ECOC 2026 | 95% of OCS Shipments Are One Customer: Google. The Next Killer App Waits for GPUs to Spill Into a Second Rack

2 days ago

8 min read

Optical circuit switching (OCS) has genuinely graduated from the science-project stage, but right now it is not a "market" — it is a customer list, and there is only one name on it. Cignal AI's call at the ECOC 2026 Market Focus was blunt: roughly 95% of OCS ports shipping today go into Google's TPU network, and that will remain true for the next two to three years. What would turn OCS into an industry-scale business is GPU scale-up — and whether that happens is not in the OCS vendors' hands. The key is whether co-packaged optics (CPO) goes multi-wavelength.

1. OCS Isn't New — Google Dragged It Out of the Lab and Back Into the Data Hall

Start with the definition, because it sets every limit that follows. OCS does exactly one thing: it moves light from one fiber to another fiber untouched, with no optical-electrical-optical (O/E/O) conversion in between. It is agnostic to protocol, wavelength and data rate; multi-wavelength works, and so do analog signals. It switches the "whole spectrum," not packets.

Plenty of companies chased this idea in the 1990s, imagining telecom network reconfiguration and protection, plus automated patch panels. It didn't take off. OCS was sent back to the lab, became a component inside smart patch panels, and went quiet for twenty years.

The one that dragged it back into the data hall was Google — and Google used it to solve an AI-cluster problem, not a telecom one.

The upside of OCS is easy to grasp: low latency, rate- and protocol-agnostic (nothing to replace when the network is upgraded), low power, and fewer human patching errors. The limits are just as hard: no sub-wavelength switching, no O/E/O regeneration, and a large blast radius — when one OCS goes down, everything behind it goes down too. That last point will keep coming back, because it is exactly why reliability ranks first among all the metrics.

Cignal AI ECOC 2026 deck — "Optical Circuit Switching Defined" (the OCS definition and pros/cons slide). Caption: the definition of OCS, its three advantages (low latency / rate- and protocol-agnostic / low power) and three limitations (no sub-wavelength switching / no O/E/O regeneration / large blast radius). Source: Cignal AI, ECOC 2026 Market Focus

2. Four Metrics Decide Whether an OCS Actually Sells

The speaker boiled down what buyers really compare to four metrics — and each maps to a different application, not a one-size-fits-all spec sheet:

Reliability. In an AI cluster, a failed OCS takes a huge blast radius down with it, so every technology camp pitches reliability as its number-one selling point — which also means it is no longer a differentiator but a table stake.

Radix (number of switchable ports). This one is the most interesting, because bigger is not always better. Some applications need 500+ ports, others are fine with 64 — and the latter care far more about power and cost. The split in radix requirements is itself a signal that applications are diverging.

Optical loss. OCS is a "go out and come back" architecture, so the optical path is inherently doubled and every dB is expensive. Long-reach applications are especially sensitive; short-reach ones are more forgiving.

Switching speed. In the past this barely mattered — reconfiguring once a day or once every few hours, a bit slower was fine. That is changing: new applications are starting to demand millisecond or even nanosecond switching.

On technology, nearly all early OCS was MEMS (Google uses MEMS), and most of the patents have expired, so today "everyone has a MEMS." Competitors now also include liquid crystal, piezoelectric, silicon photonics and robotic fiber switches, each strong in its own niche. This is not a race one technology will win; different applications will each pick a different technology.

3. Application One: Replacing the Spine Switch — The Best Numbers, but Almost Impossible to Copy

This is the one Google disclosed first, and the one that got the whole industry excited. The packet switches at the top of a traditional Clos network are the biggest, most power-hungry and most expensive layer of the entire network; Google replaced them with OCS.

The reported numbers: 40% lower power, 30% lower cost, 30% better throughput. The speaker put it precisely — from that moment on, it was no longer the CTO looking at OCS, it was the CFO.

But the next line is the real point: almost nobody can replicate this application. There are three reasons.

First, Google spent roughly ten years on traffic analysis before it was sure what its network looked like. Second, the architecture only works when traffic is deterministic — A always going to B is fine; A needing to reach B, C and D breaks the whole scheme. Third, it was originally implemented with a patch-panel mindset, so fast switching was not even a requirement at first.

The deployment uses 256×256 OCS, and the industry still wants larger radix. Optical loss must be kept low (long optical path), but it still fits within an IM-DD budget — even at 200G/lane, Coherent Lite is not yet needed.

The conclusion of this section is clear: it is the most convincing OCS demonstration ever, and at the same time the business model with the least reference value.

4. Application Two: Scale-Out (TPU Network) — Where 95% of Today's Ports Go

This is the one actually shipping. The speaker described the architecture in detail: a 4×4×4 TPU cube (64 chips) forms a pod, and 64 pods form a superpod; every "face" of every cube connects to OCS, for a total of 48 128×128 OCS units per superpod.

In essence it is a "smart patch panel embedded deep in the AI architecture." It can rearrange topology based on resource availability, route around failed nodes and carve the cluster into separate partitions. Early TPUs used a 3D torus; the latest generation moved to an architecture Google calls Boardfly.

Two easily overlooked details:

First, it is tied to TPU, not GPU. Only TPU users can replicate it, and today only Google supplies TPUs. So when someone says "the OCS market is huge," first ask whether that market has only one buyer.

Second, switching speed is going from "doesn't matter" to "matters." Early versions were manual operations — a person pressed a button. Current versions are not. Once automation goes up, speed becomes a spec.

As for volume, the speaker was conservative and honest: of the OCS ports shipping today, roughly 95%, maybe even 99%, go into this Google application. And the next-generation TPU actually needs about as many OCS units as this one — the rumored "10x" does not hold. What really drives the total up is that data centers keep getting bigger.

The three OCS applications compared — spine replacement, scale-out (TPU) and scale-up (GPU): radix, switching-speed needs, key constraints and current status. Source: Simple Tech Trend | Data: Cignal AI, ECOC 2026 Market Focus

5. Application Three: Scale-Up (GPU) — The Next Killer App, but CPO Holds the Key

This is the real heart of the talk.

Single-rack scale-up networks today run on copper. Once you span multiple racks — NVL576-class, eight racks — copper can't keep up and optics become mandatory. And when a switching layer has to sit between racks, the speaker asked a very direct question: why can't that layer be OCS?

The fit is actually quite good:

Radix only needs to be 64×64 or 72×72 (72 GPUs per rack) — no huge port counts required
GPU cluster traffic is deterministic and high-volume — exactly the shape OCS handles best
Space and power inside the rack are scarce, which amplifies the size advantage of silicon photonics OCS
It is naturally meant to be deployed alongside CPO / NPO on the switch tray

The speaker's verdict: scale-up will be OCS's next killer application after the AI reconfiguration use case. The longer-term goal is to remove the switch tray altogether, but that requires fully optical racks and is many years away — and a two-tier OCS network may need Coherent Lite.

But there is a challenge here, and it is the single line from the talk most worth remembering for Taiwan's supply chain: NVIDIA's current CPO is DR-based — single wavelength, many fibers. Connecting DR to OCS is bulky and expensive, because OCS would need eight times as many ports as with FR.

Eight times. That is not an engineering inconvenience; it means the business model simply doesn't work.

There are only two ways out: move micro-ring-based CPO to multi-wavelength (the speaker's word was "non-trivial"), or switch to an NPO approach. Whether that's a short-term or long-term fix, he said himself it is TBD. This lines up directly with our earlier breakdown, CPO / NPO / XPO Panel: Not a Roadmap Fight, but the 409.6T Bottleneck 18 Months Out — the reason NPO is still alive has never been just "open ecosystem."

The 8x gap in OCS port count between FR (multi-wavelength, single fiber) and DR (single wavelength, multi-fiber), and the two possible ways out. Source: Simple Tech Trend | Data: Cignal AI, ECOC 2026 Market Focus

6. Risks and Counterarguments: Where This Story Could Still Break

The speaker laid out the counterarguments himself, which deserves respect.

First, every other application is still small. Campus DCI reconfiguration (Oracle's "normally A to B, A to C only for backup" scenario), hybrid network bypass, multi-layer protection, telecom — all real, but lay out the shipped port counts and together they are out of proportion. And these ports carry low ASPs, so sizing the market by port count badly overstates it; only by dollars do you see the truth.

Second, he is revising his own forecast. Two years ago Cignal AI's report said "only one opportunity: Google"; now it says "two major applications." OCS vendors still call him conservative and say the opportunity is far bigger than two. That back-and-forth is itself a signal: visibility into this market is far less solid than the shipment numbers suggest.

Third, the technical debt of fast switching hasn't been paid. Asked from the floor whether OCS needs to be faster than MEMS, the answer was that some new applications already demand nanosecond switching, but those applications connect coherent devices, which have to resynchronize after every switch. Until synchronization is solved, faster switching is moot.

Fourth, the timeline is fuzzy. Asked how GPUs will connect to OCS, the speaker said scale-up has to go optical and in-rack optics "will almost certainly be CPO," so CPO's evolution and OCS scale-up are locked together. As for optics directly on the chip, he called that five to six years out — too far to discuss now.

7. Conclusion

Stack the three applications together and the real shape of OCS is this: a proven technology that, for now, has only one buyer.

For Taiwan's supply chain, three things belong on the watch list right now:

First, don't treat OCS port count as a proxy for optical-module demand — it is optical-module demand. Cignal AI forecasts more than 40 million OCS ports by 2030, and behind every OCS port is an optical interface that has to be lit. That number matters more to makers of optical transceivers, connectors and fiber array units (FAU) than to switch vendors.

Second, the scale-up OCS opportunity is really a silicon photonics OCS opportunity. A radix of just 64×64 / 72×72, space and power first, coexisting with CPO on the switch tray — that set of conditions cancels out MEMS's advantage (large radix) and pushes silicon photonics forward. Whoever is building silicon photonic optical switches belongs on the list now.

Third, and most important: the 8x between FR and DR is the pivot of the whole story. If CPO ends up multi-wavelength, scale-up OCS works and NPO's strategic room shrinks; if CPO stays stuck on DR, NPO gets a second chance and OCS's second killer app waits another generation. This affects not just switch vendors but the upstream CW lasers, wavelength-multiplexing components and external laser source modules (ELSFP) — who is betting on multi-wavelength and who is betting on single-wavelength multi-fiber is exactly where the supply chain forks. For context, see CPO Took a Decade Because It Wasn't a Real Problem Yet: How NVIDIA Pushed Co-Packaged Optics Into Mass Production and 2026 AI Infrastructure Must-Read: The Full Optical Communications and CPO Supply Chain Map.

Verdict: OCS is no longer crying wolf, but for now it is still a "one-customer business." To invest in it as an industry-scale market, what you're waiting on is not the OCS vendors delivering — it's CPO delivering.

This article is for technology and industry trend analysis only and does not constitute investment advice.