Anthropic Is Eyeing Its Own Chips. The Smarter Move Might Be the One It Has Already Made.

Anthropic Is Eyeing Its Own Chips. The Smarter Move Might Be the One It's Already Made.

When a company the size of Anthropic starts thinking seriously about designing its own silicon, the announcement tends to get ahead of the actual plan. That's roughly where things stand: three sources tell Reuters the company is "exploring" custom chip development, plans are "in early stages," no dedicated team has formed, and Anthropic may yet decide to stick with buying chips instead. Read the headline and you might think the company is ditching its partnerships. Read the fine print and you'll notice it just committed to billions in compute infrastructure with the companies it would theoretically compete against.

Anthropic's actual near-term silicon strategy was spelled out this week in a Broadcom filing: a supply assurance agreement for next-generation TPUs, starting in 2027, to the tune of 3.5 gigawatts of compute, The Register reported. That is not the move of a company pivoting away from Google's ecosystem. It is a company buying an enormous amount of capacity from it, while simultaneously keeping an exploratory finger on the custom-silicon button.

The revenue picture helps explain why. Anthropic's run-rate has crossed $30 billion, up from $9 billion at the end of 2025, according to figures the company disclosed this week. The enterprise cohort is scaling too: over 1,000 customers now spending more than $1 million annually with Claude, doubling from the 500 it reported when it closed its Series G at a $380 billion valuation in February. That kind of growth changes your leverage calculations. When you are spending that much on compute, the economics of building your own chip start to look different than they do when you are a startup renting time on someone else's cloud.

The numbers are also a data point in a broader revenue race Anthropic appears to be winning. OpenAI's current run-rate sits around $24 billion, according to figures the company has confirmed separately. Anthropic, the smaller competitor by model capability benchmarks, has pulled ahead on revenue. The gap is not the story; what it means for compute allocation is.

Custom chip design is not a theoretical exercise at this scale. Industry sources estimate a single tape-out for an advanced AI chip costs roughly $500 million, a figure that includes the engineering talent, the masks, and the defect rates inherent in leading-edge manufacturing processes. Meta and OpenAI are both on the same path, which means the engineers who know how to do this work are being recruited across the industry simultaneously. OpenAI is targeting mass production at TSMC in 2026, according to sources who spoke with Reuters last year. Anthropic has set no comparable public timeline.

What Anthropic has set is a $50 billion commitment to U.S. computing infrastructure. The Broadcom filing shows it is backing that commitment with a contractual claim on 3.5 gigawatts of next-generation TPU capacity. That is a significant bet, and Broadcom's own disclosure suggests the financial arrangement is not without risk on its end: the company flagged Anthropic's purchase commitments as a notable risk factor in its regulatory filing, an unusual level of candor for a supplier about a flagship customer.

The question Anthropic is implicitly asking is not whether to build a chip. It is whether the leverage it is accumulating through partnerships with Google and Broadcom is sufficient, or whether it needs a more direct claim on the silicon stack beneath its models. At $30 billion in annual revenue, the company can afford to explore both answers. The 3.5 gigawatt commitment suggests it has decided the partnership track is the right one for now, and the custom silicon exploration is the option it is keeping open.

That is a rational posture for a company with real revenue and real options. It is not the posture of a company that has decided the current ecosystem is broken. But if the exploration produces a viable design, the economics of the AI infrastructure landscape will look different than they do today.