Nvidia Backs iPronics in $125 Million Round as AI Data Centers Shift From Copper to Optical Switching

Nvidia has joined a $125 million financing round for iPronics, a silicon-photonics startup developing programmable optical switches for increasingly large artificial-intelligence data centers.

The Series B was co-led by Maverick Silicon and Light Street Capital and brings iPronics’ total funding to approximately $177 million. Other investors include Triatomic Capital, Bosch Ventures, Catalight Capital and the European Innovation Council Fund.

The investment targets a networking problem that becomes more difficult as AI clusters grow. Training and running frontier models increasingly requires tens of thousands—and eventually hundreds of thousands—of accelerators to operate as a coordinated system. Those processors can only work efficiently if enormous amounts of data move rapidly between them.

iPronics is developing optical circuit switching, or OCS, technology to address that problem. Unlike conventional packet switches that continuously examine and forward data electronically, an optical circuit switch can create direct light paths between parts of a computing cluster and reconfigure those paths as the topology changes.

That capability becomes valuable at extreme scale because hardware failures become frequent enough that infrastructure operators need to route around unavailable systems without stopping an entire training job. A reconfigurable optical layer can bypass failed devices or reshape connectivity for different workloads.

iPronics is also trying to shrink the physical footprint of OCS hardware substantially, allowing more switching capacity inside crowded AI racks. Space matters because modern AI racks are already dense with accelerators, power-delivery hardware and cooling equipment.

The company’s flagship product, iPronics ONE, is a rack-ready silicon-photonics optical circuit switch with integrated control, telemetry and APIs. The goal is to let infrastructure software dynamically alter optical connectivity between machines in real time.

AI clusters are not static systems. Training jobs can have different communication patterns from inference workloads. Machines can fail, new accelerators can be added and clusters may be divided among several customers. A programmable optical network can adapt instead of forcing every workload through one permanently wired architecture.

The broader transition is from copper toward optics. Copper remains essential inside computing systems, but distance, bandwidth and power constraints become more difficult as clusters scale. Moving more data optically can provide high bandwidth with lower electrical losses over longer distances.

Nvidia’s participation is strategically notable because the GPU maker increasingly invests across the infrastructure needed to make its accelerators useful. A GPU waiting on data is an expensive underutilized asset. Improving the network around the GPU can therefore improve the economics of Nvidia’s own hardware ecosystem.

iPronics says it is already scaling commercial deployment and has opened a Santa Clara office to expand its U.S. presence.

The company still faces competition from established networking vendors, hyperscalers with internal engineering teams and other heavily funded photonics startups. But the addressable problem is expanding quickly.

AI accelerators are becoming more powerful, clusters are becoming larger and rack densities are rising. Networking failures and power consumption therefore become more economically significant every generation.

That is why the iPronics round is about more than another networking startup. The AI infrastructure boom is pushing innovation out of the GPU and into every layer surrounding it.

Memory became a bottleneck. Power became a bottleneck. Cooling became a bottleneck. Networking is increasingly next—and as clusters move toward hundreds of thousands of accelerators, more of that network is likely to move from electrons to light.