A pre-revenue London chip startup just raised $312 million at a $3.3 billion valuation to attack HBM
The largest semiconductor venture round in Europe this year has gone to a company that has not yet taped out its first chip. Olix Computing, based in London, announced on August 3 that it had raised $312 million at a valuation of roughly $3.3 billion, according to SiliconANGLE — about triple the valuation it carried six months earlier.
The investor list is the most interesting part of the disclosure. SiliconANGLE reported that Arm Holdings and Netflix co-founder Reed Hastings participated, with other names undisclosed. RCR Tech's account of the round adds Hudson River Trading and a UK sovereign AI fund among the backers and names Fundomo as lead. The two outlets also label the round differently — SiliconANGLE calls it a Series C, RCR Tech and other trade coverage a Series B — a discrepancy the company has not publicly resolved.
The trajectory is steep by any standard. In February 2026, SiliconANGLE reported that Olix raised $220 million led by Hummingbird Ventures at a post-money valuation above $1 billion, with earlier investors including Plural, Vertex Ventures, LocalGlobe and Entrepreneurs First. Roughly six months and one funding round later, the paper value of the business has more than tripled without a commercial product in the market.
What investors are buying is an architectural bet rather than a performance benchmark. Olix is building what it calls an Optical Tensor Processing Unit, marketed around a chip named DX-1 and a rack-scale appliance called X-1. The design replaces high-bandwidth memory with on-chip SRAM and substitutes photonic interconnects for copper between chips, on the thesis — as summarised by RCR Tech — that AI infrastructure is no longer bottlenecked by raw arithmetic throughput but by memory bandwidth, interconnect latency and power.
The company is targeting inference specifically, and within inference the decode phase of large language model serving, where latency and memory bandwidth dominate. SiliconANGLE reported claims that DX-1 can run a 100-billion-parameter model at more than 10,000 tokens per second, and that multiple X-1 racks can be clustered to serve models in the 10-trillion-parameter range. None of these figures has been independently benchmarked, and the company has not published detailed specifications.
Timelines from the two accounts differ slightly. SiliconANGLE reported that shipments are expected in the first half of 2027. RCR Tech described a first tape-out targeted for late 2026, with customer pilots and rack deployments in the second half of 2027. Either way, the gap between the current valuation and first revenue is measured in quarters, not weeks.
RCR Tech was direct about the risk, characterising Olix as pre-product and pre-revenue and its valuation as highly aggressive by any reasonable measure. The publication noted that photonic manufacturing is notoriously difficult and that the company still has to clear tape-out, foundry lead times, packaging and full system integration while competing against incumbents with mature software ecosystems and far larger research budgets. Olix's counter-argument is that avoiding HBM and advanced packaging sidesteps the supply constraints throttling GPU deployments, and that its systems remain compatible with existing AI software stacks — historically the point at which alternative accelerators have failed to gain traction.
The round says something broader about where frontier-tech capital is flowing. Training-cluster economics have been funded at enormous scale for three years; the money is now chasing the cost and power profile of serving models rather than building them. Whether Olix is the vehicle for that thesis is unresolved, and the late-2026 tape-out is the first checkpoint at which the claims become testable rather than promissory.
