Anthropic reports revenue has tripled past $30 billion and explores custom AI server chip manufacturing partnerships.

Anthropic, the AI safety company behind the Claude family of models, is exploring the development of its own custom AI server chip. The move comes as the company's revenue has exploded, creating both the financial runway and the strategic urgency to rethink its chip supply chain from the ground up.

The company's run-rate revenue exceeded $30 billion in 2026, a staggering jump from roughly $9 billion at the end of 2025.

The initiative is still in its early stages, with no finalized design or dedicated chip team publicly announced. Anthropic is evaluating both ASICs optimized specifically for inference tasks and custom GPU configurations. Estimated development costs for these advanced AI chips sit around $500 million, according to Reuters.

Anthropic currently runs its models on a diversified hardware stack that includes Nvidia GPUs, Google TPUs, Amazon Trainium chips, and custom silicon from Broadcom. The company has committed over $10 billion to Broadcom alone and has plans to procure more than 1 million chips from Amazon and Google. Anthropic is also working with Google and Broadcom on multi-gigawatt TPU capacity expected to come online from 2027.

The inference side of the equation is particularly interesting. Training a large language model is a one-time or periodic compute-intensive event. But inference — the process of actually running the model to answer user queries — is where the ongoing costs pile up. An ASIC optimized for inference could slash per-query costs dramatically.

For Broadcom, Anthropic's $10 billion-plus commitment suggests the chipmaker is a beneficiary of this custom silicon trend. Broadcom designs and manufactures custom chips for other companies, meaning more in-house chip ambitions could translate to more Broadcom revenue.

Anthropic's revenue trajectory — from $9 billion to over $30 billion in roughly one year — suggests the demand-side economics justify massive infrastructure bets. Designing world-class silicon is a fundamentally different discipline than training language models, and as of July 2026, there has been no public announcement of an unveiling or timeline for any Anthropic ASICs.