James Dacombe

James Dacombe founded Olix in London in March 2024 to build hardware for running frontier AI models. The company was incorporated as Flux Corp and renamed Olix in January 2026. It is his second company: he also founded CoMind, a brain-monitoring startup he left school at 17 to build, as reported by Fortune.

A production line for tokens

Olix’s argument is that AI inference has outgrown the general-purpose chip. Producing each token takes hundreds of operations with different hardware needs, yet data centers run them all on the same processor. Its X-1 platform spreads a model across many specialized chips, each handling one part of the work, and connects them with an optical interconnect that moves data between chips using light rather than copper.

The first chip, DX-1, is a decode accelerator for the stage at which a model generates its output. It holds the model in on-chip SRAM and uses no high-bandwidth memory and no advanced packaging, the components the industry is shortest of. Olix says DX-1 can deliver more than 10,000 tokens per second per user on 100-billion-parameter models and aims to ship it to its first customers in the second half of 2027. It sells complete racks rather than individual chips.

Two rounds in six months

In February 2026 Olix raised $220m (£164m) in a round led by Hummingbird Ventures that valued it at more than $1bn, according to Tech Funding News. On 3 August it announced a $312m (£232m) Series B at a $3.3bn (£2.5bn) valuation, with Fundomo, Arm, Hudson River Trading and Netflix co-founder Reed Hastings joining and every existing investor increasing its commitment. The UK government took part through its Sovereign AI fund, as reported by The Next Web. Nick McKeown, the Stanford professor who co-invented software-defined networking, joined the board at the same time.

The company has more than 140 staff across the UK, US and Canada, according to The Next Web, and is hiring in London, Bristol, Austin, Toronto and San Francisco.

Why it matters to UK buyers

High-bandwidth memory and advanced packaging are the choke points in today’s AI hardware supply chain, and they keep inference expensive. A UK-designed system that avoids both could, if it performs as Olix claims, add a new source of inference capacity at a time when data center operators and cloud providers are competing for the same few suppliers.

Dacombe has raised $532m in six months on the strength of an architecture that has yet to ship. The DX-1 deliveries planned for 2027 will be the first real-world test of that design.