OLIX

Overview

OLIX is a London company building inference accelerators that compute with light rather than electricity. It calls its architecture the Optical Tensor Processing Unit, or OTPU, and it is the most heavily capitalised European attempt to build an alternative to the GPU.

The company was founded in London in March 2024 by James Dacombe, then 24, and operated as Flux Computing before rebranding to OLIX in January 2026. Its rise has been unusually fast even by the standards of this market: a $220 million round in February 2026 at a $1 billion valuation, followed on 3 August 2026 by a $312 million Series B at $3.3 billion — reported as the largest semiconductor funding round in European history.

The Series B investor list is worth reading closely. Arm participated, which is a meaningful signal from a company with deep visibility into processor design. So did Hudson River Trading, a quantitative trading firm of the kind that buys low-latency compute for its own use, and the UK government's Sovereign AI venture fund, which places OLIX inside Britain's industrial strategy as well as its startup economy. Netflix co-founder Reed Hastings invested as an angel. Alongside the round, OLIX appointed Professor Nick McKeown to its board — a foundational figure in software-defined networking, OpenFlow and P4, whose expertise is in exactly the data-movement problems the company claims to solve.

The technology

Photonic computing has a long history of promising results and limited commercial deployment, so the specifics matter more than usual here.

The headline performance claim is that DX-1 can exceed 10,000 output tokens per second per user on 100-billion-parameter models, with better output-token throughput per watt than general-purpose chips running large batch sizes. The framing of that claim is important: per user and at low batch size is the regime where GPUs are least efficient, so it is a fair place for a specialist to compete — but it is also the comparison most favourable to OLIX.

These are company figures for a product that is not yet in general availability. No independent benchmark of OTPU silicon has been published. This publication has not verified them.

Role in inference

OLIX is attacking the same bottleneck as most credible inference specialists — the memory-bound, latency-sensitive decode stage — but from an unusual direction. Where Cerebras keeps data on an enormous piece of silicon, d-Matrix computes inside the memory, and Etched hard-wires the model architecture, OLIX changes the physical medium in which computation and communication happen.

If it works, the payoff is not a marginal improvement. Energy per operation and energy per bit moved are becoming the binding constraints on inference deployment as power availability, not chip supply, increasingly determines how much capacity can be built. A genuinely photonic accelerator would compete on precisely that axis.

The reason for caution is equally clear. Photonics has repeatedly proven harder to manufacture at yield and integrate into conventional systems than its proponents expected. Optical components are sensitive to temperature and alignment, the analogue nature of optical computation raises questions about numerical precision, and the surrounding software stack has to be built essentially from scratch.

Strengths and key questions

Funding history

OLIX is private and does not disclose revenue. Any revenue figure attributed to the company should be treated as unverified. Note also that OLIX and Fractile are both UK inference startups that raised approximately $220M rounds in 2026, and both avoid HBM. They are entirely separate companies with different architectures — OLIX computes optically, Fractile computes inside SRAM — and are easily conflated in secondary reporting.

Sources

1. OLIX — $312m Series B at a $3.3bn valuation and Nick McKeown board appointment. Primary company announcement.
2. Data Center Dynamics — OLIX raises $312m at $3.3bn, backed by UK government Sovereign AI fund. Investor detail and government participation.
3. New Electronics — Series B and Nick McKeown board appointment. Leadership and technology summary.
4. Tech Times — The photonic AI chip that ditches HBM (3 August 2026). Announcement date and memory architecture.
5. Converge Digest — OLIX raises $312M to build photonic inference platform. DX-1 and X-1 platform detail.

Performance figures on this page are company claims for pre-production silicon and are labelled as such. See the editorial methodology for how vendor claims are handled.

Related pages