0

The London startup pairs photonic interconnects with on-chip SRAM to rethink AI inference

In sum – what we know:

• A two-part bet – Olix moves data between chips with a photonic interconnect while holding models in on-chip SRAM, rather than compute directly in light.
• Dodging the HBM crunch – By skipping high-bandwidth memory and advanced packaging, Olix sidesteps the supply-chain bottlenecks squeezing the rest of the industry.
• A European long shot – Olix is pre-product and richly valued, but a UK photonic chip company shipping at scale would be a rare hardware win for Europe.

New AI chip companies seem to have been popping up a fair bit, but few get as much financial attention as Olix. The London-based AI chip startup has raised a $312 million Series B, the largest semiconductor venture raise in Europe to date. The round values the company at $3.3 billion — up from roughly $1 billion just six months ago at its Series A in February.

New York-based Fundomo led the round, with participation from Arm, quant trading firm Hudson River Trading, Netflix co-founder Reed Hastings, and a UK Sovereign AI fund. For Arm and Hudson River Trading, backing Olix makes sense — both as a hedge against GPU constraints, and as a wider ecosystem bet. But the more interesting question is what Olix is actually building.

Who is Olix?

Olix is the work of James Dacombe, a British entrepreneur who left school at 16 before founding the company. It started life as Flux Corp Ltd in March 2024 and rebranded to Olix Computing Ltd in January 2026, shortly after its first major funding round. 

The company is headquartered in London with more than 140 employees spread across Bristol, Toronto, San Francisco, and Austin. Entrepreneurs First backed it early, placing Olix firmly in London’s deep-tech startup ecosystem rather than the traditional semiconductor world of industry veterans and university spinouts.

There’s a political dimension here too. Europe has historically been weak in cutting-edge semiconductors, and both the UK and EU have been pushing to build homegrown chip capacity rather than relying entirely on US and Asian suppliers. Olix fits neatly into that narrative — which probably didn’t hurt when a UK sovereign fund was deciding whether to write a check.

Photonic technology and inference architecture

Olix is focused entirely on AI inference rather than training. Its core product is the Optical Tensor Processing Unit, or OTPU, which uses integrated photonics to route and perform computations with light rather than relying solely on traditional electronic architectures.

The company’s underlying thesis is that AI infrastructure isn’t really bottlenecked by raw FLOPs anymore. The real constraints, Olix argues, are memory bandwidth, interconnect latency, and power — and there’s plenty of industry evidence backing that view, given how aggressively the likes of Nvidia and SK Hynix have been pushing HBM bandwidth and NVLink capacity to keep GPUs fed.

Olix’s answer is to sidestep the problem rather than scale through it. Instead of pairing its compute with high-bandwidth memory the way Nvidia-style accelerators do, the OTPU uses SRAM — faster than HBM, though more area-intensive. That’s a meaningful supply chain decision as much as an architectural one. HBM production is concentrated among a handful of memory vendors and has been a persistent constraint on GPU deployments, so a design that avoids it entirely dodges one of the industry’s tightest bottlenecks. Olix also uses a novel photonic interconnect fabric to link its chips, which the company claims lets clusters scale more efficiently than electrical interconnects for large language model workloads.

The net claim is that Olix systems will run AI models more efficiently and at lower cost than today’s GPU-HBM setups. That remains a claim — nobody outside the company has tested this hardware. But Olix says its architecture will remain compatible with existing AI models and software stacks, without requiring complete rewrites. If true, that removes one of the biggest barriers alternative accelerators face, since asking customers to re-port their entire stack has killed more than one promising chip startup.

Product roadmap and rack-scale integration

Olix isn’t selling standalone chips. It’s building complete rack-scale systems that integrate compute, photonics, memory, and networking — a full-stack approach the company treats as a differentiator, and one that echoes how Nvidia has shifted toward selling entire racks rather than individual GPUs. The products fall under the “X-1 platform” branding, built around a custom “DX-1” chip. Detailed specs haven’t been broadly disclosed yet.

The roadmap is aggressive. Series B funds will go toward completing tape-out of the first chip — targeted for later in 2026 — along with ramping manufacturing and locking in supply chain commitments. First customer pilots and rack deployments are scheduled for the second half of 2027. That’s roughly a three-year path from founding to shipping product, which is quick for any chip company, let alone one working with photonics.

Market risks and manufacturing challenges

It’s worth being clear about where Olix actually stands. The company is pre-product and pre-revenue, which makes a $3.3 billion valuation highly aggressive by any reasonable measure. Investors are backing a vision, not proven commercial hardware.

And the vision is genuinely hard to execute. Photonic chips are notoriously difficult to manufacture, yield, and package at scale — plenty of optical computing efforts have stalled at exactly this stage. Olix still has to navigate tape-out, foundry lead times, packaging, and full system integration, all of which are capital-intensive even with half a billion dollars in the bank.

Then there’s the competition. Olix will ultimately be selling against entrenched GPU providers with products already in the market, mature software ecosystems, and enormous R&D budgets. Nvidia isn’t standing still on inference, and neither are the other accelerator startups chasing the same cost-per-token argument. The window for Olix to prove its hardware is genuinely superior — before incumbents respond with their own answers — is short and expensive.

If the technology works and ships on schedule, Olix becomes a case study in photonic AI hardware and a rare European hardware success story. If it doesn’t, it becomes a very expensive lesson in how hard chips are.