Fractile raises $720 million for AI inference chips, targets first shipments in 2027

The UK government said on November 11 that its Sovereign AI venture fund and the British Business Bank had joined British AI chip startup Fractile’s Series C. Fractile put the round at $720 million the same day. Balderton, another participant, disclosed a $7 billion post-money valuation.

Fractile CEO Walter Goodwin said the new capital would help the company ship its first chips and systems in 2027. Its design focuses on how fast data moves between memory and the circuits that perform the calculations needed to produce an answer.

Fractile puts memory directly above the logic chip

Inference is the stage when a trained model processes a request and generates an output. A language model builds its response one token at a time. To produce each token, the hardware reads model weights, the values learned during training, and uses information held from earlier in the request. Memory capacity describes how much data can fit. Bandwidth describes how much of that data can move in a given time.

CapitalG, an investor in the round, argued that large models processing long documents or conversations need both capacity and bandwidth. A common AI accelerator layout places high-bandwidth memory, or HBM, alongside the processor in the same package. Balderton says Fractile stacks dynamic memory, or DRAM, directly over the chip that does the computing. That changes the path data takes between the two parts.

For workloads held back by how quickly the processor can read data, changing that connection can affect how long an answer takes. The model and the length of the input also affect system performance. CapitalG and Balderton described the design in their investment announcements. Neither published independent measurements of a shipping product in those announcements.

The chip is one part of the system

Balderton says Fractile develops the chip, the system it goes into and the software that runs on it. Fractile is recruiting compiler engineers in London and Bristol. A compiler turns the operations a model needs into instructions the hardware can execute.

Founded in 2022, Fractile announced a $220 million Series B for chip and system development in May. In February, UK AI minister Kanishka Narayan said Fractile had committed to invest £100 million in the UK over three years. That plan covered expanding its London and Bristol sites, opening a hardware engineering facility and hiring people to develop its next systems.

The UK government and private investors joined the round

Companies House records show an October 12 filing for a resolution creating three classes of Series C preferred shares and share allotment filings in October and November. November 11 was the date the company and investors made the size and participants public.

Thrive Capital, Founders Fund, Lightspeed and Redpoint co-led the Series C. Goodwin also named CapitalG, BlackRock and Shopify Ventures among the participants. The Cabinet Office and other UK bodies described the total round as worth more than £500 million. They did not disclose how much each of the two public investors contributed.

The government said the investment is intended to strengthen the UK’s AI hardware industry. Sovereign AI invests directly in early- and growth-stage AI companies expanding in the UK. The British Business Bank invests through venture funds as well as directly in technology companies. A separate national hardware plan had outlined purchases of high-performance inference chips for the country’s research computing infrastructure. The November 11 announcement concerned financing for Fractile. It included no purchase order for the company’s chips and no volume commitment.

Fractile gave a year for its first shipments. The November 11 disclosures gave no shipping quarter, ordering date, price or initial volume.