Cloud giants lock supply and customise stacks
Major cloud players are deepening long-term hardware ties and customising infrastructure rather than buying generic servers, signaling a more bespoke compute stack. Google struck a multi‑year deal to keep deploying Intel Xeon alongside custom IPUs, and Alphabet has extended TPU and networking supply agreements tied to partners like Broadcom while expanding compute access for AI labs. That trend reshapes vendor selection: long supply agreements and custom silicon roadmaps matter as much as raw model performance. (tomshardware.com) (simplywall.st)
Google just made two different bets at once: on April 9 it expanded a multiyear deal to keep Intel Xeon chips inside Google Cloud, and three days earlier Broadcom disclosed long-term agreements to keep building future Google Tensor Processing Units and networking gear through 2031. (intel.com) (sec.gov) That is not the old cloud model where a provider buys a pile of mostly standard servers and swaps vendors when prices move. Google is locking in a custom parts list years ahead, the way a carmaker locks engines, transmissions, and electronics before the factory line starts. (cnbc.com) (sec.gov) The Intel side of the deal is about the central processing unit, which is the general-purpose brain that schedules work, moves data, and keeps a server running even when a graphics chip gets the attention. Intel said Google will align across multiple generations of Xeon processors for artificial intelligence, inference, and general-purpose workloads. (intel.com) (cnbc.com) Google is not pairing those central processors with off-the-shelf support chips either. Intel said the two companies are expanding co-development of custom infrastructure processing units, which are chips that take over networking, storage, and security jobs so the main processor can spend more time on customer workloads. (intel.com) (newsroom.intel.com) The Broadcom side is about Google’s own artificial intelligence accelerator, called the Tensor Processing Unit. In an April 6 filing, Broadcom said it signed a long-term agreement to develop and supply future generations of those chips and a separate supply assurance agreement for networking and related components through 2031. (sec.gov) (cnbc.com) Those Tensor Processing Units are already central to Google’s cloud pitch. Google said in November 2025 that frontier models including Gemini, Veo, Imagen, and Anthropic’s Claude train and serve on Tensor Processing Units, and it said its Ironwood chip is the seventh generation of that line. (cloud.google.com) (blog.google) Ironwood shows what “custom stack” means in practice. Google said the chip was built specifically for inference, can scale to 9,216 liquid-cooled chips in one system, and is tied to Google’s own inter-chip network and software stack instead of a generic server recipe that any cloud company could copy. (blog.google) (cloud.google.com) Then Google turned that internal hardware roadmap into a customer product. Anthropic said on April 6 that it signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation Tensor Processing Unit capacity starting in 2027, and Broadcom’s filing put Anthropic’s share at about 3.5 gigawatts. (anthropic.com) (sec.gov) So the cloud sale is changing shape. Instead of just renting raw computing by the hour, Google is selling access to a stack where the central processing unit comes from Intel, the artificial intelligence accelerator is Google’s own design, the networking is locked in with Broadcom, and a lab like Anthropic reserves capacity years before the machines are fully online. (intel.com) (sec.gov) (anthropic.com) That shifts the pecking order for suppliers. A chip company now needs more than a fast benchmark score; it needs a multiyear roadmap, packaging, networking, software hooks, and enough manufacturing certainty that a cloud giant will bet an entire data-center design on it through the end of the decade. (cnbc.com) (sec.gov) And it means the next fight in artificial intelligence infrastructure may be less about one superstar chip and more about who can assemble the best full machine. Google’s recent deals say the winning cloud may look less like a warehouse of interchangeable boxes and more like a custom-built factory where every part is chosen years in advance. (intel.com) (cloud.google.com)
