OLIX Raises $312M for Photonic AI Inference Chips

The race to build cheaper AI hardware has a new heavyweight bet, and it runs on light rather than electrons. On August 3, 2026, London-based OLIX announced a $312 million Series B at a $3.3 billion valuation, one of the largest funding rounds ever for a British semiconductor company and a sizeable wager that photonic computing can carve out a place in AI inference next to GPUs and custom silicon.

The round was led by Fundomo and brought in two conspicuous new names — chip-design firm Arm and quantitative trading house Hudson River Trading — alongside existing backers Hummingbird Ventures, Crane, Plural, Creandum, Phoenix Court, and Transition, each of which increased its commitment. Reed Hastings, the co-founder of Netflix, participated as an angel investor, and reports indicate a UK government-backed sovereign AI venture fund also took part.

What OLIX is building

OLIX is designing what it calls Optical Tensor Processing Units (OTPUs) — chips that perform the matrix math at the heart of AI models using light instead of electricity. In a conventional accelerator, data moves as electrical charge through transistors; in a photonic design, information is encoded in beams of light routed through on-chip waveguides. The pitch is that optics can move and multiply data with far less energy lost to heat, attacking two of the biggest constraints on AI infrastructure at once: power draw and interconnect bandwidth.

The company is aiming squarely at inference — the job of running a trained model to generate outputs — rather than the training phase where Nvidia dominates. That focus matters. Training is a periodic, capital-intensive event; inference is the recurring, high-volume workload that runs every time a user sends a prompt, and its cost increasingly determines whether an AI product is profitable. A chip that lowers the cost per token on inference addresses the part of the bill that never stops growing.

OLIX also makes a pointed architectural claim: its design is built to sidestep high-bandwidth memory, the expensive, supply-constrained stacked DRAM that has become a bottleneck for AI accelerators and a driver of the industry’s memory shortages. By combining photonic interconnects with what the company describes as deterministic rack-scale computing, OLIX says it can keep large models fed with data without leaning on the same HBM supply chain everyone else is fighting over.

The capital will fund development of the company’s first inference accelerator, the DX-1, with initial customer systems targeted for the second half of 2027. That timeline is the crux of the bet: revenue, and proof the physics works at production scale, are still more than a year out.

The people behind it

OLIX was founded in 2024 by James Dacombe, making the $3.3 billion valuation remarkable for a company barely two years old with no product yet in market. The round also came with a marquee technical hire: Nick McKeown, a Stanford professor and one of the pioneers of software-defined networking — whose earlier ventures reshaped how data centers move traffic — has joined the company to work on its optical interconnect stack. For a startup whose thesis rests on networking and data movement as much as raw compute, landing a figure of McKeown’s stature is a credibility signal aimed at both customers and future investors.

Arm’s presence on the cap table is its own tell. The Cambridge-based architecture licensor rarely takes direct startup stakes, and its involvement points to strategic interest in how inference silicon evolves beyond the GPU. Hudson River Trading, meanwhile, is the kind of latency-obsessed customer that would benefit early from faster, cheaper inference — the sort of backer whose money doubles as a design partnership.

A crowded field with a hard problem

OLIX is not the only company trying to unseat general-purpose GPUs on inference economics. A wave of startups has raised enormous sums pitching purpose-built inference silicon, from Etched’s roughly $20 billion valuation for its transformer-specialized ASIC to a broader cohort chasing custom accelerators. Most of those bets are still electronic; OLIX’s distinction is going photonic, a more ambitious and less proven path.

That ambition cuts both ways. Photonic computing has been “five years away” for a long time. The physics is real — optics genuinely can move data with less energy — but turning laboratory demonstrations into manufacturable, reliable, software-supported products has defeated well-funded efforts before. Analog and optical designs face thorny challenges around precision, calibration, and integration with the digital control logic and software stacks that real workloads require. And a chip is only as useful as the toolchain around it; Nvidia’s durable advantage is as much its software ecosystem as its hardware.

The valuation leaves little margin for error. $3.3 billion pre-product prices in a great deal of future success, and it lands amid an unusually generous funding climate for anything AI-adjacent — the same climate that produced Together AI’s $800 million raise and a record run of AI-infrastructure financing. Whether that enthusiasm reflects durable demand or a cycle that eventually cools is the question hanging over every large raise in the space.

What it means

OLIX’s $312 million round is a bet that the AI hardware market is big enough, and inference costs painful enough, to reward a genuinely different architecture — even one that has to prove its physics at scale before it can prove its economics.

Who wins if it works. Cloud providers and AI companies gain a second axis of competition against Nvidia’s GPUs, with the prospect of lower energy costs and freedom from the HBM bottleneck that constrains everyone today. The UK gets a flagship semiconductor champion at a moment when governments are treating domestic compute as sovereign infrastructure. And a photonics ecosystem that has struggled to reach commercial scale gets its best-funded shot yet.

Who’s exposed if it doesn’t. The obvious risk sits with OLIX’s investors: a $3.3 billion valuation on a pre-revenue company with a 2027 delivery target is priced for near-flawless execution, and photonic computing’s history is littered with promising demos that never shipped. More broadly, the raise is a data point in the debate over whether AI-hardware valuations have detached from what the underlying businesses can deliver.

What to watch. First, the DX-1’s real specifications when they arrive — energy per token and throughput on actual models, not laboratory benchmarks. Second, whether OLIX can build a software stack that makes the chips usable without heroic engineering from customers. Third, the 2027 delivery date: slippage is common in silicon, and a miss would test investor patience quickly. Fourth, whether the HBM-free claim holds up under production workloads, or whether memory reasserts itself as the binding constraint. The inference-chip field is filling with ambitious bets; OLIX has just made one of the boldest, and the most physically difficult, of them all.