Big capacity deal report

A report moving around markets says Anthropic is lining up an enormous block of Google artificial intelligence chip capacity, with Broadcom tied to the hardware stack behind it, and the scale being discussed is large enough to be measured in gigawatts of power rather than racks of servers. The report is not an official securities filing, so it deserves caution before anyone treats every detail as settled fact. (markets.financialcontent.com) The basic idea is simple: the best artificial intelligence companies now need so many chips that “how many can you get” matters almost as much as “how smart is your model.” Anthropic said in October 2025 that it was expanding its use of Google Cloud tensor processing units, or Google’s in-house artificial intelligence chips, in a deal worth tens of billions of dollars that was expected to bring well over a gigawatt of capacity online in 2026. (anthropic.com) Those chips are called tensor processing units because Google built them specifically for the matrix math used in machine learning, instead of for general computing. Google says its cloud tensor processing units are designed for both training models and running them after training, which is the difference between building the engine and keeping the car on the road every day. (cloud.google.com) Anthropic did not start this week. In February 2023, it said it had selected Google Cloud as a cloud provider and would use Google’s graphics processing units and tensor processing unit clusters to train, scale, and deploy its systems. (anthropic.com) The newer wrinkle is that Anthropic now says Google is only one lane in a three-lane setup. In an April 2026 post, Anthropic said it trains and runs Claude across Amazon Web Services Trainium chips, Google tensor processing units, and NVIDIA graphics processing units so it can place different workloads on the hardware that fits them best. (anthropic.com) That sentence tells you what smaller companies are up against. If Anthropic is spreading work across Amazon chips, Google chips, and NVIDIA chips at the same time, it is doing that because frontier artificial intelligence demand has outgrown the era when one supplier could comfortably handle everything. (anthropic.com) Broadcom sits one layer lower in the stack, where the economics get brutal. Broadcom says custom artificial intelligence accelerators are built from compute, memory, network input and output, and advanced packaging, which means the bottleneck is not just the chip design but the whole assembly line that turns designs into usable systems. (broadcom.com) Packaging matters because a modern artificial intelligence chip is less like a single processor and more like a tiny neighborhood wired together at very high speed. Broadcom says its newer packaging methods are aimed at improving performance, lowering power, and shrinking size for large artificial intelligence systems, which is exactly the kind of engineering that becomes scarce when a few customers start reserving capacity years ahead. (broadcom.com) Google has also been pushing more powerful versions of its own chips. In November 2025, Google Cloud said its Ironwood tensor processing unit offered 10 times the peak performance of TPU v5p and more than four times the performance per chip versus TPU v6e for training and inference, which helps explain why a model developer would want to lock in access before everyone else does. (cloud.google.com) Put that together and the picture is less “one flashy partnership” than “specialized supply is being spoken for in bulk.” Anthropic’s official statements confirm a very large Google tensor processing unit expansion and a multi-platform chip strategy, while the market report adds the more aggressive claim that Broadcom-linked infrastructure and dedicated capacity are being locked in at extraordinary scale. (anthropic.com 1) (anthropic.com 2) (markets.financialcontent.com) For everyone outside that top tier, the practical lesson is boring and expensive at the same time: you cannot assume the cheapest or fastest chip will be available when you need it. Anthropic is explicitly talking about matching workloads to different chips, and that is usually what companies do when supply, cost, latency, and resilience all have to be managed at once. (anthropic.com)