Olix raises $312M: your token price floor is a 2027 bet

A two-year-old British company just raised $312 million to build a chip that moves data between silicon with light instead of copper. You will never buy one. You will still pay a price shaped by whether it works — because inference silicon sets the floor under your token bill, and every credible new entrant is a bet on that floor dropping.

What actually happened

Olix announced a $312 million Series B at a $3.3 billion valuation on August 3, as reported by The Stack. Investors include Arm, Hudson River Trading, and Netflix co-founder Reed Hastings, with the UK government's Sovereign AI venture fund also participating — per Data Center Dynamics, it's the largest semiconductor venture round ever raised by a European company.

The architecture is the interesting part. Olix's thesis is that the binding constraint on inference isn't arithmetic throughput — it's memory bandwidth and the cost of shuffling data between chips. Its answer is to split the job: a dedicated decode accelerator for the reasoning-and-output stage, connected by optical die-to-die links rather than electrical ones. The company's own framing, quoted by The Stack: instead of one chip doing everything, build specialized silicon and "unroll models across racks like an assembly line."

The number that matters more than the valuation: first DX-1 systems are targeted for the second half of 2027. Olix has published throughput claims for the chip, but we haven't seen them independently verified, so treat them as marketing until someone runs a benchmark.

Why inference silicon matters for your business

Draw the chain. A fab makes the wafer. Nvidia, Etched, Cerebras, and now Olix turn it into inference capacity. Your model vendor rents that capacity and marks it up. Your per-token invoice is the last link, and it's the only one you see.

Two things follow, and neither of them is "buy a chip."

First, 2027 is not a plan. Between Olix's H2 2027 ship date, Etched's transformer-only ASIC, and AMD's 2GW Anthropic buildout landing in 2027, the industry keeps promising you cheaper tokens roughly eighteen months out. Budget for today's prices. Treat any drop as upside.

Second, you only capture a price drop if switching is cheap. When specialized silicon works, it shows up in your world as a new provider with better economics — not as a press release you act on. If moving from one model endpoint to another is a config change, that's a Tuesday. If it's a rewrite because prompt formats, tool schemas, and retry logic are welded into your application code, you'll stay on the expensive option out of inertia. We've made this argument about open weights and model aliases. It's the same discipline: one abstraction layer, and log cost per completed task, not cost per token, so you can actually tell when the floor moves.

Sources: The Stack, Data Center Dynamics.