Fractile

Overview

Fractile builds inference chips that perform computation inside memory. The matrix multiplications that dominate transformer inference are executed within SRAM cells sitting alongside the compute logic, rather than by shuttling weights back and forth to separate DRAM. The company calls this memory-compute fusion, and the practical consequence is that its design uses no HBM stacks and no off-chip DRAM movement at all.

That is an unusually clean statement of the thesis every inference specialist is working from. The decode stage of language model inference is limited by memory bandwidth, not arithmetic; a GPU spends much of its capability waiting for weights to arrive. Fractile's response is not to add faster memory but to remove the journey.

The company was founded in 2022 by Dr Walter Goodwin, then a PhD student at the University of Oxford's Robotics Institute, and emerged from stealth in July 2024 with $15 million in seed funding. It raised a $220 million Series B in May 2026 at roughly $1 billion post-money. In August 2026 it became the subject of considerably more attention, for reasons covered below.

The architecture

Fractile has not published detailed silicon specifications, and no independent measurement of its hardware exists. What follows is the architecture as the company describes it.

The constraint implied by the third row is the one to watch, and it is the same one that shaped Groq. SRAM is fast but not dense. A design that holds all weights in SRAM needs a great deal of silicon to hold a large model, which pushes cost back up even as it removes the memory bottleneck. How Fractile resolves that trade at frontier model sizes is the most important undisclosed detail about the company.

Performance claims, and how they have changed

Fractile's public claims have moderated as it has moved toward silicon, which is worth recording plainly.

• At stealth exit in 2024, the company described running large language models up to 100 times faster than existing hardware, with operational costs lower by around 90%.
• More recent investor materials have been reported as framing the comparison as roughly 25 times faster at one-tenth the cost.

A four-fold reduction in the headline speed claim is not evidence that the technology fails to work. Early figures are typically simulated or derived from a narrow benchmark, and claims tightening as a design approaches manufacturable silicon is the normal and healthy direction of travel — the opposite pattern would be more worrying. But it does mean the widely repeated “100× faster” figure should not be treated as the company's current position.

All figures in this section are company claims. Fractile silicon is not commercially available, no third party has benchmarked it, and this publication has verified none of these numbers.

The Anthropic agreement and the $6.5B talks

On 19 August 2026 Bloomberg reported that Fractile had reached an initial agreement to sell approximately $250 million of chips to Anthropic, with the stated intention of expanding the contract, and that the company was in advanced talks to raise about $600 million at a $6.5 billion pre-money valuation, reportedly co-led by Redpoint Ventures and Lightspeed Venture Partners.

Both elements are reported rather than confirmed. Fractile and Anthropic declined to comment, and the round has not closed. Terms could change or the raise could fail to complete.

Taken at face value, the sequence is striking. Fractile was valued at roughly $1 billion in May. Three months later it is in talks at $6.5 billion — a rise of more than six times — on the strength of a supply agreement for chips that are not expected to be commercially ready until 2027. Anthropic is, in effect, committing $250 million to silicon that does not yet exist.

There is a coherent reading of why it might do so. Anthropic is diversifying its inference supply across Nvidia, Google, Amazon and AMD, and has separately been reported to be co-designing custom silicon with Samsung. Memory is the binding constraint on inference cost, and DRAM pricing has been under acute pressure. An architecture that removes HBM and off-chip DRAM entirely is therefore strategically valuable to a large model provider well before it is proven — the option is worth buying early, and $250 million is a modest hedge against inference costs at Anthropic's scale.

The less comfortable reading is that a pre-revenue contract for unshipped hardware is being used to justify a six-fold valuation increase, and that if the silicon slips or underperforms, both the contract and the valuation rest on the same unproven assumption.

Strengths and key questions

Leadership and investors

Dr Walter Goodwin founded Fractile in 2022 and leads it. His background is in AI and robotics rather than semiconductor design — he completed a PhD at the University of Oxford's Robotics Institute — which is less unusual in this cohort than it sounds; several inference architectures have come from people who approached the problem as a systems question rather than a chip-design one.

The investor list is a genuine signal. Pat Gelsinger joined as an angel investor and operating adviser in January 2025, shortly after leaving Intel. Hermann Hauser, who co-founded Arm, is also an angel, as are Stan Boland and Amar Shah, co-founder of the autonomous driving company Wayve. For a UK chip company, having both an Arm co-founder and a former Intel chief executive on the cap table is about as strong a domain endorsement as is available.

Funding history

Fractile is private and does not disclose revenue. It is pre-revenue with respect to chip sales, since its silicon is not expected to be commercially ready until 2027. The reported Anthropic agreement is a forward supply commitment, not recognised revenue.

How Fractile compares

Fractile is one of several companies attacking data movement, and the distinctions matter because they are frequently blurred in coverage.

• d-Matrix also uses digital in-memory compute, but pairs it with external memory and has already reached full production — a stage Fractile will not reach until 2027.
• Cerebras keeps data on-chip by making the chip enormous, and streams weights from external subsystems.
• Groq built an SRAM-based design with no HBM, and encountered precisely the capacity-versus-cost trade Fractile will face.
• OLIX is the other well-funded UK company avoiding HBM, but it computes optically rather than in SRAM. The two are separate companies with different architectures and are often confused in secondary reporting.

The Groq comparison is the instructive one. Groq had a real architectural advantage and a genuine customer base, and the resolution was a licensing transaction that transferred the technology and team to Nvidia. Anyone assessing Fractile at a reported $6.5 billion is implicitly taking a view on whether an independent path exists for an SRAM-based inference architecture, or whether the realistic ceiling is absorption by a larger player.

Sources

1. The Next Web — Fractile raises $220m to take its in-memory-compute inference chip into production. Series B size, investor list, valuation and architecture description.
2. Electronics Weekly — UK's Fractile raises $220m for inference ICs (May 2026). Round detail and participating investors.
3. Bloomberg — Fractile in talks for $6.5 billion value after Anthropic deal (19 August 2026). The reported Anthropic agreement and funding talks.
4. Tom's Hardware — Anthropic in early talks to buy DRAM-less inference chips from Fractile. Architecture context and memory-market pressure.
5. EU-Startups — Fractile exits stealth with seed funding (July 2024). Seed round and founding detail.
6. Data Center Dynamics — Pat Gelsinger invests in Fractile. Angel investment and adviser role.
7. Fractile. Company site. Note: it returned HTTP 403 to automated retrieval at the time of review, so company statements here are taken from reporting that quotes them directly.

The August 2026 Anthropic agreement and funding round are reported by Bloomberg and repeated widely, but are not confirmed by either company and the round has not closed. They are recorded here as reported, and this profile will be updated when the position is settled. See the editorial methodology.

Related pages