Fractile raises $720M to build faster, cheaper chips for running AI

Fractile, a British startup building chips for running AI models, has raised a $720 million Series C backed by the UK’s Sovereign AI Fund, TechCrunch reported on November 11, 2026. The round, led by Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint, makes Fractile one of the best-funded startups building purpose-built AI chips, alongside companies such as Axelera, Etched and OLIX.

The money is meant to get the company’s first product out the door. In a LinkedIn post quoted by TechCrunch, CEO Walter Goodwin wrote that the funding will enable the startup to “ship [its] first chips and systems next year”. For developers, that means Fractile is still a pre-product company: the significance today is about where AI inference costs and speeds may head, not about hardware anyone can use yet.

What Fractile’s $720M Series C includes

According to TechCrunch, the UK government said the investment is about “ensuring the U.K. has a greater stake in the hardware that will underpin its future economy”. The government put the company’s valuation at “over £5 billion,” and TechCrunch notes that $7 billion has been reported both by the Financial Times and by a Reuters source. Private-market data firm Caplight also estimates Fractile’s valuation at $7 billion.

Item	Detail
Round	$720M Series C
Lead investors	Thrive Capital, Founders Fund, Lightspeed Venture Partners, Redpoint
Government backing	UK Sovereign AI Fund
Valuation	Over £5bn (UK government); $7bn reported by FT and a Reuters source
Total raised to date	$986.7M, per Caplight’s estimate
Headquarters / founded	London, 2022
Employees	About 105, per Caplight

Caplight lists further Series C participants, including Nebular, Bain Capital Ventures, the UK Department for Science, Innovation and Technology, Factorial Capital, DST Global, SV Angel and Entropy Industrial Capital.

The numbers do not all match

Sources differ on the exact size of the round. TechCrunch reports $720 million. Caplight records a “$774M round closed” on November 11 and, in an October 30 entry, $773 million in Series C preferred-share financing, citing Companies House filings that show £577 million allocated since September. The gap may reflect currency conversion, tranches or rounding, but the sources do not explain it, so the $720 million figure here is TechCrunch’s.

How Fractile’s chip design differs from rivals

Purpose-built AI chips are one of several approaches being explored to cut the cost of inference, which TechCrunch describes as using increasingly demanding AI models in production. Inference, for readers new to the term, is the work a model does when it answers a request, as opposed to training, when it learns from data.

Fractile’s main differentiator, in TechCrunch’s assessment, is its hardware design. After a pivot, the company’s approach now centres on DRAM, the most common type of RAM, stacked on top of each chip in a way the company hopes will get around the usual trade-off between memory and capacity. Caplight describes Fractile’s hardware in similar terms, as specialised inference hardware that speeds up large-scale model execution by integrating memory and compute on a single chip.

Earlier descriptions of the company framed the target as memory bandwidth, the rate at which data can move between memory and the processor. LinkedIn posts summarising Fractile’s previous round describe chips “designed to overcome memory bandwidth limitations”.

The company also works on more than the chip itself. TechCrunch reports a full-stack approach that includes a machine learning compiler, the software that translates a model into instructions for specific hardware, intended to maximise the speed at which large language models run on Fractile’s chips. On that basis, the company frames itself as a frontier AI company, and TechCrunch says the round was priced accordingly.

Why agentic AI raises the stakes for inference hardware

TechCrunch links Fractile’s bet to the rise of agentic AI, systems that carry out multi-step tasks on their own. The trade-off between memory and capacity has become more of a problem as companies try to run longer chains of tasks with large models without losing speed. As compute struggles to keep up, TechCrunch writes, hopes are high for AI chips, despite their limited availability to date.

The speed claim, and how strong it is

Fractile’s headline performance figure comes from the company, not from independent testing. LinkedIn posts summarising coverage of its earlier funding say the company claims its hardware can deliver speeds of up to 1,200 tokens per second. One post describes this as raising processing speeds from 40 tokens per second to about 1,200. A token is a small chunk of text, roughly a word or part of a word, so tokens per second is a common measure of how fast a model produces output.

The same posts say the company pitches this as enabling long sequential reasoning and multi-million-token workloads to be processed in days instead of months. None of the sources include an independent benchmark, and the TechCrunch report on the new round does not repeat the figure. Until chips ship and outside testers can run real models on them, the number is a design target.

Fractile’s funding path and the reported Anthropic deal

The Series C caps a fast year for the company. Caplight’s records show a seed round in July 2024, a grant in October 2024, early funding rounds in January and March 2026, and a Series B on May 13, 2026, before the November 11 Series C.

LinkedIn posts describe that Series B as $220 million, led by Accel, Factorial Funds and Founders Fund, at a valuation of around $1 billion. In August, posts relaying earlier reports said Fractile was in advanced talks to raise around $600 million at a $6.5 billion pre-money valuation. The final round came in larger, at $720 million per TechCrunch.

Those August posts also say the jump followed an initial deal worth about $250 million for Fractile to supply AI inference hardware to Anthropic, with potential for the relationship to expand. The posts are second-hand summaries, and TechCrunch’s report on the new round does not mention Anthropic, so the deal should be treated as reported rather than confirmed. Anthropic’s own financial moves are covered in Unicorn’s report on its IPO pricing.

The company has also been expanding. According to the LinkedIn posts, Fractile plans to invest £100 million in a hardware engineering facility and software testing lab in Bristol, and operates bases in San Francisco and Taiwan alongside its London headquarters. Caplight records the Bristol plan as a $134 million investment.

What the round means for developers building on AI APIs

Most developers never choose the chips their models run on; they pick an API and a price. But the hardware underneath shapes both. Inference cost and speed decide how many tokens a product can afford to generate, and how long a user waits for an agent to finish a multi-step task. TechCrunch frames purpose-built chips as one way the industry is trying to bring those costs down.

Several points matter for anyone planning around this:

• No product yet. Fractile’s first chips and systems are due next year, according to its CEO. Caplight likewise records plans for a first commercial chip in 2027.
• Performance is unproven. The 1,200 tokens-per-second figure is a company claim relayed in LinkedIn posts, without independent testing.
• Software matters as much as silicon. Fractile’s compiler work suggests it aims to run existing large language models well, rather than require new ones, but the sources do not say which models or frameworks it will support.
• A crowded field. TechCrunch names Axelera, Etched and OLIX as other top-funded players, and LinkedIn posts list Cerebras, Etched and Groq as competitors. Caplight’s closest comparables are d-Matrix, OLIX and Etched.

For context on how large cloud providers are approaching the same problem, see Unicorn’s report on Google’s two new TPU 8 systems for long-context work.

Risks and open questions

Goodwin himself sounds a cautious note: in the LinkedIn post cited by TechCrunch, he wrote that “there is still a lot of work ahead” for the company. Building chips is slow and capital-intensive, and Fractile has already pivoted its approach once, according to TechCrunch.

Open questions the sources leave unanswered include:

• How the stacked-DRAM design performs on real workloads compared with existing hardware.
• Who the first customers will be, and whether the reported Anthropic agreement is confirmed or expanded.
• Whether chips will be sold as hardware, offered through cloud partners, or both. Caplight lists the business model as B2B hardware sales.
• Why the reported round size differs between TechCrunch and Caplight’s filing-based data.

There is also a policy dimension. The UK government’s direct involvement through the Sovereign AI Fund ties Fractile to a national goal of owning more of the AI hardware stack, which may bring extra scrutiny over where chips are made and sold.

What to watch next

The next milestone is the one Goodwin has set: shipping first chips and systems next year. Signs to watch for are named customers, independent performance numbers on real models, details of the compiler and the models it supports, and progress on the Bristol engineering and testing facility. Whether Fractile turns its funding into hardware that measurably lowers the cost of running large models, especially for long agentic tasks, will decide whether the valuation holds.