AI infrastructure is getting expensive
The market shows longer, bigger infrastructure commitments: Meta agreed to spend an additional $21 billion with CoreWeave while data‑centre capex reached roughly $770 billion in 2025, signaling runaway demand and heavy long‑dated vendor exposure. At the same time, Broadcom’s multi‑year deals with Google and Anthropic highlight how non‑Nvidia suppliers are becoming strategic infrastructure partners. (cnbc.com) (pv-magazine.com) (markets.financialcontent.com) (247wallst.com)
Meta just agreed to rent another $21 billion of artificial intelligence computing from CoreWeave through December 2032, on top of an earlier $14.2 billion arrangement, pushing the two companies’ total commitment to about $35 billion. That means one social media company is now locking in years of outside computing capacity the way an airline locks in planes before the travel season starts. (cnbc.com) (coreweave.com) CoreWeave is not a consumer app company. It is a landlord for artificial intelligence chips, renting out clusters of Nvidia processors in data centers to companies that need more computing than they can build fast enough themselves. (coreweave.com) (cnbc.com) The new Meta contract is aimed at inference workloads, which is the part where a trained model answers questions, writes code, or generates images for real users. Training is like teaching the model in a classroom once; inference is like staffing the call center every day after class ends. (coreweave.com) That daily serving step is becoming so large that global data-center capital spending reached $770 billion in 2025, according to Rystad Energy. Rystad said data-center spending had already passed solar photovoltaic investment in 2024 and in 2025 also surpassed upstream oil and gas. (pv-magazine.com) (rystadenergy.com) This is why the new contracts are getting longer. CoreWeave said the Meta deal runs through December 2032 and includes some of the first deployments of Nvidia’s Vera Rubin platform, so buyers are now reserving future chip generations years before the hardware is fully rolled out. (coreweave.com) The squeeze is no longer only about Nvidia. Broadcom said on April 6 that it signed a long-term agreement to develop future versions of Google’s artificial intelligence chips and related data-center networking products through 2031. (cnbc.com) (broadcom.com) Google’s chips are called tensor processing units, which are custom processors built for Google’s own artificial intelligence systems. Broadcom is the manufacturing and design partner helping turn those Google designs into the hardware that runs inside Google’s data centers. (broadcom.com) Anthropic then plugged into that same supply chain. The company said it signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation tensor processing unit capacity starting in 2027, and Broadcom said Anthropic will access about 3.5 gigawatts of that capacity. (anthropic.com) (cnbc.com) A gigawatt is power-plant scale, not server-rack scale. When an artificial intelligence company is reserving 3.5 gigawatts, it is effectively booking the electricity, buildings, cooling equipment, networking gear, and chips for an industrial system, not just buying software licenses. (anthropic.com) (pv-magazine.com) That is the shift underneath this week’s deals: artificial intelligence infrastructure is starting to look less like normal cloud spending and more like utilities, railroads, and aircraft fleets. The companies winning now are not only the ones with the best models, but the ones willing to sign six-year and seven-year contracts for chips, power, and buildings before demand arrives. (cnbc.com 1) (cnbc.com 2) (rystadenergy.com)
