Anthropic is discussing with the British startup Fractile to integrate its AI inference chips into its infrastructure. The creator of Claude is thus looking for a fourth silicon supplier, in addition to Nvidia, Google, and Amazon. The information, published by The Information on May 1, 2026, marks a strong signal for the specialized AI semiconductor sector.
According to sources close to the matter cited by The Information, discussions remain at a very preliminary stage. No volumes, delivery schedules, or contractual commitments have been defined. The news nevertheless confirms a fundamental trend: the growing pressure from the demand for Claude is pushing Anthropic to explore all avenues to secure its computing power.
An exploratory discussion still
The talks between Anthropic and Fractile are based on unofficial sources. Neither Anthropic nor Fractile have released a statement at the time of publication of this article, and neither company has denied the information. This silence is consistent with a discussion described asexploratory : Fractile is not yet able to deliver its chips in volume, and the creator of Claude has no interest in formalizing a partnership until the actual performance of the silicon is validated.
According to information reported by several specialized media outlets, including Tom's Hardware, Fractile's chips are not expected to reach their commercial maturity before 2027. This deadline roughly corresponds to the large-scale deployment of Google-Broadcom TPUs ordered by Anthropic earlier this year. The timeline therefore leaves a narrow window to turn an intention into a firm contract.
Fractile, a technical bet on SRAM memory
Fractile was founded in 2022 by Walter Goodwin, a former doctoral student at the Oxford Robotics Institute. The London-based company is developing a chip architecture known as memory-compute fusionThe principle is simple to understand. Instead of constantly transferring data between the processor and external DRAM memory, Fractile places the memory and the computing unit on the same integrated circuit, using static memory. SRAM This choice greatly reduces latency and energy consumption, two major bottlenecks in current AI inference.
According to Fractile, this approach allows a large language model to run up to 100 times faster and at a tenth of the cost of a traditional Nvidia GPU. These figures, communicated by the company based on internal simulations, have not been validated by independent benchmarks. The company has not yet commercialized its chips; the first test chips have been in tape-out phase from its London and Bristol teams, as Walter Goodwin indicated in an interview with eeNews Europe in early 2025.
Pat Gelsinger, former CEO of Intel and now an investor in the company, has publicly supported this approach in a LinkedIn post. According to him, the rise of reasoning models, which generate thousands of tokens in output, accentuates the limitations of existing hardware roadmaps. His endorsement gives Fractile rare industrial credibility for a startup at this stage of development.
Anthropic's multi-supplier strategy
The publisher of Claude has always avoided dependence on a single silicon supplier. Today, its models run on three complementary chip families: the Nvidia GPUs, Amazon's Trainium processors (deployed as part of project Rainier), and Google's TPUs, for which an order of over one gigawatt was announced in October 2025.
Fractile's potential arrival as fourth source not only addresses a need for cost optimization. It is part of a strategic coverage logic. The shortage of AI chips and the soaring prices of HBM memory, which equips Nvidia GPUs, have eroded Anthropic's gross margins in 2025. A chip designed specifically for inference, and which would avoid external DRAM, represents a serious avenue for improving the economic unit of each Claude request.
Growth that forces Anthropic's hand
The financial context explains the urgency. Anthropic's annualized revenue has gone from approximately 9 billion dollars at the end of 2025 to over 30 billion dollars in April 2026, according to figures reported by several specialized media outlets. This surge places considerable pressure on existing infrastructure and makes any credible alternative to GPUs strategically relevant.
This dynamic is part of a broader trend. The market for chips dedicated to inference is now valued at over $50 billion by 2026, according to industry estimates. Inference workloads alone account for about two-thirds of total AI computation, compared to a much smaller share two years ago when model training still dominated budgets. The economic center of gravity for AI is shifting towards the production phase, and players who do not secure efficient inference infrastructure risk seeing their margins melt away.
A 2027 calendar and several serious reservations
Several limitations temper the enthusiasm. Fractile has not yet produced a real silicon chip. The announced performance is based on internal simulations and projections. The company employs a small team, even though it has recently been strengthened with former engineers from Graphcore, Nvidia, and Imagination Technologies.
The precedent of Graphcore, considered a few years ago as the British flagship of hardware AI, should be recalled. The company was acquired by SoftBank in 2024 for approximately $600 million, less than its total venture capital fundraising. The road from a promising architecture to a commercial product at the scale of a hyperscaler remains extremely difficult, and several startups in the sector have learned this the hard way.
Fractile also faces head-on competition. Groq, acquired by Nvidia in December 2025 for $20 billion, and Cerebras are pursuing similar approaches based on SRAM or near-memory computing. Tenstorrent and Olix, which raised $220 million earlier this year, also occupy this niche. The accelerated inference market attracts top engineers and the most patient capital, but it will only leave room for a small number of winners.
The UK seeks its place in the AI chip race
Anthropic's interest in Fractile highlights the dynamics of the British semiconductor sector. In February 2026, Fractile announced an investment of £100 million over three years to expand its operations in London and Bristol, including the opening of a new hardware engineering facility. This announcement was hailed by Kanishka Narayan, the British Minister for AI, as a model for national technology investment.
The company is also in discussions to raise $200 million at a valuation exceeding $1 billion, according to the Financial Times. The round would be led by Accel, with participation from Founders Fund and 8VC among potential investors. Pat Gelsinger is also among the supporters, in a strategic advisor role. If the fundraising materializes and the first test chips pass the technical milestones, Fractile will officially enter the league of major specialized foundries.
For Anthropic, opening a fourth hardware track, even by 2027, is a long-term bet consistent with the industry's industrial trajectory. The next concrete test will be to see if discussions with Fractile go beyond the exploratory stage and result in a defined volume or a firm delivery schedule before Google-Broadcom TPUs absorb most of the new Claude workloads. Above all, the announcement confirms that the future profitability of language models now depends as much on specialized silicon as on algorithmic progress, and the coming months will tell if the British startup truly joins the very small list of suppliers capable of powering a leading player like Anthropic.
No Comment! Be the first one.