Anthropic Mulls Custom AI Chip Design as Claude Revenue Tops $30 B Run Rate
Key Points
- Anthropic is evaluating the design of proprietary AI chips but has not yet formed a dedicated team.
- The company’s annualized revenue run rate for Claude has exceeded $30 billion, up from $9 billion at the end of 2025.
- Current Claude workloads run on a mix of Google‑Broadcom TPUs, Amazon custom chips, and Nvidia GPUs.
- Anthropic signed a long‑term deal for 3.5 GW of TPU compute capacity beginning in 2027, three times its 2026 usage.
- The TPU agreement is contingent on Anthropic’s continued commercial success, per Broadcom’s SEC filing.
- Industry peers such as Meta and OpenAI are also pursuing custom AI silicon.
- Developing an advanced AI chip is estimated to cost about $500 million.
- Anthropic remains unprofitable, but its soaring revenue makes a large chip‑design investment more feasible.
San Francisco‑based Anthropic is weighing the development of its own artificial‑intelligence chips, according to three sources familiar with the effort. The move comes as the company’s annualized revenue run rate for its Claude models surged past $30 billion, up from roughly $9 billion at the end of 2025. Anthropic still runs workloads on a mix of Google‑Broadcom TPUs, Amazon‑custom silicon and Nvidia GPUs, and has just secured a long‑term deal for 3.5 gigawatts of TPU capacity beginning in 2027. The firm has not yet formed a dedicated chip team and may continue buying off‑the‑shelf silicon.
Anthropic, the San Francisco AI startup behind the Claude series of conversational agents, is quietly exploring the possibility of designing its own custom chips. Reuters cited three sources who said the effort is still in an early stage: no dedicated engineering team has been assembled and the company has not committed to a particular architecture.
The timing is notable. Anthropic announced that its annualized revenue run rate has climbed beyond $30 billion, a dramatic jump from the roughly $9 billion reported at the close of 2025. That rapid growth has generated a scale of compute demand that makes the economics of proprietary silicon worth a closer look.
Today the firm runs Claude on a heterogeneous fleet of hardware. It matches each workload to the processor that best fits the task, drawing from Google’s Tensor Processing Units built in partnership with Broadcom, Amazon’s custom silicon and Nvidia’s graphics‑processing units. The diversity reflects Anthropic’s strategy of leveraging the strengths of each platform rather than betting on a single supplier.
Just days before the chip‑design story broke, Anthropic signed a long‑term agreement with Google and Broadcom that will grant it access to about 3.5 gigawatts of TPU‑based compute capacity starting in 2027. The deal represents roughly three times the one gigawatt the company was using in early 2026, according to Broadcom’s SEC filing. The filing also noted that the expanded deployment hinges on Anthropic’s continued commercial success, an unusual hedge for a regulatory document.
The new TPU commitment builds on Anthropic’s November 2025 pledge to invest $50 billion in U.S. computing infrastructure. Broadcom, already a chip‑design partner for OpenAI, now finds itself at the center of a burgeoning custom‑silicon market that aims to provide alternatives to Nvidia’s general‑purpose GPUs.
Industry peers are moving in the same direction. Meta has been developing its own AI training chips, and OpenAI is reportedly working on custom silicon as well. Sources familiar with the broader market estimate that creating an advanced AI chip can cost around $500 million, reflecting the expense of hiring specialized engineers and validating a manufacturing process.
For Anthropic, the financial calculus is mixed. The company remains unprofitable, yet its revenue base has more than tripled in a four‑month span, making a half‑billion‑dollar investment appear more manageable than it would have a year ago. Whether Anthropic ultimately decides to build its own silicon or continue purchasing from established vendors will depend on how its compute needs evolve and how the economics of custom chips compare to the rapidly expanding TPU supply it has secured.