Fractile Raises $220 Million Series B to Build In-Memory AI Inference Chips, Reaches Unicorn Valuation

Overview

London-based AI chip startup Fractile has closed a $220 million Series B funding round, reaching a post-money valuation of approximately $1 billion and entering unicorn territory, according to Yahoo Finance. The round, co-led by Accel, Factorial Funds, and Peter Thiel’s Founders Fund, will fund the tape-out of Fractile’s first inference chip, expand its software stack, and support early customer integration ahead of a targeted 2027 commercial launch.

Fractile, founded in 2022 by Oxford robotics PhD Walter Goodwin, is developing a chip architecture that performs inference computation directly inside SRAM cells co-located with compute logic — eliminating the costly round trips to off-chip DRAM that currently throttle large language model throughput.

What We Know

The funding round. The $220 million Series B exceeded the $200 million target Fractile had been pursuing earlier this year, as reported by The Next Web. In addition to the three lead investors, participation came from Conviction, Gigascale, O1A, Felicis, Buckley Ventures, and 8VC, alongside existing backers, according to SiliconAngle. The round brings Fractile’s total funding to approximately $237.5 million, following a $15 million seed led by Kindred Capital, the NATO Innovation Fund, and Oxford Science Enterprises, as Tom’s Hardware previously reported.

The technology. Fractile’s design approach, which the company describes as in-memory compute, “performs the matrix multiplications that dominate transformer inference inside SRAM cells located alongside the compute logic,” as described by The Next Web. This contrasts with conventional GPU-based inference, where the chip must repeatedly fetch model weights from separate high-bandwidth DRAM modules. Tom’s Hardware described the approach as storing data “directly next to the transistors that perform the arithmetic, rather than relying on off-chip DRAM.”

Goodwin explained the expected gains in an earlier interview with Network World: “the benefits are a hundred-fold increase in effective bandwidth and much higher energy efficiency.” In 2024, before test chips had been manufactured, he projected the approach could run large language models “100 times faster and 10 times cheaper than Nvidia’s GPUs,” as reported by Tom’s Hardware.

The inference bottleneck. Goodwin articulated the core market thesis in remarks accompanying the funding announcement: “Compressing a month of work into a day, a weekend of lab computation into a coffee break, will make all that work happen radically faster, but it will also make far more ambitious AI use cases economically viable,” according to SiliconAngle. He added that “the defining work of the 21st century will be marked by the engine of inference delivering immense and diffuse chains of intellectual inquiry, in drug discovery, in software engineering, in materials discovery.”

The Next Web noted that the Series B proceeds are allocated specifically for chip tape-out, software-stack development, and early customer integration, with no full production ramp funding included in this round.

Team and backers. Tom’s Hardware reported the engineering team includes veterans from Graphcore, Nvidia, and Imagination Technologies, and the company is building a proprietary software stack alongside its hardware. Notable angel investors and advisers include Pat Gelsinger, the former Intel chief executive, according to The Next Web.

UK expansion. In February 2026, Fractile announced plans to invest £100 million in its UK operations over three years, including the growth of existing sites in London and Bristol and the creation of a new hardware engineering facility in Bristol, as The Next Web reported. UK AI Minister Kanishka Narayan described the latest investment as “a strong vote of confidence in British AI,” according to Yahoo Finance.

Customer interest. Tom’s Hardware reported that Anthropic is in early discussions to purchase Fractile’s chips when they become available, which would make Fractile a fourth compute supplier for the AI lab alongside Nvidia, Google TPUs, and Amazon Trainium and Inferentia.

What We Don’t Know

Fractile has not disclosed manufacturing partners, wafer cost targets, or yield projections for its first chip. The performance claims of 100x speed and 10x cost reduction relative to current GPU setups were derived from simulations conducted before any test chips had been manufactured; independent validation of those figures has not yet been published. The company has also not announced confirmed production customers or a purchase agreement with Anthropic.

The current Series B does not include funding for full production ramp, meaning additional capital will be required before commercial volumes can be reached. Fractile’s chips are not expected to reach datacenter deployment until 2027, according to Tom’s Hardware.

Analysis

Fractile enters the inference chip market at a moment when the compute economics of large language model deployment are under intensifying scrutiny. As frontier models grow longer in context and more expensive to run, the data movement between processors and off-chip memory has emerged as a significant cost and latency driver — a constraint that SRAM-on-die designs are specifically positioned to address.

The company faces an established field of inference-focused competitors including Cerebras, Groq, Tenstorrent, SambaNova, and Etched, as Yahoo Finance noted. Nvidia holds an estimated 80 to 90 percent share of AI accelerator sales. What distinguishes Fractile’s approach is the architectural bet that SRAM-based in-memory compute can remove the memory bandwidth bottleneck entirely, rather than optimizing around it.

With $220 million in hand and a 2027 tape-out target, Fractile now has the runway to move from modeled performance to silicon. Whether the architecture delivers on its projected gains at production scale — and whether it can attract customers beyond early discussions — will determine whether this round is a launching pad or a high-water mark.