Anthropic has held discussions with London-based chip startup Fractile about purchasing its AI inference chips, according to The Information, citing people familiar with the matter. The talks would add Fractile as a fourth chip supplier alongside Google, Amazon, and Nvidia as Anthropic’s infrastructure struggles to keep pace with surging Claude demand.

Seeking Alpha confirmed the report, noting Fractile’s chips are designed specifically to run AI models more efficiently at inference time, the computationally expensive step where trained models generate responses.

Why Anthropic Needs More Silicon

The timing tracks with months of visible capacity strain. Anthropic told Fortune on April 24 that “demand for Claude has grown at an unprecedented rate, and our infrastructure has been stretched to meet it, particularly at peak hours.” The company’s annualized revenue run rate tripled from roughly $9 billion at year-end 2025 to $30 billion by April 2026, driven primarily by Claude Code adoption among developers and enterprise engineering teams.

That growth has come with costs. Anthropic has throttled Claude usage during peak hours, limited access to its Mythos model, and faced weeks of user complaints about degraded Claude Code performance before publishing a detailed engineering postmortem. In the same Fortune statement, Anthropic acknowledged that “compute is a constraint across the entire industry” and pointed to expanded partnerships with Amazon and Google as the near-term solution.

Adding Fractile would represent a different kind of fix: purpose-built inference silicon from outside the hyperscaler ecosystem.

What Fractile Builds

Founded in 2022, Fractile uses in-memory compute architecture to fuse processing and memory on the same chip, attacking the memory bandwidth bottleneck that limits how fast GPUs can run large models. The company claims its approach delivers inference up to 50 times faster at roughly 10% of the cost of GPU-based systems, according to Computer Weekly.

The pitch caught the attention of Intel’s former CEO Pat Gelsinger, who invested personally and wrote on LinkedIn that Fractile’s approach “overcomes the memory bottleneck that holds back today’s GPUs, while decimating power consumption.” He called the inference performance improvement “equivalent to years of lead on model development.”

Fractile is backed by the NATO Innovation Fund, Kindred Capital, and Oxford Science Enterprises. In February, UK AI minister Kanishka Narayan announced the company’s £100 million expansion across London and Bristol, including a new engineering facility for industrial-grade hardware, according to Computer Weekly. The team has grown to roughly 70 people with plans to add 40 more.

By late March, Sifted reported that Fractile was in talks to raise $200 million at a $1 billion valuation from Accel and Oxford Science Enterprises, citing the Financial Times.

The Supplier Concentration Problem

Anthropic currently relies on three chip suppliers, all of which are either competitors, investors, or both. Google supplies TPUs and just committed up to $40 billion in investment. Amazon provides custom Trainium chips and has invested up to $25 billion. Nvidia sells GPUs to nearly everyone.

Each relationship comes with leverage dynamics that work against Anthropic. Google and Amazon both run competing AI products. Nvidia allocates capacity across the entire industry. A deal with Fractile, if it materializes, would give Anthropic access to inference silicon from a supplier whose only business is selling chips, not running a competing cloud or AI service.

The deal also carries risk. Fractile has never shipped production silicon at scale. Its chips are expected to be ready sometime in 2026, according to Computer Weekly, but “expected” is not “deployed.” Anthropic would be betting on a 70-person startup to deliver the kind of manufacturing reliability that typically takes years to establish.

Inference as the Bottleneck

The talks underscore a broader shift in the AI industry: the binding constraint for frontier labs is no longer model capability but the infrastructure to run models at the scale users demand. Anthropic’s $30 billion run rate proves the demand exists. The question is whether the supply chain can keep up.

For agent platforms and developers building on Claude, the capacity constraints are not abstract. They show up as rate limits, degraded performance during peak hours, and throttled access to the most capable models. If Fractile’s technology works as advertised, Anthropic gets a path to serving more requests per dollar. If it doesn’t, the capacity crunch continues through at least late 2026, when Amazon and Google’s expanded infrastructure is expected to come online.