Olix bets photonic inference chips dethrone GPUs: $312M at $3.3B before tape-out

Olix designs AI inference chips: its X-1 platform unrolls models across custom chips linked by optical interconnects that move data with light instead of copper, starting with the DX-1 decode accelerator built on fast on-chip SRAM rather than HBM or advanced packaging.

That AI inference needs specialized chips, not general GPUs: Olix bets photonic interconnects plus on-chip SRAM beat GPU+HBM economics for token generation.

$312M Series B at a $3.3B valuation announced August 2026 (Fundomo, Arm, Hudson River Trading, Reed Hastings and existing backers), following a $220M round in February 2026 at just over $1B.

James Dacombe dropped out of school at 16, founded neurotech company CoMind at 17 - which has raised about $100M and still lists him as CEO - and launched Olix, initially called Flux, in London in 2024 at age 23. His argument, as he told the Financial Times: 'The era of a single Nvidia chip handling every AI task was over.' AI workloads were becoming too demanding and varied for one general-purpose processor, and inference - generating model responses - was where compute cost would concentrate.

Olix attracted a deep backer list early - Hummingbird Ventures, Crane, Plural, Creandum, Phoenix Court and Transition - and raised $220M in February 2026 at just over a $1B valuation. In early August 2026 it closed a $312M Series B at $3.3B, roughly tripling its valuation in six months; Fundomo led, Arm, Hudson River Trading and Netflix co-founder Reed Hastings joined, and the UK government's Sovereign AI venture fund backed the company. Olix hired former Wise CFO Matt Briers, appointed Professor Nick McKeown to its board, and grew past 140 employees across London, Bristol, Toronto and Austin.

Still pre-revenue as of 2026-09-02: no chip has shipped - Olix plans to finish its tape-out later in 2026 and deliver the DX-1 to launch customers in the second half of 2027.

Inference demand is structural: as AI shifts from training to serving, inference cost concentrates, giving specialized silicon a growing market rather than a hypothetical one.

Supply-chain arbitrage: by avoiding HBM and advanced packaging - the scarce inputs bottlenecking the industry - the architecture promises a buildable product even in a constrained market.

Photonic interconnect is a credible moat: moving data between chips with light targets the scaling wall of copper interconnects as model deployments grow.

The founder brought proof of nerve: a Thiel-backed dropout with a $100M-raised neurotech company made a pre-product chip startup fundable at $3.3B.

The doubling of valuation from about $1B to $3.3B in six months, with Arm and the UK government joining, signals industrial rather than purely financial conviction.

Raising rounds before shipping works when investors can price the bottleneck: Olix sold a structural argument - inference demand grows while GPU and HBM supply is constrained - not a demo.

As of 2026-09-02 Olix has not shipped a chip. It employs more than 140 people across London, Bristol, Toronto and Austin, plans to complete its tape-out later in 2026, and targets delivery of the DX-1 to launch customers in the second half of 2027. It competes with other specialized-inference chip startups such as Etched and Fractile, and the open question is whether SRAM-based decode silicon with photonic interconnects can actually undercut Nvidia's incumbent economics once real workloads benchmark the hardware.