UK Startup OLIX Raises $312M to Scale Optical Interconnect AI Chips for 10-Trillion-Parameter Models

The artificial intelligence industry is currently constrained by a massive physical bottleneck: moving data between chips using copper wires. As frontier models scale toward 10 trillion parameters, the energy and latency costs of these traditional electrical connections are becoming prohibitive. Enter OLIX, a London-based semiconductor startup that just raised $312 million in Series B funding to commercialize a radically different architecture. The round, led by New York investment firm Fundomo, catapults the two-year-old company to a $3.3 billion valuation—more than tripling its worth in just six months.[1][8]

The funding round attracted a heavy-hitting roster of backers, including chip design giant Arm, quantitative trading firm Hudson River Trading, and Netflix co-founder Reed Hastings. The rapid influx of capital highlights a growing industry consensus that general-purpose graphics processing units (GPUs), while excellent for training AI models, may not be the most efficient hardware for running them at scale. OLIX is specifically targeting this "inference" phase, aiming to build rack-scale systems that treat data centers like specialized token-producing factories.[1][5][8]

At the heart of OLIX's pitch is a fundamental shift in how chips communicate. The company's X-1 platform replaces traditional copper connections with a "slow and wide" optical interconnect, moving data directly between chips using pulses of light. By utilizing silicon photonics, the architecture dramatically reduces the power consumption and latency associated with electrical signaling, which currently accounts for a massive portion of an AI cluster's energy budget.[3][4]

This optical approach allows OLIX to distribute a single massive AI model across hundreds of specialized chips without suffering the severe performance penalties that typically occur when data leaves a processor. Instead of forcing a single chip to handle every part of a model's computation, the X-1 platform fully unrolls the model into a continuous production line. Each chip in the sequence is dedicated to a specific stage of the token generation process, passing the data along the optical fabric at ultra-low latency.[3][5][8]

The first silicon component of this platform is the DX-1, a specialized "decode accelerator" designed specifically for the stage where an AI model reasons and generates its output. For models in the 100-billion-parameter class, OLIX claims the DX-1 can deliver over 10,000 tokens per second per user, operating at a significantly higher throughput-per-watt than general-purpose GPUs running large batch sizes. The company states that the architecture is designed to scale seamlessly to models containing 10 trillion parameters and beyond.[3][7]

Beyond performance claims, OLIX's architecture makes a strategic end-run around the most severe bottlenecks in the global semiconductor supply chain. Modern AI accelerators rely heavily on High-Bandwidth Memory (HBM) and advanced packaging techniques like CoWoS (Chip-on-Wafer-on-Substrate). These components are currently in critical shortage, dominating the manufacturing constraints of industry leaders and slowing the deployment of new data centers.[1][4]

OLIX completely bypasses these bottlenecks by abandoning HBM altogether. Instead, the DX-1 stores the model entirely in fast, on-chip Static Random-Access Memory (SRAM). Because the optical interconnect allows chips to share data so efficiently, the system can pool the SRAM across multiple processors, eliminating the need for complex advanced packaging. This design choice not only reduces energy consumption but theoretically allows OLIX to scale manufacturing volumes without waiting in line for constrained supply chain components.[1][4][5]

To manage this distributed architecture, OLIX relies on a fully deterministic compiler. Unlike traditional systems that dynamically schedule workloads on the fly—introducing unpredictable delays—the OLIX compiler maps out the exact flow of data across the entire rack ahead of time. This rack-scale hardware and software co-design ensures that the optical links are utilized with maximum efficiency, preventing the data traffic jams that plague conventional AI clusters.[3][4]

Despite the massive valuation and technical promise, OLIX remains a pre-revenue company that has yet to ship a physical product. The startup, founded by 25-year-old James Dacombe, is currently operating on architectural simulations and component testing. The new $312 million injection is designed to fund the incredibly expensive process of finalizing the silicon design—known as "taping out"—which is scheduled for later in 2026.[1][6]

If the tape-out is successful, OLIX plans to deliver its first commercial DX-1 systems to customers in the second half of 2027. Until those systems are deployed in real-world data centers, the company's headline performance claims remain untested against live, unpredictable AI workloads. Industry analysts note that the gap between a funded architecture and reliable, manufacturable production hardware is notoriously difficult to cross.[2][6][7]

To navigate this transition from theoretical design to commercial manufacturing, OLIX is bringing in seasoned industry veterans. The company recently appointed Matt Briers, the former finance chief who guided UK fintech Wise through its massive scaling phase, as its new Chief Financial Officer. Briers is tasked with managing the complex capital requirements and supply chain commitments necessary to build custom silicon at scale.[1][3]

On the technical side, OLIX has added Stanford University professor Nick McKeown to its board of directors. McKeown is widely regarded as a pioneer in modern networking, having co-invented software-defined networking (SDN) and the P4 programming language before leading Intel's networking business. His involvement lends significant credibility to OLIX's optical interconnect strategy, signaling to the market that the networking architecture is grounded in proven principles.[3][4]

The broader market context for OLIX's raise is a mix of explosive demand and growing financial anxiety. While venture capital continues to pour into AI hardware startups—including recent massive rounds for European competitors like Axelera and Fractile—public markets have grown increasingly jittery. Investors are beginning to scrutinize the massive debt and capital expenditures funding the global AI infrastructure buildout, questioning when the software revenue will catch up to the hardware costs.[1][2]

This economic pressure is precisely why OLIX's focus on inference efficiency is resonating with investors. While training a frontier model is a massive one-time capital expense, running that model in production incurs continuous, compounding costs in electricity, memory bandwidth, and data center space. As AI agents become embedded in everyday software and handle increasingly complex reasoning tasks, the cost of inference threatens to outpace the economic value of the outputs.[5][8]

If OLIX can successfully manufacture its optical architecture and deliver on its efficiency claims, it could fundamentally alter the economics of artificial intelligence. By breaking the reliance on general-purpose GPUs and constrained memory supply chains, the company aims to make the deployment of 10-trillion-parameter models not just technically feasible, but commercially viable for a much broader range of applications.[3]