The minister spoke, the chip is still 18 months out
Joséphine Kant, who runs the UK's Sovereign AI unit, described Olix as completely rethinking what an inference chip should be from first principles. The AI minister, Kanishka Narayan, framed it more simply: the future of AI will be built on the chips that power models. Both statements were made about a product no customer has run. Olix says the DX-1 reaches initial customer access in the second half of 2027.
That gap is the story. A government fund has taken an equity position in a company founded in 2024, in a round valuing it at 3.3 billion dollars, on the strength of silicon that will not sit in a buyer's rack for roughly eighteen months. Nothing about that is improper, and venture investing works precisely this way. The trouble starts when the announcement is read as a supply event rather than a financing event.
The valuation tripled, the roadmap did not
Olix, which operated as Flux Computing until it rebranded in January 2026, raised 220 million dollars in its Series A and carried a valuation above a billion dollars. Six months later the Series B brought in 312 million dollars, roughly 290 million euros, at 3.3 billion. The company is led by James Dacombe, who is 25, and has just added Matt Briers, previously nine years as chief financial officer at Wise, with Nick McKeown of Stanford joining the board.
What did not move across those six months is the date. The DX-1 is an optical processor, using light rather than electricity to carry out the arithmetic behind model inference, and Olix claims more than 10,000 tokens per second per user with higher throughput per watt than general-purpose parts. Those are design targets on a part that has not been delivered. A valuation can triple in two quarters. A tape-out cannot.
Arm's cheque is the line worth reading
The investor list has been reported as a curiosity: Fundomo led, Hudson River Trading joined, Reed Hastings put in money, the state came along. The name carrying the most information is Arm. Arm licenses the instruction set underneath most of the world's mobile silicon and a growing share of its server silicon, and it does not make speculative consumer bets.
An optical interconnect that Arm sits close to is a standards question rather than a startup question. If the way the DX-1 moves data between chips becomes something Arm designs around, that shapes what a 2029 inference rack looks like and who can supply parts for it. Buyers who read only the valuation will miss that the lock-in conversation has already started, eighteen months before anyone can place an order.
Two calendars, and only one has your budget on it
Sovereignty announcements run on a political calendar and silicon runs on a fabrication calendar, and the two are rarely in step. The Sovereign AI unit is itself four months old, launched on 16 April 2026 with 500 million pounds, equity cheques capped at 20 million pounds per company, a million GPU hours of national supercomputing per startup and visas processed in a working day. Its first equity investment went to Callosum. Olix is early in a programme that has not yet produced a delivered product from anyone.
For an operator sizing inference capacity the practical consequence is narrow, and worth stating plainly. Nothing in this announcement changes what you can buy in 2026 or 2027. If a supplier, a consultant or a government communication implies a domestic option is close enough to wait for, the question that settles it is the shipment date. Here the answer is the second half of 2027 at the earliest, before volume, before pricing, and before anyone outside the company has benchmarked the part.
Read next: Nvidia Stopped Renting Its CPU Core From Arm | Nvidia's Real Lock-In Just Went Open Source



