London's OLIX Raises $312 Million to Put Light Inside the Inference Chip

The largest semiconductor venture round in European history has gone to a company that has not yet shipped a chip. OLIX Computing, founded in London two years ago by a 25-year-old Thiel fellow, announced on August 3 that it raised $312 million at a $3.3 billion valuation — roughly triple the price investors paid six months earlier. The entire bet rests on a claim that the bottleneck in AI inference is no longer arithmetic but the copper carrying data between chips, and that the fix is to replace it with light.

The round brought in chip designer Arm, the trading firm Hudson River Trading and the venture fund Fundomo as new backers, alongside angel money from Netflix co-founder Reed Hastings. Existing investors Hummingbird Ventures, Crane, Plural, Creandum, Phoenix Court and Transition all increased their commitments. The UK government's Sovereign AI venture fund, a £500 million vehicle launched in March, disclosed an investment as well. It follows a $220 million round in February led by Hummingbird that valued OLIX at $1 billion — meaning the company has tripled its price tag in under six months.

OLIX's argument begins with a metaphor its founder, James Dacombe, uses constantly: a data center running inference is a factory, and its product is the token. Producing a single token takes hundreds of operations, each placing different demands on hardware, and any other factory would give each stage a machine built for it. The token factory instead runs every stage on the same general-purpose chip, and each new generation of that chip is an even better generalist and never a specialist. OLIX's response is the X-1 platform, which unrolls a model across a large number of chips so each handles one part of the network — an assembly line rather than a soloist — linked by what the company calls a slow and wide optical interconnect that moves data directly from die to die in light rather than copper. Workloads are scheduled by a fully deterministic compiler, and OLIX says as many as 10,000 chips can be tied into a single multi-rack scale-up domain.

The first product is DX-1, a decode accelerator built for the stage where a model reasons and generates its output. On 100-billion-parameter models, OLIX claims DX-1 can deliver more than 10,000 tokens per second per user while simultaneously beating general-purpose chips on output tokens per watt — notable because GPUs generally reach their best efficiency only at large batch sizes, trading away per-user latency to get there. The architecture is designed to scale to models of 10 trillion parameters and above. DX-1 holds the model in fast on-chip SRAM and uses no high-bandwidth memory and no advanced packaging, the two components the industry is shortest of.

“Olix's mission is to build the infrastructure for frontier AI. Instead of forcing a single chip to do everything, we build specialized silicon and an optical network to connect it, unrolling models across racks like an assembly line,” Dacombe wrote in a LinkedIn post announcing the round. “A different answer to the industry's supply chain constraints, built on components available at scale today.”

The financing came with two hires that read as credibility purchases. Nick McKeown, co-inventor of software-defined networking, OpenFlow and P4, joined the board; he is a Stanford professor emeritus, the 2025 Marconi Prize winner, and a co-founder of Nicira and Barefoot Networks, sold to VMware and Intel respectively. Matt Briers, chief financial officer of Wise for nine years through its 2021 London direct listing at £8.75 billion, became CFO.

Why It Matters

McKeown's stated reason for joining is the cleanest articulation of the thesis anyone in this sector has offered.

“Inference is bounded by moving data, not by arithmetic,” he wrote. “Olix's approach stands out because they refuse to treat the chip, the interconnect, and the compiler as separate problems. Designing them together, with light doing what copper no longer can, is how you get a step change instead of just another increment.”

That is the whole photonics case in two sentences. Transistor density keeps improving; the energy cost of pushing a bit down a copper trace does not, and it gets worse as data rates climb. In a rack-scale inference system, a growing share of the power budget goes to electrical serialization and the optical modules at the edges rather than to the multiply-accumulate operations anyone is actually paying for. Move the light closer to the die and that overhead falls, which is why capital has poured into the category: Lightmatter has raised roughly $850 million at a $4.4 billion valuation, Ayar Labs closed a $500 million Series E in March that lifted it to about $3.75 billion, and Marvell acquired Celestial AI in December.

OLIX is attempting something more aggressive than any of them. Lightmatter and Ayar Labs sell optical I/O to whoever is building the accelerator. OLIX is building the accelerator, the lasers and the network, and selling complete racks — a full-stack wager against Nvidia rather than a component sold into its ecosystem. Skipping HBM entirely is the same instinct applied to procurement: if you cannot outbid hyperscalers for memory, design a chip that does not need it.

What To Watch

Everything here is still a claim. There is no shipping silicon, no named customer and no third-party benchmark; the 10,000 tokens per second figure is OLIX's own, on its own workload, for a chip due in the second half of 2027. That is the date to hold the company to. Before then, watch for a named launch partner, which would convert the thesis into a purchase order, and for any independent measurement of tokens per watt against a current-generation GPU rack. Watch, too, whether OLIX's hiring plan — from roughly 70 engineers in February toward 200-plus across London, Bristol, Austin, Toronto and San Francisco — actually lands. Photonic interconnect has been eighteen months away from changing everything for about a decade now. What is different in 2026 is that the power ceiling has stopped being theoretical.