Fractile's $220 million funding round comes as Anthropic targets its UK semiconductor operations.

      Accel led the funding round for the London-based chip startup, with Pat Gelsinger joining as an angel investor shortly after reports indicated that Anthropic was in preliminary talks to become a client. Fractile, which focuses on designing inference chips that integrate compute and memory on a single die, announced on Tuesday that it has secured $220 million to move its hardware into production. This funding surpasses the $200 million target the company had circulated in late March, as noted by Electronics Weekly, and positions Fractile among European chip firms presenting themselves as viable alternatives to Nvidia for inference solutions.

      The investor profile adds credibility to this round. Accel is reported to have taken the lead, with former Intel CEO Pat Gelsinger also contributing as an angel investor and operational advisor. Existing investors, including Kindred Capital, the NATO Innovation Fund, and Oxford Science Enterprises—which co-led Fractile’s $15 million seed funding in July 2024—are participating in this round as well.

      Fractile’s technology challenges existing architectural paradigms. Traditional AI accelerators, such as Nvidia’s H- and B-series GPUs, separate the compute die from high-bandwidth memory, incurring energy and latency costs when moving data between the two. In contrast, Fractile's design performs the matrix multiplications central to transformer inference within SRAM cells that are placed next to the compute logic, adopting an in-memory-compute strategy that the company asserts significantly reduces reliance on DRAM, which is currently a key limitation on inference expenses.

      Fractile contends that its chip can execute cutting-edge models up to 100 times faster and at a tenth of the cost compared to current GPU configurations, with updated investor documents presenting a comparison of 25 times faster and one-tenth the cost. However, the main technical challenge remains whether these metrics hold true under production conditions, as the company has only shared simulation and small-scale results, not extensive benchmarks against deployed GPU clusters.

      Fractile's initial commercial chip is not anticipated to be available until 2027, a timeline the company has confirmed publicly. The $220 million funding is aimed at advancing the design through tape-out, building the software stack, and facilitating early customer integration rather than ramping up for full production.

      The funding round is timely for the customer aspect as Anthropic is reportedly in initial discussions to purchase Fractile chips upon their release. If this relationship solidifies, Fractile would become the fourth officially named compute supplier for Anthropic, joining Nvidia, Google’s TPUs, and Amazon's Trainium and Inferentia. While Anthropic has explored creating its own custom AI chips, its interest in Fractile suggests a continued multi-supplier strategy.

      Fractile is among a select group of European chip startups that argue the inference market is fundamentally different from training, thus presenting winning opportunities. Over the past year, TNW has tracked three such companies. The argument posits that while training demands the largest and most advanced systems, where Nvidia's CUDA advantage is strongest, inference—the workload that incurs most expenditures once a model is in place—benefits from specialized architectures optimized for throughput and energy efficiency rather than peak floating-point operations.

      The competitive landscape for this thesis is becoming increasingly crowded. Groq has delivered its language-processing units to various model providers and recently secured funding at a $6.9 billion valuation; Etched is developing silicon specifically for transformers; Cerebras and SambaNova have raised capital focused on the same workloads from different perspectives. Google is also assembling a four-partner inference-chip supply chain with Broadcom, MediaTek, and Marvell to rival Nvidia in the inference market. Fractile asserts that its in-memory architecture is superior in terms of cost-sensitive inference measured by watts per useful token.

      This funding round follows Fractile’s announcement in February of a £100 million ($132 million) initiative to expand its London and Bristol operations over three years, which includes establishing a new hardware engineering site in Bristol, and aligns with the broader UK sovereign AI initiative, which also led to partnerships involving BT, Nscale, and Nvidia in April.

      Founder and CEO Walter Goodwin, a PhD from the Oxford Robotics Institute and now in his late twenties, has been the public spokesperson for the company. The team has recruited engineers from Graphcore, Nvidia, and Imagination Technologies and is developing its software stack concurrently with the silicon. The next key milestones are expected to be tape-out and customer integration.