Fractile, a U.K.-based AI chip startup trying to reduce the time it takes advanced models to answer complex queries, has raised $220 million in a Series B round backed by several prominent technology investors, adding fresh capital to the race for faster and cheaper AI inference hardware.

The company announced the financing on May 13, saying the round was led by Accel, Factorial Funds and Founders Fund. Other participants included Conviction, Gigascale, O1A, Felicis, Buckley Ventures and 8VC, with existing investors also taking part. Fractile said the money will help accelerate the path to getting its first chips and systems into customers’ hands.

The fundraise is notable because it targets one of the most urgent technical and economic problems in generative AI: inference. Training a large model remains enormously expensive, but once a model is built, the business depends on running it repeatedly for users. Every chatbot response, coding-agent step, document analysis, scientific query or automated workflow requires inference. As AI systems move from short answers to longer reasoning chains, the computing cost and delay attached to each output become central to whether products are practical at scale.

Fractile’s public pitch is built around that constraint. In its announcement, the company said the world’s most capable AI systems are increasingly limited by how long they take to produce useful outputs. It argued that frontier models are being pushed toward longer sequences, more intermediate reasoning and more computationally intensive workflows, making speed and unit economics decisive factors rather than secondary infrastructure concerns.

The company said existing chips tend to run advanced models at roughly 40 tokens per second in some demanding settings, while very long outputs can require much higher throughput to become commercially useful. Fractile framed the challenge in stark terms: compressing a workload that might otherwise take a month into a day would require output closer to about 1,200 tokens per second, while still handling large models and long contexts.

Fractile’s proposed answer is new inference hardware that rethinks the relationship between compute and memory. Data Center Dynamics reported that the company is developing chips using in-memory compute, an approach designed to let processors perform calculations directly in memory rather than constantly moving data between separate compute and memory components. That approach is intended to improve performance, reduce power consumption and lower the cost of running inference at scale.

The memory bottleneck is a central issue in AI hardware. Large language models move enormous amounts of data as they generate responses, and hardware performance is often constrained not only by raw arithmetic capability but by how efficiently information can be transferred, stored and reused. Fractile says it is working across the full stack, from AI research and foundry process work to chip microarchitecture, to break the trade-off between cost and latency.

The U.S. relevance of the deal is substantial even though Fractile is headquartered in London. Founders Fund, co-founded by Peter Thiel, is one of the best-known Silicon Valley venture firms. Felicis and 8VC are also closely associated with the U.S. technology investing ecosystem. Accel has deep roots in U.S. venture capital, even as it operates globally. Fractile also said it is hiring in San Francisco, along with London, Bristol and Taipei, placing the company’s expansion inside the same cross-border supply chain and talent market that defines much of the AI semiconductor sector.

For Silicon Valley, the deal adds another well-funded challenger to a market dominated by Nvidia’s GPUs and by large cloud providers building or buying custom AI accelerators. Startups are trying to compete by specializing for narrower workloads, especially inference, where customers may value lower latency, better power efficiency or lower per-query cost over general-purpose flexibility. The opportunity is large, but the bar is high: chip startups must design hardware, secure manufacturing capacity, build software tooling and convince customers to take a platform risk.

Fractile is entering that contest as demand for reasoning models and agentic AI systems grows. These systems often generate more tokens than earlier chatbots because they may plan, test, revise, call tools and perform multi-step tasks before returning a final answer. That makes inference cost more visible to AI labs, enterprise software vendors and cloud platforms. If a model’s output is slow or expensive, customers may use it less, limit the length of tasks or avoid deploying it in real-time products.

Data Center Dynamics reported that Fractile was founded in 2022 by Dr. Walter Goodwin, then a PhD student at the University of Oxford’s Robotics Institute. The outlet also reported that former Intel CEO Pat Gelsinger had previously invested in the startup and that Fractile announced plans in February to invest £100 million, or about $135 million, to expand U.K. operations over three years, including growth in London and Bristol and a new hardware engineering facility in Bristol.

The latest round follows wider investor interest in AI infrastructure beyond model developers themselves. Venture firms have continued to fund companies working on chips, data centers, networking, energy efficiency and AI deployment software, reflecting the view that the next stage of AI competition will depend as much on infrastructure economics as model quality. In that environment, inference hardware is seen as a strategic layer because it affects the cost of serving every user request.

Still, Fractile’s announcement is an early-stage technology milestone rather than proof of commercial adoption. The company has not disclosed broad customer deployments, production volumes or detailed performance benchmarks in the announcement. As with other AI chip startups, its prospects will depend on whether its architecture can be manufactured reliably, integrated with software frameworks and deployed in real customer environments against incumbent hardware and cloud-based alternatives.

The timing also reflects a shift in how AI companies discuss performance. During the first wave of large-model competition, much of the focus was on training scale, model parameters and benchmark scores. Now, as businesses try to turn AI into products and workflows, the emphasis is moving toward latency, throughput, power use and cost per output. Fractile is positioning itself directly inside that shift, arguing that advanced AI will be constrained unless the hardware running inference changes.

For U.S. startups and enterprise buyers, Fractile’s round is another signal that the AI hardware market remains open despite the strength of existing suppliers. If companies can lower inference costs or make long-running AI agents more responsive, they could influence how software firms price AI features, how cloud providers allocate capacity and how quickly enterprises adopt more ambitious AI workflows. The challenge for Fractile is to turn a well-funded thesis into hardware that customers can use at scale.