Nvidia joined a $125m round for the switch that makes GPUs wait less
iPronics, a Valencia silicon-photonics spinout, raised a Series B co-led by Maverick Silicon and Light Street, with Nvidia in the syndicate and two interconnect veterans on the board.
iPronics, a silicon-photonics spinout from the Universitat Politècnica de València, announced $125m in Series B funding on 2 September, co-led by Maverick Silicon and Light Street Capital. Nvidia joined the round. That takes the seven-year-old company's total funding to $177m.
Seven of every ten dollars it has ever raised arrived in one week.
Note the verb. Nvidia "joined"; it did not lead. Reuters' Stephen Nellis put it as "$125 million from Nvidia and others", which is how the market will remember it, and the release puts Nvidia third in its own sentence.
Spain's @iPronics has raised $125 million from Nvidia and others to further its optical circuit switches. These are the switches that Google helped make famous for re-routing around hardware failures in large clusters and that several startups are helping bring to the wider world. A couple of interesting notes about this one:
1. The EU has a ton of design talent that's finally getting some funding in the current chip boom - and a lot of it is happening in fields like this that are a bit niche but potentially hugely valuable in the AI overall of data center infrastructure.
2. It's also interesting that both iPronics and nEye, one of the other startups pursuing 2nd Gen OCS, are essentially academic spinouts. It's fascinating to see VC money pulling advances in silicon photonics integration from lab to tapeout to hyperscaler evals on a time scales measured in months.
The product is an optical circuit switch you can rack: iPronics ONE, a programmable optical layer that lets an AI cluster rewire its own connectivity in real time for training or inference. Google made the category famous by building its own to route around dead hardware in big clusters. Everyone else has been buying electrical switches and copper (which is cheap, well understood, and running out of headroom), and the company's own framing is that scale-up demand is "accelerating the transition from copper to optics".
Which is the part worth pausing on.
Optical switching does not sell more GPUs. It makes the ones already bolted in more useful, by cutting the number of idle accelerators waiting on a link. How keen would you be, selling accelerators, to fund the thing that makes each one go further?
Our read: Nvidia is buying cheap optionality on the layer where its pricing power is thinnest, and $125m across a syndicate of ten is about the cheapest way to hold a seat. Look at who arrived with the money. Geoffrey Tate, advising Light Street and joining the board, was the founding CEO of Rambus and took it public, then ran Flex Logix until Analog Devices bought it, and sits on the board of Ayar Labs. Young Sohn, now an adviser, was Samsung's chief strategy officer and CEO of Inphi, and sits on the boards of Arm and Cadence. That is not a growth board (nobody recruits a Rambus founder to help with hiring). That is a group of people who have sold optical and interconnect companies to larger companies before.
The obvious counter is that this is simply what a good Series B looks like in silicon: you hire the greybeards because tapeout is hard and hyperscaler procurement is harder. Fair, and probably the main effect.
But the two readings are not exclusive, and the second one is cheaper to test. Nellis makes a second point that deserves more attention than the funding number. Both iPronics and nEye, the other startup chasing second-generation optical circuit switching, are academic spinouts, and venture money is now dragging silicon photonics from lab to tapeout to hyperscaler evaluation on timescales measured in months. European chip design has spent roughly a decade being described as talent without capital. This is what it looks like when that stops.
What we would watch, and when. iPronics has opened a Santa Clara office and is hiring for customer deployments. If within twelve months there is a named hyperscaler deployment rather than another evaluation, the thesis is working. If instead the next headline is an acquisition, remember which two people joined the board.