Anthropic is in early discussions with UK-based chip startup Fractile to buy its inference accelerators, The Information reported on May 2, 2026. If a deal materialises, Fractile would become Anthropic’s fourth chip supplier, alongside Nvidia, Google’s TPUs, and Amazon’s custom silicon.

The talks are at an early stage and no agreement has been reached. Fractile’s chips are not expected to reach commercial readiness until around 2027. But the fact that Anthropic is exploring the relationship at all says something about how seriously the company is treating its compute supply chain as its revenue grows.

What Fractile Actually Does

Founded in 2022 by Oxford PhD Walter Goodwin, Fractile is building inference chips with an architecture that is meaningfully different from conventional AI hardware.

Most inference chips today separate compute and memory. The processor does the calculation, then retrieves weights and activations from external high-bandwidth memory chips. That data transfer is a significant bottleneck, particularly for large language models where the sheer volume of parameters being moved between compute and memory limits how fast inference can run.

Fractile’s approach co-locates memory and compute on the same die using SRAM, eliminating the need to shuttle data to separate DRAM chips. The company claims this architecture delivers inference up to 25 times faster at one-tenth the cost of incumbent hardware.

Those are significant claims, and they are not yet independently verified at commercial scale. But the underlying physics of the approach is sound: keeping memory closer to compute reduces latency and energy consumption, and SRAM is faster than DRAM even when the raw capacity is smaller.

Why Anthropic Is Interested

Anthropic’s annualised revenue run rate crossed $30 billion in March 2026, up from approximately $9 billion at the end of 2025. With that growth comes a compute bill that now scales to tens of billions of dollars per year.

Inference is a significant part of that cost. Every time a Claude user sends a message, generates code, or runs an agentic task, Anthropic pays for the chips that process it. At current scale, shaving even a few percentage points off inference costs is worth hundreds of millions of dollars annually.

Anthropic already has contracted Google and Broadcom TPU capacity coming online from 2027, as part of an April 2026 partnership covering approximately 5 gigawatts of compute. Exploring Fractile in parallel is not a signal that the Google deal is inadequate. It is a signal that Anthropic wants options.

Diversifying suppliers gives Anthropic negotiating leverage. It reduces the risk of being dependent on any single hardware vendor. And if Fractile’s claims about inference efficiency hold up at commercial scale, the economics would compound significantly as Claude traffic grows.

The Broader Context

Anthropic is not alone in developing hardware relationships outside the standard Nvidia and hyperscaler stack. The entire AI industry is reckoning with the fact that general-purpose GPU infrastructure was not designed specifically for transformer model inference. Purpose-built inference chips, whether from established vendors like Groq and Cerebras or newer entrants like Fractile, represent a genuine bet that the hardware layer can be substantially improved for this specific workload.

For businesses running AI at scale, the infrastructure investment happening at this level has real implications. The inference costs that feel expensive today are being actively targeted by multiple competing approaches. Within two to three years, the cost curve for running large language models at enterprise scale is likely to look materially different.

What This Means for Business

The Fractile story is not actionable for most businesses today. No one should be making procurement decisions based on chips that are not yet commercially available.

What it does reinforce is a broader pattern: the major AI providers are investing seriously in reducing the cost of inference. Anthropic exploring a fourth chip supplier, Google developing next-generation TPUs, Amazon building custom silicon, OpenAI pursuing its own hardware programme, these are not separate stories. They are the same story about a technology where the hardware economics are still being figured out.

For businesses deploying AI now, the relevant implication is that your AI costs in 2027 and 2028 will likely be lower than your costs today, possibly substantially lower. That should factor into how aggressively you are willing to invest in AI infrastructure and tooling now, knowing the unit economics are likely to improve.

The window for building genuine AI capability is not closing. It is being held open by hundreds of billions of dollars in infrastructure investment from companies that need AI to be affordable enough to deploy everywhere.

---

For a deeper walkthrough of tools like this and how they fit together, the free Working With Claude field guide covers the ecosystem end to end. Get the guide.