Anthropic is in early-stage discussions to buy AI inference chips from Fractile, a London-based semiconductor startup, according to a report from The Information published on May 3, 2026. The potential deal would give Anthropic access to specialized chips designed to run AI models more efficiently — and marks the Claude maker’s most concrete step yet toward diversifying away from its current reliance on Nvidia, Google, and Amazon for its computing infrastructure.
No deal has been finalized, and the talks are described as early-stage. However, the conversations reveal the intense pressure Anthropic is under to bring down the cost of running its AI models at scale.
What Fractile Makes and Why It Matters
Fractile is a UK-based chip startup building what are known as inference accelerators — specialized processors designed not to train AI models, but to run them after training has been completed. Unlike general-purpose GPUs, inference chips are engineered for high-throughput, repetitive workloads with lower power consumption and reduced latency. That combination makes them particularly attractive for applications where cost per query and response speed matter most, such as enterprise AI tools, consumer chatbots, and large-scale API deployments.
Fractile’s approach uses SRAM rather than separate high-bandwidth memory chips, which is expected to produce significant efficiency gains. The company is currently in talks to raise more than £160 million at a valuation of around £800 million, placing it among the most prominent AI chip startups in Europe.
Fractile’s chips are not yet commercially available. According to The Information, they could ship as soon as 2027 — meaning any agreement between Anthropic and Fractile would be a forward-looking procurement deal rather than an immediate supply arrangement.
Why Anthropic Needs This
Anthropic’s revenue has grown dramatically in recent months. The company’s run-rate revenue has more than tripled from approximately $9 billion at the end of 2025 to around $30 billion in mid-2026, according to reporting by industry analysts. That growth has put enormous strain on its compute infrastructure.
According to The Information, Anthropic expects its annual spending on servers and chips to reach tens of billions of dollars. Its gross profit margin for AI product operations fell short of internal targets last year because inference costs were higher than expected — a problem that OpenAI has also faced with its own products.
This is the structural challenge driving demand for inference chips. Training a large language model is an enormous one-time cost, but running that model in response to millions of user queries every day is an ongoing and rapidly scaling expense. For AI companies whose products depend on fast, affordable inference, the cost of running their models has become a competitive variable.
Anthropic currently uses a mix of Google’s TPUs and Amazon’s custom chips. Earlier this year, it signed a long-term deal with Google and Broadcom — which helps design TPUs — as part of a broader $50 billion commitment to US computing infrastructure. However, as the company put it, depending heavily on any single supplier creates both pricing pressure and strategic vulnerability.
This is the same logic that Google applied when building TPUs and Amazon applied when developing Trainium and Inferentia. It is now trickling down to AI labs like Anthropic.
The Broader Chip Diversification Race
Anthropic’s interest in Fractile fits a wider pattern. AI companies at every level of the stack are actively seeking alternatives to Nvidia for inference workloads. Cerebras and Groq, along with Fractile, are all developing inference-focused chips that use SRAM architectures to sidestep the latency bottleneck of external high-bandwidth memory.
For Anthropic specifically, the Fractile discussions build on a broader diversification strategy. Unlike OpenAI and xAI — which depend heavily on Nvidia for their compute — Anthropic has deliberately spread its chip supply across multiple providers. The Fractile talks would extend that strategy into specialized inference silicon.
This has direct implications for users of Claude, Anthropic’s AI model. Lower inference costs translate to lower prices and higher availability for end users — the same dynamic that has driven Google Gemini to expand its capabilities and offer more features at lower price points to consumers and developers.
The push to reduce inference costs also sits at the center of a much larger industry conversation. Big Tech companies collectively are expected to spend over $700 billion on AI infrastructure in 2026. As that buildout scales, companies that can run frontier models efficiently and cheaply will have a structural advantage over those that cannot.
As Alphabet’s Q1 2026 earnings demonstrated, the returns from AI infrastructure investment can be extraordinary — but only for the companies that control enough of the stack to capture the value they create.
Frequently Asked Questions
What is Fractile?
Fractile is a London-based AI chip startup building inference accelerators — processors designed to run trained AI models efficiently using SRAM architecture rather than high-bandwidth memory. The company is in talks to raise over £160 million at a £800 million valuation.
Why is Anthropic talking to Fractile?
Anthropic is looking to diversify its chip supply beyond its current dependence on Nvidia, Google TPUs, and Amazon chips. Specialized inference chips like those Fractile is developing could help Anthropic reduce the cost of running its Claude AI models at scale.
When would Fractile’s chips be available?
According to reports, Fractile’s chips could ship as early as 2027. Any deal between Anthropic and Fractile would therefore be a forward-looking procurement agreement rather than an immediate supply arrangement.
What is an inference chip and how is it different from a GPU?
An inference chip is a processor optimized specifically for running trained AI models, rather than training them from scratch. Unlike general-purpose GPUs, inference chips prioritize low latency, high throughput, and energy efficiency for the repetitive queries that production AI systems handle at scale.
Has Anthropic confirmed the Fractile deal?
No. As of May 3, 2026, the talks are at an early stage and no agreement has been announced by either Anthropic or Fractile. The original report came from The Information, citing people familiar with the matter.
Conclusion
Anthropic’s reported discussions with Fractile reflect the growing recognition across the AI industry that inference costs are a strategic problem — not just an operational one. As Claude continues to scale and Anthropic’s revenue grows, securing a reliable, cost-efficient pipeline of inference chips is becoming as important as model quality itself. The Fractile talks may or may not result in a deal, but they confirm that the Claude maker is thinking seriously about owning more of the compute stack it depends on to serve its users.
Leave a Reply