Featured
SEMICONDUCTORS • This year
Gates Frontier Led a $110 Million Bet Against Silicon
On January 22, an Austin startup called Neurophos closed an oversubscribed $110 million Series A led by Bill Gates’ venture arm, with Microsoft’s M12, Aramco Ventures and Bosch Ventures alongside. Total funding reached $118 million.
Worth being precise about what happened, because the shorthand gets it wrong. Gates Frontier led the round rather than writing the whole cheque, and the deal was announced publicly, covered at the time by the trade and technology press. What makes it interesting is not secrecy. It is the direction of the bet.
Neurophos is trying to compute with light instead of electricity. Its optical processing unit packs more than a million micron-scale optical elements onto a single chip, using metamaterial modulators the company says are 10,000 times smaller than conventional photonic elements — the miniaturisation that would make large-scale photonic computing manufacturable at data-centre scale for the first time. Its chief executive describes a single on-chip photonic sensor measuring 1,000 by 1,000, roughly fifteen times the 256-by-256 matrix used in most AI accelerators.

Funding raised by photonic AI chip startups in 2026.
The reason capital is moving is the constraint we wrote about on Tuesday from the other end. Traditional silicon is hitting a thermal wall: performance now costs heat and electricity in quantities that data centres struggle to supply. Light produces far less heat, travels faster and is less sensitive to temperature. The catch has always been manufacturing — optical components are physically large and awkward to mass-produce, and they need converters to move data between digital and analogue.
And this is a field, not a single company. AI chip startups have taken $8.3 billion globally this year. Two weeks ago a London firm raised $312 million at a $3.3 billion valuation on a photonic interconnect design — nearly three times the Neurophos round, seven months later.
The sober part belongs in the same paragraph as the enthusiasm. Neurophos has no shipping product, no revenue and no customers running production workloads. First commercial systems and a manufacturing ramp are targeted for 2028. What it has is a claim and a roster of backers betting the claim survives contact with a foundry.
Here’s why it lands on your desk: for three years the way to own the AI build-out has been to own the incumbent chipmaker. The people with the best view of what comes after are now funding the thing designed to replace it — and they are doing it at the stage where the companies are still private.
Source: Company announcements / TechCrunch / Dealroom via CNBC