Where the fabric is going.

Optical circuit switching became a merchant market in 2026. The next slot opens in 2028, and the switches shipping today are too slow for it. Here is the record, with dates.

Two years, in order.

  1. Sep 2025

    MixNet, at SIGCOMM, reconfigures an optical fabric at runtime for mixture-of-experts training on 32 GPUs at about half the network cost of a non-blocking fat tree. ACM

  2. Nov 2025

    ACTINA, from Columbia and Nvidia at SC25, shows that reconfiguring inside a workload beats reconfiguring between jobs by up to 2.3 times, but only when the switch reconfigures in under 100 microseconds. ACM

  3. Feb 2026

    Google's Ironwood pods, 9,216 chips each, are connected through its Apollo optical circuit switches. TrendForce

  4. Apr 2026

    The Open Compute Project publishes its optical circuit switching white paper, backed by Microsoft, Nvidia, Coherent, Polatis, nEye, and Oriole. Its technology table scores mirror-based switches in milliseconds and names microseconds as the frontier. nEye raises $80 million. UALink publishes its 2.0 specifications. OCP

  5. Jun 2026

    Lawrence Berkeley National Laboratory's 2025 update projects US data centers at 649 TWh, 11.8% of national electricity, by 2030, with networking already 4.5% of data center electricity. LBNL

  6. Jul 2026

    Cignal AI raises its forecast to more than $8 billion a year of optical circuit switching by 2030 and reports that Nvidia will integrate optical circuit switches into its NVL576 architecture. Google has spent well over $1 billion on them. Cignal AI

  7. Aug 2026

    Nvidia's co-packaged optics switches enter full production at a fifth of the power of pluggable optics. Lumentum reports optical circuit switch shipments doubling quarter on quarter and guides to its first $100 million quarter. StorageReview, Semiconductor Today

  8. Sep 2026

    iPronics raises $125 million with Nvidia on the cap table for a sub-millisecond silicon-photonics switch. The Register

  9. Oct 2026

    Salience Labs raises $125 million with AMD and Intel, framing optical switching around "the token economics that AI workloads demand." SiliconANGLE

The microsecond question.

Mirror-based switches reconfigure in milliseconds. That is enough to rewire a cluster between jobs, which is how Google uses them, and it is the whole of today's market. The question that decides the next market is whether a fabric can reconfigure inside a job.

Two papers frame it honestly. MixNet shows that for some mixture-of-experts training, millisecond switching is enough if the reconfiguration is hidden behind compute. ACTINA shows that for in-workload steering of collective operations, the payoff appears only below 100 microseconds, and that it is worth up to 2.3 times in throughput and 1.75 times in tokens per joule when it does.

The workloads that cannot hide a reconfiguration are the ones growing fastest: steering bandwidth within a single training step, rerouting around a failed link in the middle of a collective, and resizing inference capacity as demand moves. These are what a twelve-microsecond switch is for. We do not claim every workload needs it.

The companies funded this year to chase the same slot use silicon photonics, which the OCP table scores as fast but low in port count and high in loss, or metasurfaces, which it scores as microsecond and large in port count. A diffractive free-space switch is in the second group, with the loss to bring down.

The fabric in 2028.

Racks at 600 kW

Nvidia's Rubin Ultra NVL576 is expected in the second half of 2027 at about 600 kW per liquid-cooled rack, five times today's. Every watt of network per rack is a watt of cooling and a liter of water the facility must provide.

Optics on the switch chip

Co-packaged optics shipped in volume in 2026, at a fifth of the power of pluggable transceivers. It cuts the energy of the electrical hop; it does not remove the hop. An optical circuit switch removes it.

Optical scale-up

Scale-out fabrics went optical first. The scale-up fabric inside the rack is next, and it is where reconfiguration time matters most, because the collectives that run there are measured in microseconds.

What it means for us.

  1. Lead with the threshold, not the multiplier

    ACTINA's 100 microseconds is a published line. We are on the right side of it by a factor of eight. Mirror-based switches are on the wrong side by a factor of a hundred.

  2. Publish loss, port count, and crosstalk next to speed

    The first objection from any network architect is insertion loss, because it limits an optical switch to one or two hops. Our measured and designed figures are on the technology page.

  3. Name the 2028 slot

    The scale-up optical circuit switch inside the next rack generation is the market entry. The incumbents are too slow for it by their own datasheets, and they could license our method to catch up.

  4. Speak the operator's language

    Tokens per joule and computing per megawatt are what a grid-constrained buyer measures. Our energy page is written in those units.

  5. Finish the chip

    Everything above depends on the phase-modulating mirror array that brings loss from 16 dB to 3 dB. That is where the next dollar goes.