Olix Computing, a London-based startup developing optical inference hardware, raised $312 million at a $3.3 billion valuation — more than tripling its worth in months — as investors channel capital toward specialized chips designed to challenge Nvidia's GPU-centric hold on AI compute2,3. ANALYSIS The raise signals a market thesis that the inference era demands purpose-built silicon, not repurposed training accelerators, and that the window for challengers is open.
Why it matters
Nvidia's dominance of AI compute has been built on GPUs designed as general-purpose accelerators. ◆ As AI workloads shift from training toward inference — the phase where models generate outputs at scale — a growing cohort of investors is betting that specialized architectures can undercut GPUs on cost, latency, and energy per token. Olix raised $312 million at a $3.3 billion valuation, more than tripling its just-over-$1-billion valuation from a $220 million round in February. The round was led by New York VC Fundomo, with participation from Arm, Hudson River Trading, and Netflix co-founder Reed Hastings as an angel investor. Existing backers — including the UK government's Sovereign AI fund — all increased their commitments. The raise represents a concentrated bet that inference-specific hardware can carve out territory Nvidia currently holds by default.
The big picture
Olix's approach is architectural, not incremental. The company is developing an optical digital processor with a "novel memory and interconnect architecture" compatible with current AI models. Its flagship product, DX-1, is a decode accelerator designed for the stage when a model reasons and generates its output4. DX-1 uses no advanced packaging and no high-bandwidth memory; instead, it holds a model in fast on-chip SRAM for higher energy efficiency and lower latency. For models with 100 billion parameters, Olix claims DX-1 can deliver over 10,000 tokens per second per user. The company says the architecture scales to models of 10 trillion parameters and above.
Olix designs every part of its systems: the chips, the lasers, and the network that connects them. It plans to sell its chips as part of a server rack that includes software and networking equipment1. The company explained its philosophy in stark manufacturing terms: "A datacenter is a factory whose product is the token. Producing a single token takes hundreds of operations, each placing different demands on hardware. Any other factory would give each stage a machine built for it. Instead, the token factory runs every stage on the same general-purpose chip".
ANALYSIS That framing — the datacenter as a token factory running mismatched equipment — is a direct challenge to the GPU-for-everything paradigm. It positions Olix not as a faster GPU but as a different category of machine for a different stage of the pipeline.
Between the lines
Olix founder James Dacombe, who started the company in London in 2024 at age 23, told the Financial Times that "we are moving into a world of specialists". ◆ The investor roster reinforces that thesis from multiple angles. Arm's participation suggests the chip architecture giant sees optical inference as complementary to its own licensing model rather than competitive. Hudson River Trading, a quantitative trading firm, brings the perspective of a latency-obsessed compute buyer. Reed Hastings brings the credibility of someone who scaled one of the world's largest streaming inference workloads at Netflix.
The executive hires underscore a company preparing for commercialization, not just research. Former Wise CFO Matt Briers, who took Wise public, has joined as CFO. Nick McKeown — Stanford professor emeritus, winner of the 2025 Marconi Prize, co-founder of Nicira (acquired by VMware) and Barefoot Networks (acquired by Intel), and former head of Intel's networking division — has joined the board. ◆ McKeown's background in software-defined networking and programmable data planes aligns precisely with Olix's claim that it designs the full system stack, not just the chip.
Olix plans to deliver DX-1 to its first customers in the second half of 2027. The company is hiring across silicon, photonics, compiler, and systems engineering in London, Bristol, Austin, Toronto, and San Francisco. ◆ The geographic spread across five cities in three countries suggests Olix is building a distributed engineering organization to tap specialized talent pools in photonics and semiconductor design wherever they exist.
What's next
Olix has stated it will use the funding to bring its chips to market, build out the silicon platform, and fund manufacturing and supply chain commitments. The H2 2027 delivery target for DX-1 is the first hard milestone against which the company's claims can be tested. ◆ Until racks reach customer datacenters, the $3.3 billion valuation prices a thesis — that inference workloads will fragment away from GPUs toward purpose-built accelerators — rather than proven revenue. The speed of Olix's valuation tripling, from just over $1 billion in February to $3.3 billion now, reflects how urgently the market wants that thesis to be true.