The Photonics Gamble: Can iPronics Outpace the Hyperscalers?

In the race to scale AI, the bottleneck isn't just the chip; it's the plumbing. As Nvidia and other players push toward systems cramming hundreds or thousands of accelerators into single massive units, the industry is revisiting a concept reminiscent of the old telephone switchboard: Optical Circuit Switching (OCS).

As first reported by The Register, the current state of OCS is a mixed bag. Google has successfully utilized OCS in its TPU clusters for years, employing 2D and 3D torus topologies to manage some of the largest compute domains in the industry. This allows Google to virtually hot-swap failed accelerators or adjust pod sizes on demand. However, as The Register notes, traditional OCS appliances—often relying on micro-electromechanical systems (MEMS) and microscopic mirrors—are slow, with reconfiguration times typically around 100 ms, and physically bulky.

Enter iPronics. The startup is positioning its "second-gen" OCS as the solution to these hardware inefficiencies. By ditching MEMS and LCD-based systems in favor of a silicon photonics-based design with no moving parts, iPronics claims it can achieve sub-ms reconfiguration times. From a P&L perspective, the value proposition here is density and speed: iPronics says its technology allows for mid-workload topology changes by hiding latency during compute cycles.

Furthermore, iPronics is targeting a massive reduction in footprint. While existing 300-port appliances from Coherent generally take up eight or more rack units, iPronics contends that high-density connectors and multi-chips could enable it to house as many as 720 port pairs within a single rack unit. The company's current iPronics One chip supports 32 ports, with 72- and 144-port chips currently in development.

But is this a sustainable moat, or is it simply a bridge to a future where hyperscalers build this in-house? Nvidia seems to be hedging its bets. In March, Nvidia invested $6 billion—split as $2 billion each—into Marvell, Lumentum, and Coherent. Now, as reported by The Register, Nvidia has joined Light Street Capital and Maverick Silicon to invest an additional $125 million into iPronics.

***

**Opinion:** From where I sit, the risk is clear. OCS is most effective where network paths are relatively stable, which contrasts with the all-to-all connectivity prioritized in systems like AMD's Helios or Nvidia's NVL72. While iPronics' silicon photonics approach solves the "clunky hardware" problem, it doesn't necessarily solve the architectural tension. If the industry moves toward hybrid environments—using copper-based switched fabrics inside the rack and OCS-based meshes for rack-to-rack communication—iPronics has a window. But once the hyperscalers prove the silicon photonics model works at scale, the incentive to pay a startup premium vanishes. Nvidia isn't just investing in a product; they are investing in the R&D to ensure their massive LPU clusters don't hit a physical wall.