Fractile raises $220 million for AI inference chips as Anthropic eyes UK silicon

Fractile's $220m round arrives as Anthropic eyes its UK silicon

Accel led the London chip startup's round, with Pat Gelsinger joining as an angel investor, weeks after Anthropic was reported to be in early discussions to become a customer. Fractile, the London-based startup designing inference chips that put compute and memory on the same die, has raised $220 million to take its hardware to production, the company said on Tuesday. The round closes above the $200 million reported target the company was understood to be sounding out in late March, as Electronics Weekly first noted, and lifts Fractile into the cohort of European chip companies pitching themselves as alternatives to Nvidia at the inference layer. The investor profile is what gives the round its weight. Accel is understood to have led, with former Intel chief executive Pat Gelsinger participating as an angel and operating adviser. Existing backers Kindred Capital, the NATO Innovation Fund, and Oxford Science Enterprises, which co-led Fractile's $15 million seed in July 2024, are part of the round. The technology argument runs against the prevailing architecture. Conventional AI accelerators, including Nvidia's H- and B-series GPUs, separate the compute die from high-bandwidth memory and pay an energy and latency tax shuttling data between them. Fractile's design instead performs the matrix multiplications that dominate transformer inference inside SRAM cells located alongside the compute logic, an in-memory-compute approach the company says removes most of the DRAM dependence that is currently the binding constraint on inference cost. Fractile claims the resulting chip can run frontier models up to 100 times faster and 10 times cheaper than current GPU setups; more recent investor materials, frame the comparison as 25 times faster at one-tenth the cost. Whether those numbers hold under production loads is the central technical question. The company has so far disclosed simulation and small-silicon results rather than at-scale benchmarks against deployed GPU clusters. F ractile's first commercial chip is not expected to be available until 2027, a timeline the company has reiterated publicly, and the $220 million is sized to take the design through tape-out, software-stack build, and early customer integration rather than full production ramp. The customer side is where the round arrives at the right moment. Anthropic is in early discussions to buy Fractile chips when they are available, multiple outlets reported earlier this month. If the relationship formalises, Fractile would become Anthropic's fourth named compute supplier alongside Nvidia, Google's TPUs, and Amazon's Trainium and Inferentia parts. Anthropic has separately been exploring building its own custom AI chips, but the Fractile track suggests it is still pursuing a multi-supplier hedge. Fractile is also part of a small group of European chip startups whose pitch is that the inference market is structurally distinct from training and therefore winnable. TNW has tracked three such companies across the past year. The argument is that training will continue to require the largest, most exotic systems and that Nvidia's CUDA moat is strongest there, while inference, the workload that actually consumes most of the dollars once a model is deployed, rewards specialised architectures tuned for throughput and energy per token rather than peak FLOPs. The competitive set on that thesis is becoming crowded. Groq has shipped its language-processing units to multiple model providers and recently raised at a $6.9 billion valuation; Etched is building transformer-specific silicon; Cerebras and SambaNova have raised against the same workload from different angles. Google itself is assembling a four-partner inference-chip supply chain with Broadcom, MediaTek, and Marvell to challenge Nvidia at the inference layer. Fractile's claim is that its in-memory architecture wins on the metric that matters most for cost-sensitive inference, watts per useful token. The round follows Fractile's February announcement of a £100 million ($132 million) three-year expansion of its London and Bristol operations, including a new hardware-engineering site in Bristol, and fits the wider UK sovereign-AI push that also produced the BT, Nscale, and Nvidia data-centre partnership in April. Founder and chief executive Walter Goodwin, an Oxford Robotics Institute PhD now in his late twenties, has been the public face of the pitch. The team has drawn engineers from Graphcore, Nvidia, and Imagination Technologies, and is building its software stack alongside the silicon. Tape-out and customer integration are the next visible milestones.

British inference chip startup Fractile bags $220M to accelerate token consumption - SiliconANGLE

British inference chip startup Fractile bags $220M to accelerate token consumption U.K.-based artificial intelligence inference chip startup Fractile Ltd. said today it has closed on a $220 million Series B round of funding. The round was co-led by Accel, Factorial Funds and Founders Fund, and saw the participation of Conviction, Gigascale, O1A, Felicis, Buckley Ventures, 8VC and existing backers. The company was founded in 2022 by the Oxford University-trained chip engineer Walter Goodwin (pictured), who serves as its Chief Executive. He helped to design Fractile's specialized inference chips, targeting workloads powered by trained AI models. According to Goodwin, the company is targeting what it believes is the key constraint facing the world's most advanced frontier models, namely the time it takes for them to generate outputs after being prompted. As AI models grow in size and become increasingly sophisticated, they now require tens of millions of "tokens," which are the basic unit of measurement in advanced computing, to solve hard problems or complete assigned tasks. However, these tokens require lots of data to be moved between the processor and its associated memory, which increases the latency, and therefore the response time to queries. To get around this, Fractile has developed a novel logic chip complete with an architecture that attaches memory to fit inside a standard server rack. The design helps to reduce latency and maximize bandwidth without sacrificing speed, the company said. While Fractile has kept the technical specifications of how its chips work close to its chest, Goodwin told the Wall Street Journal that it doesn't use traditional high-bandwidth memory, nor on-chip static random-access memory or SRAM. That suggests it's based on an entirely novel design. According to Goodwin, Fractile's chips can help to dramatically accelerate AI workloads, and also enable entirely new ones that aren't possible on standard graphics processing units, which are the most widely used AI chips today. "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," Goodwin wrote in a blog post announcing the round. "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, in any field where humanity will benefit from sheer intellectual work to resolve complex problems." Such tantalizing claims will be put to the test in due course, and Fractile will need to show that it can live up to them if it wants to stand out in an increasingly competitive market for specialized inference chips. In recent years, a host of inference chip specialists have emerged, hoping to take market share away from Nvidia Corp., the AI chip market leader. Those rivals include Cerebras Systems Inc. and its dinner plate-sized WSE-3 chips, which also target AI production workloads. Cerebras is set to go public tomorrow via an initial public offering that will raise at least $5.5 billion, the industry's biggest public debut in years. Other rivals include SambaNova Systems Inc., which recently cemented a partnership with Intel Corp., and startups such as Untether AI Inc. and Graphcore Inc. It's also going up against Nvidia itself, which recently launched the Groq 3 language processing unit for inference workloads, and public cloud giants like Amazon Web Services Inc. and Google Cloud, which also have dedicated inference chips. OpenAI Group PBC could be another rival, as it's currently believed to be working with Broadcom Inc. and Taiwan Semiconductor Manufacturing Co. Ltd. on its own inference chip design.