OLIX Raises $312M Series B to Build the 'Token Factory' for AI Inference

Posted on 20th Aug 2026 06:05:16 in Artificial Intelligence, Machine Learning

Tagged as: AI inference, AI chips, semiconductors, startup funding

London-based OLIX Computing has closed one of the largest funding rounds in the history of European AI hardware. The two-year-old startup announced a $312 million Series B on August 3, 2026, at a valuation of $3.3 billion — roughly tripling its worth since February, when it raised a $220 million Series A that valued the company at $1 billion. The capital will fund the delivery of OLIX's first chip, the DX-1 decode accelerator, to its initial customers by the second half of 2027, along with the manufacturing and supply chain commitments that shipping frontier inference hardware demands.

What makes the round noteworthy is not just its size. It is a bet on a specific, contrarian thesis: that AI inference — the process of running trained models to generate answers — has outgrown the general-purpose chips that currently do the work, and that a purpose-built "production line" of specialised silicon and optical interconnects can deliver a step change in both performance and cost.

A $312 Million Vote of Confidence

The Series B drew a striking mix of investors. New participants include Arm, the British chip design giant whose architectures power most of the world's smartphones; Hudson River Trading, one of the most sophisticated quantitative trading firms on the planet; Fundomo; and angel investors such as Netflix co-founder Reed Hastings. Every existing institutional backer — Hummingbird Ventures, Plural, Creandum, Crane, Phoenix Court and Transition — increased its commitment. Hummingbird, which led the Series A, is a London-based fund whose portfolio includes Kraken, Revolut and Deliveroo.

OLIX also strengthened its leadership in ways that signal commercial intent. Professor Nick McKeown, co-inventor of software-defined networking, OpenFlow and P4, has joined the board. McKeown is Professor Emeritus of Computer Science and Electrical Engineering at Stanford, a 2025 Marconi Prize winner, and a repeat company-builder: he co-founded Nicira (acquired by VMware) and Barefoot Networks (acquired by Intel), where he led Intel's networking business. As CFO, OLIX appointed Matt Briers, who spent nine years as CFO of Wise, steering the fintech from a loss-making startup to a 2021 direct listing on the London Stock Exchange that valued the company at roughly $12 billion — the first such listing by a technology company in London.

Founder James Dacombe, now 25, has built a track record that investors clearly trust. He founded the company as Flux Computing in March 2024 and rebranded it to OLIX in January 2026 after the Series A. He also serves as CEO of CoMind, a brain-monitoring startup he launched as a teenager that has separately raised about $100 million. Across both companies he has assembled roughly $350 million in funding — an unusual record at any age, and an extraordinary one at 25.

The Token Factory: Rethinking How Inference Hardware Is Built

OLIX's core argument is a manufacturing analogy. A data centre, the company says, is a factory whose product is the token — the smallest unit of AI output. Producing a single token takes hundreds of operations, each placing different demands on hardware: some stages read enormous amounts of memory, others burn compute on vector arithmetic, others shuttle results between chips. Any conventional factory would give each stage a machine purpose-built for it. The AI industry instead runs every stage on the same general-purpose GPU, and each new chip generation tries to make that generalist a little better.

That compromise, OLIX argues, is why tokens remain scarce and expensive, and why high throughput and low latency are so hard to achieve simultaneously. The company's X-1 platform inverts the approach: models are fully "unrolled" across a large number of specialised chips, so each chip concentrates on a single part of the model, like a workstation on an assembly line. Because AI architectures are still evolving rapidly, each chip keeps a flexible compute fabric rather than hard-coding a specific model's design. A fully deterministic compiler schedules workloads across entire racks.

This systems-level thesis has backing from veterans of internet infrastructure. Jonathan Heiliger, general partner at Vertex Ventures and a former Facebook infrastructure executive who helped launch Meta's Open Compute Project, has described the GPU status quo as forcing "a compromise between speed and cost," and called the demand for ground-up rearchitecture "brutally hard" work that Dacombe's team is executing faster than companies with ten times its resources.

DX-1: The Decode Accelerator

The first product of the X-1 platform is DX-1, a decode accelerator aimed at the stage where a model reasons and generates its output — the most latency-sensitive part of serving a frontier model. For models around 100 billion parameters, OLIX claims DX-1 can deliver more than 10,000 tokens per second per user, with higher output-token throughput per watt than general-purpose chips running large batch sizes. The architecture is designed to scale to models of 10 trillion parameters and beyond through a multi-rack scale-up domain architecture.

Two design choices set DX-1 apart. First, it holds the model in fast on-chip SRAM instead of off-chip high-bandwidth memory (HBM), trading capacity for dramatically better energy efficiency and latency. Second, it deliberately avoids two of the most supply-constrained components in the semiconductor industry: HBM and advanced packaging. In OLIX's telling, a startup cannot compete for scarce HBM capacity against hyperscalers, so a viable new architecture must be built from parts that are actually available. That choice, the company argues, lets it scale volumes even as the rest of the industry fights over allocations.

The Photonic Bet: Light Instead of Copper

Moving data between chips has always been the bottleneck that made a production-line approach impractical. OLIX's answer is a novel "slow and wide" optical interconnect that moves data directly between chips using light rather than copper, at what the company says is ultra-low latency and energy cost. This is enabled by rack-scale co-design across every part of the link — the chips, the lasers and the network — all of which OLIX designs itself. The company is hiring silicon, photonics, compiler and systems engineers across London, Bristol, Austin, Toronto and San Francisco to execute on it.

The photonics strategy places OLIX in a small but well-funded group of companies — including Neurophos, Lightmatter, OptoML, Rayd and Volantis — trying to make optical computing and interconnects commercially viable. Analysts at Jon Peddie Research note that the sector's open question is timing: whether these companies solve the problem the market actually cares about by the time their silicon arrives.

What It Means for the AI Infrastructure Race

No startup has yet loosened Nvidia's grip on the AI accelerator market, and OLIX faces genuine execution risk: DX-1 does not reach first customers until H2 2027, an eternity in a market moving this fast. But investor sentiment has shifted. Nvidia's recent partnership with Groq — which brought in one of the original founders of Google's AI chip program — has reignited interest in a category long written off as too capital-intensive for startups. Rising inference costs from AI agents and reasoning models, which consume vastly more compute per query than simple chatbots, have made cheaper inference a board-level priority at every major AI lab.

If OLIX delivers, the payoff would extend well beyond its own balance sheet. Specialised, SRAM-and-photonics-based inference racks that sidestep the HBM and advanced-packaging crunch could make frontier AI meaningfully cheaper to serve — changing the economics of every company that ships AI products. The token factory is still under construction, but after a $312 million round, a Stanford networking legend on the board and a Wise veteran on the finance team, it has the capital and the crew to try.

Sources

• OLIX — OLIX raises $312m at a $3.3bn valuation and appoints Professor Nick McKeown to its board
• Verdict via Yahoo Finance — UK chip startup OLIX lands $312m to scale frontier inference
• Jon Peddie Research — An AI processor by any other name
• Quantum Zeitgeist — OLIX Secures $312M Series B at $3.3 Billion Valuation
• OLIX — Building the infrastructure for frontier AI