Anthropic Locks In 3.5 Gigawatts of Google TPUs from Broadcom for 2027

Anthropic has signed an expanded compute deal with Google and Broadcom that secures 3.5 gigawatts of next-generation TPU capacity starting in 2027. The numbers come from a Broadcom SEC filing, and they confirm what has been a months-long buildout: the Claude maker is stacking one of the largest captive compute commitments in the industry.

The announcement lands as Anthropic confirms its annual revenue run rate has crossed $30 billion — a near-doubling in under a year.

The deal, in numbers

The 3.5 gigawatts are additive. Anthropic already has roughly 1 GW of Google Cloud TPU capacity coming online in 2026 under an agreement announced last October. The new agreement stacks another 3.5 GW on top, deployed starting in 2027.

For context: 3.5 GW is enough electricity to power several million U.S. homes. Routed into AI infrastructure, it represents one of the largest single compute commitments ever disclosed by an AI lab.

Broadcom is the silicon supplier. Google Cloud is the operator — the TPUs will physically run inside Google data centers. Anthropic is the tenant. It's a three-way arrangement that gives each party exactly what it needs: Anthropic gets priority capacity, Google gets a marquee customer locked into its TPU roadmap, and Broadcom gets a long-duration silicon contract.

Why Anthropic is moving now

The easiest way to read this deal is through revenue. Anthropic's run rate at $30 billion means the company is consuming compute faster than any forecast from a year ago. Claude Code alone is now a significant fraction of enterprise AI spend, and Claude's Opus and Sonnet models are increasingly the default for agentic workloads inside large enterprises.

Anthropic's CFO Krishna Rao, speaking to the broader compute strategy earlier this month, framed the expansion as necessary to meet demand that has "outpaced even aggressive internal forecasts."

Why this matters: frontier model labs no longer compete primarily on algorithms. They compete on whether they can get enough power and chips in 2027 and 2028. A 3.5 GW commitment today is the kind of move that determines who can serve the enterprise market in three years.

Broadcom's dual-track AI bet

The Anthropic expansion is part of a bigger Broadcom strategy. The same chipmaker has committed to designing and supplying future generations of Google's TPUs through 2031, and separately has a $10 billion custom silicon program with OpenAI announced as a 10 GW co-development effort.

That makes Broadcom the rare company directly plumbed into two of the three frontier labs' roadmaps. It is also one of the clearest signals that custom AI silicon — not general-purpose GPUs — is where the hyperscaler and lab spend is now flowing.

Nvidia is still the dominant vendor by revenue and by installed base. But the gigawatt-scale commitments increasingly go to custom chips designed for specific workloads.

What analysts are saying

Futurum Group called the deal a "structural advantage" for Anthropic, arguing that locking in gigawatt-scale TPU supply this early forecloses one of the biggest risks the lab faces — not being able to serve demand.

The Motley Fool framed it as a dual win for Alphabet and Broadcom shareholders, noting that Anthropic's commitment effectively underwrites multiple generations of TPU demand for Google Cloud. Tom's Hardware reported that Anthropic's revenue disclosure alongside the compute expansion is likely to accelerate the next round of fundraising conversations.

Industry impact

For customers, the practical effect is straightforward: the companies that can access the most frontier-grade compute over the next three years are Anthropic, OpenAI, Google DeepMind, and Meta. This deal reinforces Anthropic's place on that short list.

For Nvidia, it's another data point in the slow erosion of "GPUs are the only option" as the default assumption. Anthropic's models already run on a mix of Nvidia GPUs, Google TPUs, and Amazon's Trainium. The TPU expansion tilts the future mix further toward custom silicon.

For enterprises building on Claude, the capacity commitment reduces one of the real anxieties of betting on a specific model provider: whether the provider can scale inference to meet production demand.

What's next

The first of the new capacity comes online in 2027 and ramps over multiple years. That lines up with Google's rollout of its 8th-generation TPUs — TPU 8t and TPU 8i — which were announced at Cloud Next this week and are expected to be the silicon backing much of this Anthropic capacity.

Watch for two things. First, whether Anthropic discloses further capacity expansions later in 2026 — a 3.5 GW commitment still might not be enough if revenue keeps doubling. Second, whether the company announces an equivalent deal on the GPU side, which would signal it wants to keep its compute footprint diversified rather than consolidating on TPUs.

Bottom line

Anthropic just bought itself runway through 2028. The 3.5 GW TPU commitment is the kind of move only a lab with $30 billion in run-rate revenue can make, and it resets the bar for what "serious" frontier compute looks like. The next AI race isn't only about model quality — it's about which labs locked in the power and silicon early.