Analysis
iPronics, a Spanish photonics startup, raised $125 million in a Series B round co-led by Maverick Silicon and Light Street Capital, with Nvidia joining as a new strategic investor, the company announced. The Register reported the round also drew participation from Triatomic Capital, Bosch Ventures, Catalight Capital, the European Innovation Council Fund, and several existing backers, taking iPronics's total funding to $177 million.
What iPronics Actually Builds
Founded in 2019 as a spinoff from the Universitat Politecnica de Valencia, iPronics builds optical circuit switches -- its flagship product, iPronics ONE, is a rack-mounted device that uses silicon photonics to physically reconfigure which servers talk to which other servers, in sub-millisecond time, without ever converting the signal from light to electricity. Conventional data-center networking does that conversion at every switch hop, which adds latency and burns power at a scale that matters enormously once a training cluster spans tens of thousands of GPUs.
“- Ayar Labs -- optical I/O chiplets that attach directly to GPUs and CPUs, backed by Nvidia, Intel, and AMD among others.”
That's the pitch to Nvidia specifically. As frontier labs push single training runs across 100,000-plus GPU clusters -- the scale Anthropic and OpenAI are now building toward with tens of billions of dollars in compute commitments Pulse has tracked all year -- the interconnect increasingly determines effective utilization more than the chips themselves do. A GPU cluster that spends 20% of its time waiting on network reconfiguration is running at an effective 20% discount on every dollar of Nvidia hardware inside it, exactly the kind of inefficiency a chip vendor has a direct financial interest in fixing upstream.
The Photonics-for-AI Field Is Getting Crowded
iPronics isn't alone chasing this bottleneck, and the approaches diverge:
- Lightmatter -- photonic compute and interconnect, raised at a multi-billion-dollar valuation, focused on optical interposers between chips.
- Ayar Labs -- optical I/O chiplets that attach directly to GPUs and CPUs, backed by Nvidia, Intel, and AMD among others.
- Celestial AI -- a "photonic fabric" aimed at disaggregating memory from compute across a data center using light.
iPronics's optical circuit switching is a narrower, more mechanical bet than some of these -- reconfiguring which servers connect, rather than replacing the electrical interconnect inside a chip package -- which may be exactly why it's cheaper to deploy into an existing data-center build and why the round closed at $125 million rather than the multi-billion valuations photonic-compute plays have drawn.
The unresolved question is manufacturing scale: silicon photonics has a long history of components that work beautifully in a lab and struggle to yield at the volumes hyperscale data centers actually need. iPronics's new capital is earmarked specifically for scaling manufacturing and deployments -- the company all but admitting that's the current constraint, not the underlying physics.
