointed to high industrial electricity prices and regulatory uncertainty. (cnbc.com) (thenextweb.com) The energy gap was blunt: The Next Web reported that industrial power prices in Britain were about four times higher than in the United States. A model can be copied in software, but a data centre still has to sit on real land, pull real power, and survive real local rules. (thenextweb.com) Amazon added the chip angle on April 9, 2026, when Chief Executive Officer Andy Jassy said Amazon’s custom silicon business already has an annual revenue run rate above $20 billion. In the same shareholder letter, Jassy said that if the chip unit sold this year’s output to Amazon Web Services and outside buyers like a stand-alone chip company, the run rate would be about $50 billion. (aboutamazon.com) Those chips are not side projects. Amazon Web Services says Graviton handles general computing, Trainium handles artificial intelligence training, Inferentia handles artificial intelligence inference, and Nitro offloads networking and security tasks so servers can do more useful work. (aws.amazon.com 1) (aws.amazon.com 2) Jassy also said Amazon is not planning roughly $200 billion of capital spending in 2026 “on a hunch,” and GeekWire reported that two large customers asked to buy all available Graviton capacity for 2026. That is what an infrastructure race looks like: customers trying to lock up next year’s supply before the buildings are even finished. (aboutamazon.com) (geekwire.com) Meta is also buying power, not just processors. In January, Meta said its nuclear energy agreements could support up to 6.6 gigawatts of new and existing clean power by 2035 for the grids that support its operations, including its Prometheus supercluster in Ohio. (about.fb.com) So the competition is starting to look less like the smartphone app wars and more like railroads or electric utilities. The winners are lining up long contracts for chips, power, land, and construction, because the next breakthrough model is increasingly the output of whoever secured the biggest industrial base first. (coreweave.com) (openai.com) (aboutamazon.com)
