OCS (Optical Circuit Switching): The Next Optical Revolution in AI Data Centers

Over the past two years, the AI boom has made the phrase "the data center bottleneck" sound like more than a cliché for the first time. The problem is not the GPU, not HBM, not PCIe — it is the network at the very bottom: the switching architecture. AI clusters keep growing, from a single room packed with GPUs to many rooms stitched together into an AI super-factory. And every time scale grows 4x, the cabling, power and latency of the switching network explode exponentially.

The whole industry now sees the same thing:

The core of next-generation AI networking is not faster SerDes, but smarter topology plus lower-power switching.

That is why OCS (Optical Circuit Switching) is drawing attention.

Why Is Everyone Taking OCS Seriously?

In my view, the reason is:

Light should not be forced into electrons and then back into light.

A traditional switch (i.e., OEO switching) does the following:

1. Optical → electrical

2. Switching handled inside the chip

3. Electrical → optical

That sounds reasonable, but at AI-scale data volumes it turns into three big problems:

1. Electrical switching is the bottleneck

All traffic must pass through the switch chip, so:

• When SerDes speeds can't keep rising, your entire network upgrade stalls

• Chip power becomes unreasonably high

• Switching equipment becomes big, hot and expensive

2. Real-time OEO processing is too costly

1,000+ racks and tens of thousands of data flows going in and out, each requiring OEO conversion — the cost scales linearly, but traffic grows exponentially.

3. Performance can't scale with the cluster

Everyone knows GPU scaling is hitting diminishing returns, but few notice that:

The network is the main reason diminishing returns arrive early.

OCS is like replacing one of the rail tracks with a frictionless one.

The Essence of OCS: "Optical Path Reconfiguration" That Skips O/E/O

It is not a "faster switch"

Nor is it a "new type of optical module"

OCS is:

• A way to connect optical paths directly to optical paths

• Every connection is a "dedicated optical path"

• Data rate, protocol and modulation format don't matter

In other words:

You connect fiber A to fiber B without touching electronics in between.

This also means:

Fully Rate-Agnostic

→ No need to replace the OCS for 800G / 1.6T / 3.2T or beyond

Truly Low Power

→ No DSP, no SerDes, no switch chip

Highly Scalable

→ Want more ports? You're just connecting more fiber

→ No need to redeploy the spine layer

Ultra-Low, Stable Latency

→ No packet processing, no buffers, no ASIC pipeline

The Four Technology Paths of OCS

There is a lot of engineering detail here, but I'll summarize it as simply as possible:

1. MEMS: mature, but with limits

• Rotating mirrors, ~25ms switching

• Low insertion loss, large scale, first to commercialize

• Adopted in Google TPU v4 / v5 / Ironwood

→ The most mature OCS in the near term, but with clear long-term bottlenecks.

2. Liquid crystal (LC / DLC): extremely high reliability

• Derived from WSS technology

• World-leading reliability

• Slower switching (~100ms)

→ Suited to large-scale spines that need high reliability and stability.

3. Piezoelectric (Piezo / DLBS): stronger physical limits

• No mechanical moving parts at all

• Lowest insertion loss and return loss

• Harder to scale to ultra-high port counts

→ Very promising long-term, but the challenge is scaling.

4. Silicon photonic waveguides (SiPh): the one to watch

• Fastest switching (<100 µs)

• Insertion loss can be compensated with SOAs

• Greatest cost-reduction potential (CMOS process)

→ Very likely the ultimate solution.

Why OCS Is Really Taking Off: All Three AI Networks Are Hitting Limits

AI is not one network but three:

• Scale-up (within a rack / pod)

• Scale-out (within a data center)

• Scale-across (between data centers)

OCS has found a place in all three.

1. Scale-up: TPU's 3D torus can't be held together without OCS

The evolution of Google's TPU topology shows why OCS is necessary.

TPU v4

• 4,096 TPUs

• 48 units of 136-port MEMS OCS

• 3D torus topology relies on OCS to reconfigure optical paths

• Both latency and power drop significantly

Ironwood

• 9,216 TPUs (more than double the scale)

• Requires double the OCS ports

• Architecture fully retains OCS rather than CPO

This shows one thing:

Google isn't "trialing" OCS — it has made it part of the TPU system design.

2. Scale-out: redesigning the data center spine layer

Google embedded Apollo OCS into its Jupiter network, with results that shook the industry:

• Latency down 10%

• Throughput up 30%

• Overall power down 40%

• Cost down 30%

The impact of optical switching in large-scale data centers is far more pronounced than expected.

And this is not "partial traffic optimization"

It is Google's conclusion from real-world operation:

OCS in the spine layer = less pressure on the entire DC network

For hyperscalers, that is an incentive impossible to ignore.

3. Scale-across: NVIDIA's third network (across data centers)

With Spectrum-XGS, NVIDIA introduced a new concept:

Scale-across is the third critical network of the AI era.

When multiple data centers need to become "one AI cluster":

• Massive DCI (long-haul optical links) is needed

• Dynamically reconfigurable topology is needed

• Low-latency synchronization across regions is needed

These needs map almost perfectly to OCS characteristics:

DCI requirement	OCS characteristic
High bandwidth	Rate-agnostic, scalable ports
Dynamic topology	Real-time optical path reconfiguration
Long distance	Low loss, works with C-band
Heterogeneous environments	Fully protocol-agnostic

Coherent has further announced:

→ A DCI-dedicated C-band OCS launching in 2026

This means:

Scale-across will be the next big OCS growth market.

OCS vs. CPO: Not Rivals, but the Twin Cores of AI Networking

This is the most commonly misunderstood part today.

• Wrong view: OCS will replace CPO

• Right view: CPO and OCS are complementary

NVIDIA's data makes it clearest:

Network architecture	Power
Pluggable	83 pJ/bit
Pluggable + OCS	50 pJ/bit
CPO	48 pJ/bit
CPO + OCS	31 pJ/bit (lowest)

This proves:

• CPO handles high-speed, short-reach switching (rack / ToR / leaf)

• OCS handles topology reconfiguration and spine / DCI

Combined → the end-state of the AI SuperFabric.

Supply Chain View: Which Positions Are Worth Watching?

I break the supply chain into three layers:

Layer 1: Core switching technology

MEMS

• Globally dominated by one or two players (e.g., Calient, Lumentum)

Liquid crystal (LC / DLC)

• Coherent has the most mature technology

Piezoelectric (DLBS)

• Strongly pushed jointly by Polatis + Lingyun Photonics

Silicon photonic waveguides

• iPronics

• China's Taclink is catching up fast (offering 32×32 with in-house SOA capability)

The path most worth watching over the next decade: SiPh OCS

Layer 2: Optical / passive components

Including:

• Fiber arrays

• FAU

• Lens arrays

• Polarization components

• Optical wedges (YVO₄)

• Coupling assemblies

• WDM（Z-block）

Layer 3: Complete / system-level OCS

• Calient / Polatis (mature)

• Coherent (liquid crystal systems)

• TeraHop (an Innolight subsidiary)

• Advanced Fiber Resources (entering via contract manufacturing)

• Taclink (SiPh OCS prototype)

This layer will grow very fast, especially with:

• Hyperscalers building their own OCS

• Chinese AI data centers investing in all-optical networks

Finally: OCS Is Not About Being "Faster" but "Simpler"

Networking in the AI era is evolving from:

Electrical switching → hybrid opto-electrical → all-optical network

OCS is not a "faster switch"

but rather:

The core of a new architecture that makes network topology flexible, low-power and scalable again.

If you work on:

• AI Infrastructure

• GPU/TPU/ASIC clusters

• Hyperscale data center architecture

• Optical modules / SiPh

• Cloud Networking

then the term OCS will show up more and more in your meetings and roadmaps over the next five years.

Summary

OCS is foundational infrastructure for next-generation AI data centers, playing a role similar to the GPU's in AI. It won't replace everything, but it will redefine the logic of everything.

• It will be a must-have for TPUs / AI ASICs

• It will reshape the spine layer

• It will be the standard option for next-generation DCI

• It will coexist with CPO — even become its best partner

• It will drive a reshuffle of the optical supply chain

• It will become a new battleground for the SiPh industry