3 SEP 2026 — Nvidia has joined Maverick Silicon and Light Street Capital in a US$125m round for iPronics, six months after putting US$6bn into Coherent, Lumentum and Marvell. Optical circuit switching is the technology packet switching replaced. It is coming back because AI training traffic looks nothing like the internet traffic packet switching was built for.
What was funded
The round closed on 2 September. iPronics builds optical circuit switches in silicon photonics with no moving parts, where earlier generations used microscopic mirrors, piezoelectric actuators or liquid crystal to physically redirect light.
Its second-generation design reconfigures in under a millisecond, against roughly 100 milliseconds for first-generation MEMS systems. The iPronics One chip carries 32 ports, with 72-port and 144-port parts in development, and the company projects up to 720 port pairs in a single rack unit. Coherent's 300-port appliances typically occupy eight or more rack units.
Nvidia's March 2026 photonics investments were US$2bn each into Coherent, Lumentum and Marvell. Google has run optical circuit switching in its TPU clusters for years, using 2D and 3D torus topologies to change pod size on demand and hot-swap failed accelerators.
Why the industry abandoned this and is returning to it
Circuit switching reserves a path between two endpoints for the duration of a conversation, which is how the telephone network worked. Packet switching replaced it because internet traffic is bursty and unpredictable. Holding a dedicated path open for millions of short flows to arbitrary destinations would waste most of the capacity.
AI training traffic has the opposite shape. A small number of enormous flows run between known endpoints for hours at a time, in patterns determined by the parallelism strategy rather than by anything a user does. There is very little to statistically multiplex, because the traffic is not statistical.
For that workload, the flexibility of packet switching is wasted overhead. Every hop through an electrical switch means converting light to electricity, making a forwarding decision and converting back, and paying power and latency for a decision that was the same as the last ten million.
The reconfiguration time is what makes it usable
The obvious objection to circuit switching is rigidity. A fixed path is efficient until the job changes, and a hundred milliseconds to rearrange the fabric means the topology is effectively static.
Sub-millisecond reconfiguration changes what the technology is for. At that speed a cluster can rearrange itself between phases of a training run, isolate a failed accelerator and re-form around it, or hand capacity between jobs, all without the reconfiguration being visible as a stall.
Google's TPU clusters are the existence proof, and Google built that capability itself over years rather than buying it. The startups and merchant silicon arriving now exist to sell the same thing to operators who do not want to repeat the engineering.
Nvidia is buying position in a layer it does not sell
Six billion dollars into three optical component makers in March, then a venture round in a switching startup in September, is a pattern rather than a set of opportunistic bets. Nvidia does not sell optical switches and has no obvious intention of starting.
Nvidia sells accelerators, and a customer can only usefully buy as many as their network can connect. A network that cannot keep the parts of a cluster fed makes the marginal accelerator worthless, which puts a ceiling on Nvidia's own market that has nothing to do with Nvidia's own products.
Investing across the optical supply chain relieves that ceiling and does something else at the same time. A company with positions in the components, the transceivers and now the switching gets to see the roadmaps of everyone building that layer, which is worth holding whether or not any single investment pays off.
Power is the reason this has become urgent
An optical switch that redirects light without converting it does not run the transceivers, serialisers and forwarding logic that an electrical switch does. In a facility where the binding constraint is the grid connection, every watt removed from the network is a watt available for accelerators.
This is the same power argument arriving from several directions at once. We reported that tape is being reconsidered because it consumes nothing at rest, and that operators are installing on-site gas turbines to bypass interconnection queues. Reducing demand, shifting it and generating it privately are three responses to one constraint.
What this means for the region
The optical layer is quietly becoming the contested part of AI infrastructure, and this is the third transaction in a week pointing at it. We reported that a private optical company is taking control of GoPro partly for its transceiver business, and that the rare-earth exposure that reaches a data centre runs through erbium in optical amplifiers.
For operators building in Johor, Batam or Thailand, none of this changes a procurement decision this year. Optical circuit switching is a cluster-scale technology that matters to whoever is designing the fabric, which for most regional facilities is the tenant rather than the landlord.
This does change the questions a facility operator should ask a prospective anchor tenant. A hyperscaler designing around optical switching has different power, cabling and rack-unit assumptions than one buying conventional leaf-spine, and a building specified for the second is not trivially suitable for the first.
View author profile → · Editorial policy
