Anthropic's 3.5GW TPU Lock Is Not a Compute Deal. It Is a Moat.

The Real Story Is Not the Revenue Number

Anthropic's run-rate revenue crossing $30 billion is striking. But operators who stop there miss the harder signal: a new expanded agreement will allow Anthropic to tap 3.5GW of Google's TPU capacity from Broadcom. This is the largest public compute allocation disclosed by any AI lab, and it restructures the competitive landscape in ways revenue figures alone do not capture.

Broadcom disclosed in a securities filing that it will supply Anthropic with roughly 3.5 gigawatts of Google TPU capacity starting in 2027, in addition to the 1GW already coming online in 2026 under the Google Cloud agreement announced last October. The arrangement routes Google-designed TPUs to Anthropic via Broadcom as part of a multi-gigawatt commitment for next-generation TPU-based compute.

This is a supply chain lock-in, not a financing event. Anthropic traded optionality for certainty, and Google secured something no investment memo captures: a guaranteed, multi-year anchor customer whose compute dependency is now structurally woven into Google's TPU roadmap.

The Revenue Trajectory Reveals the Demand Signal

Anthropic's revenue trajectory has been: $87 million run rate in January 2024, $1 billion by December 2024, $9 billion by end of 2025, $14 billion in February 2026, $19 billion in March, and $30 billion in April. This is not a software growth curve—it is demand for compute at a scale most infrastructure teams have never planned for.

More than 1,000 business customers now spend over $1 million annually. Companies integrate Anthropic's models into internal workflows and customer-facing tools via APIs, with smaller revenue from premium chatbot subscriptions. API-driven enterprise inference is high-volume and margin-sensitive in ways subscriptions are not. Every new enterprise workload is a compute consumption event, which explains why the TPU capacity commitment exists.

Claude Code, the agentic AI coding tool launched publicly in mid-2025, became the fastest-growing product in company history, hitting $1 billion in annualized revenue within six months. Its run-rate revenue has since grown to over $2.5 billion, more than doubling since the beginning of 2026. Coding agents are computationally intensive—long context windows, multi-step reasoning, frequent tool calls. The product driving revenue acceleration is also the product straining capacity most.

What 3.5GW Actually Means

Gigawatts measure power, not chip count, and that framing is deliberate. It captures not just accelerators but the full infrastructure: power delivery, cooling, networking, physical space. The expanded partnership will come online starting in 2027 under a multiyear deal running until 2031, according to SEC filing.

Last October, Anthropic and Google announced a deal worth "tens of billions of dollars" for 1 million TPUs, expected to bring more than 1GW of capacity online this year. The April agreement triples that baseline to 3.5 gigawatts.

A consumption clause matters here. Broadcom disclosed that Anthropic's TPU capacity consumption is dependent on its continued commercial success. The full 3.5GW is not a guaranteed burn—it is a capacity reservation whose utilization scales with revenue. If growth decelerates, capacity sits idle; if it continues, Anthropic has headroom without competing for scarce silicon on the open market.

Non-Google-backed labs cannot make the same bet.

Google's Asymmetric Play

Google secures multiple advantages simultaneously: a major anchor customer for its custom silicon program, deepened dependency on Google Cloud as the primary inference substrate, and a credible counter-narrative to Nvidia's GPU dominance at frontier scale.

OpenAI has separately committed to 6GW of AMD GPU capacity, with the first gigawatt expected in the second half of this year. The contrast is revealing. OpenAI's commitment runs through AMD and Azure—Microsoft's infrastructure. Anthropic's runs through Google. Two frontier labs, two silicon bets, two cloud dependencies. The infrastructure layer is now the competitive moat.

Broadcom confirmed a long-term agreement with Google to supply custom TPUs and networking components through 2031. Broadcom's role as intermediary is not incidental. CEO Hock Tan has publicly targeted more than $100 billion in AI revenue by 2027, and multi-year supply commitments provide the demand signal required to justify capacity expansion at TSMC. Anthropic's commitment effectively pulls forward the manufacturing capacity it needs.

Who Is Exposed

The gap this deal creates is not about model quality. It is about serving inference at scale without capacity constraints during demand spikes. Labs without committed gigawatt-scale compute face a compounding problem: as enterprise customers grow, inference costs become unpredictable, spot market access tightens, and SLA commitments become harder to guarantee.

Anthropic now captures more than 73% of spending among first-time buyers of AI tools, while OpenAI's share has dipped to about 27%. If that new-customer capture rate holds, demand compounds faster than labs without pre-committed capacity can react.

Open-source projects and smaller startups face the steepest version of this problem. They cannot secure multi-year gigawatt commitments without the revenue to backstop consumption clauses. They will compete on the spot GPU market, where Nvidia's pricing power remains strongest.

What to Watch

1. Capacity utilization disclosure, Q3 2026. When the first gigawatt comes fully online, watch for changes in Anthropic's inference pricing. Margin expansion requires high utilization; pricing pressure signals underutilization.

2. OpenAI's TPU strategy. OpenAI relies on Azure and has committed to AMD GPUs. Watch whether it secures any Google TPU allocation—that would signal either Google's willingness to play both sides or OpenAI hedging its Azure dependency.

3. 2027 capacity activation. When 3.5GW of new capacity comes online, watch for announcements of additional commitments before then. The consumption clause protects Google but does not cap Anthropic's demand upside.

4. Regulatory scrutiny on compute concentration. Anthropic is already contesting a federal designation that labeled it a potential supply chain risk, warning it could cost billions in lost revenue. Compute concentration at this scale will attract antitrust attention. Watch for DOJ or FTC inquiries into exclusive capacity arrangements.

5. Second-tier lab consolidation. Labs without comparable capacity commitments face a narrowing window. Expect M&A activity or strategic partnerships as smaller labs seek shelter under hyperscaler agreements before the 2027 capacity cliff hardens.