Nvidia invests $125M in iPronics to scale AI clusters via optical switching
The funding targets second-generation silicon photonics to replace traditional packet switching in massive AI compute domains.
Nvidia, Maverick Silicon, and Light Street Capital have invested $125 million into iPronics to develop second-generation optical circuit switching (OCS) technology. The move signals a strategic shift toward "telephone switchboard-style" light switching to support the scaling of massive AI clusters.
Leveraging silicon photonics, iPronics is creating OCS systems that eliminate the moving parts found in traditional micro-mirror (MEMS) systems. This architectural change allows the company to claim sub-millisecond reconfiguration times. Currently, the iPronics One chip supports 32 ports, but the company is developing expanded 72- and 144-port chips. The long-term target is to achieve 720 port pairs per rack unit.
The shift from packet switching
As AI clusters scale from dozens to thousands of GPUs, traditional copper and packet-switched fabrics are hitting critical limits in power consumption, latency, and physical density. Unlike packet switching, which processes individual data packets, OCS reconfigures the network by switching entire light paths. This approach reduces the overhead associated with traditional networking and allows for denser, more flexible topologies.
While the industry is now pursuing "second-gen" OCS, the concept is not entirely new. Google has historically utilized OCS within its TPU clusters to manage pod sizes and virtually hot-swap failed accelerators without disrupting the broader system. Nvidia's current investment follows a broader pattern of aggressive spending on optical infrastructure; the company previously invested $6 billion—split as $2 billion each—into Coherent, Lumentum, and Marvell to advance optics.
Why it matters for AI scale
For the "massive compute domains" required to train trillion-parameter models, the network often becomes the primary bottleneck. The ability to reconfigure network topologies in sub-milliseconds allows data center operators to optimize the fabric for specific AI workloads mid-run. Furthermore, it simplifies the maintenance of these clusters by allowing failed hardware to be swapped with minimal disruption to the overall compute task.
What's next
Industry attention now turns to the deployment of these higher-port-count chips and whether silicon photonics can fully displace MEMS-based systems in production environments. As the demand for denser accelerator clusters grows, the success of these second-generation OCS technologies will likely determine how efficiently the next generation of AI models can be trained and deployed.