Fractile

Market position

Fractile is one of a group of startups challenging Nvidia's dominance in AI chips, alongside Groq, Mythic, Rain AI, Cerebras and Graphcore, and in competition with AMD and in-house silicon from Microsoft, Google and AWS. It is frequently profiled as a leading UK entrant in the AI inference chip race, appearing in EE Times' Silicon 100 2025 and Sifted lists of European hardware and future-of-compute startups to watch. Coverage notes structural risks for specialised AI chip startups, including long design-to-production cycles and the entrenchment of Nvidia's CUDA developer ecosystem.

The company's stated differentiator is combining memory and processing into a single component, going beyond approaches that merely move memory closer to compute, in order to deliver low latency and high throughput at the same time and at lower power. A second stated differentiator is a purpose-built, simpler software stack developed alongside the hardware, intended to avoid the overhead that CUDA incurs compensating for GPUs' general-purpose design.

Technology

Fractile's processors physically interleave memory and compute, an in-memory computing approach in which arithmetic is performed directly where data is stored, next to the transistors doing the computation. The company contrasts this with GPUs that rely on off-chip DRAM, and with Groq's use of on-chip SRAM co-located with processing, saying Fractile goes further by combining memory and processing into a single component. Fractile also builds its own software stack alongside the hardware, spanning transistor-level circuit design up to cloud inference server logic, and has worked on foundry process innovation and chip micro-architecture. Public claims include running LLMs up to 25x faster at one-tenth the cost (company website), 100x faster and 10x cheaper than Nvidia GPUs in simulation (2024), and a target of 20x better performance per watt than existing AI hardware.

Go-to-market

The company is pre-commercial, focused on delivering first chips and systems to customers. Public sources indicate direct engagement with large AI model providers: The Information reported, as cited by DCD, that Anthropic held discussions with Fractile about purchasing its inference chips when available in 2027. Fractile also maintains a high public profile through press coverage, conference appearances and policy engagement in the UK.

Operators of frontier AI model inference at scale, including generative AI model developers and data centre/cloud inference providers.

Geography

Headquartered in the UK with two offices in central Bristol and central London. As of the May 2026 funding announcement, Fractile was hiring in London and Bristol (UK), San Francisco (US) and Taipei (Taiwan). Its February 2026 expansion adds a new hardware engineering facility in Bristol.

History

Founded in 2022, Fractile spent about two years in stealth before emerging in July 2024 with $15 million in seed funding led by Oxford Science Enterprises, with Kindred Capital, the NATO innovation fund, Cocoa and Innovia Capital participating; at that point it employed 14 people and planned to reach 18 by the end of August 2024. In October 2024 Business Weekly reported ARIA awarded the company £5m, and noted senior hires from NVIDIA, Arm and Imagination. Former Intel CEO Pat Gelsinger disclosed an angel investment in January 2025, and former Arm and Acorn Computers executive Stan Boland had also previously invested. Through 2025 the company featured in coverage by CNBC, the Financial Times, Sifted, EE Times' Silicon 100 and others. In February 2026 Fractile announced a £100 million ($135m) three-year UK investment covering expansion of its London and Bristol sites and a new Bristol hardware engineering facility, an announcement used by the UK government to showcase its AI plans. In May 2026 the company raised a $220 million Series B.

Risks & controversies

Performance figures published by Fractile are company claims, and as of July 2024 were based on simulation rather than manufactured silicon; the company had not stated when chips would be in production. Reporting highlights general risks for AI chip startups: multi-year lead times from design to production mean AI workload requirements may shift before a chip ships, and Nvidia's CUDA ecosystem creates high switching costs for developers. The reported Anthropic discussions are sourced to a third-party report and are not confirmed sales.