Optical Circuit Switches Explained | Photonics

Back to Technology, Institution & Company Guides

TECHNOLOGY EXPLAINER

Optical Circuit Switches (OCS)
— rerouting light with mirrors or liquid crystal, without turning it back into electricity

An ordinary network switch turns the incoming light back into electricity, reads the destination, then turns it into light again to send it on. An optical circuit switch (OCS) physically reconnects an input fibre to an output fibre while the signal stays as light. Google reports that it has built and deployed tens of thousands of its own OCS units, which steer light with tiny MEMS mirrors. Inside are gold-coated mirrors, windows that let light through and magneto-optic crystals — a device made almost entirely of optical materials.

Built from primary sources: Google's preprint (arXiv) and ISCA 2023 paper, and official announcements from Coherent and Lumentum / Last updated September 2026

Abstract conceptual image of many thin beams of light crossing diagonally between two flat surfaces facing each other in a dark space, each folded back in a different direction — **Conceptual image (AI-generated).** An impression of the idea of rerouting light by folding its path with mirrors. **It does not show the internal structure, mirror count or light paths of a real OCS.**

What this article covers

What an optical circuit switch is (the short version)
How it differs from an electrical switch — reading versus connecting
The approaches compared — MEMS, liquid crystal, robotic, piezoelectric
Inside a MEMS OCS — two mirror arrays and camera-based control
How data centres use it — Apollo and TPU v4
Our calculation: counting ports, and mirror yield
A materials engineer's view (1): an OCS is a bundle of optical materials
A materials engineer's view (2): circulators and garnet — the crystal that doubles the ports
What is still hard
Glossary / References / Claim-to-source audit

How claims are labelled in this article

Sourced = stated in a paper or a company's official announcement (link given). Companies' claims about performance or superiority are treated as that company's claims
Our calculation = a figure this article derived, with the assumptions spelled out
Not yet confirmed = a plan or outlook with no confirmed track record
Structural readings and materials-design interpretations are marked separately as Commentary.

1. What an optical circuit switch is (the short version)

What it does: it guides light arriving on an input-port fibre to a chosen output-port fibre while it stays as light. Google's Apollo paper describes a MEMS-based OCS as one that "simply deflects light from the input port to the desired output port"Sourced
Why it helps: because light passes straight through, the same paper says, the OCS is independent of data rate and wavelength, so the same OCS can stay in service as optical transceivers move through generationsSourced
Where it is weak: reconfiguration takes milliseconds, too slow to switch packet by packet. The paper puts the switching time of commercial OCSes at typically 10 to 20 msSourced

How this fits with the rest of the series

The optical modules at either end of an OCS are covered in our explainer on optical transceivers, WDM, which puts several wavelengths on one fibre, in our explainer on wavelength-division multiplexing and microrings, and connections between fibres and components in our explainer on optical connectors and fibre attach. The electrical scale-up network that links GPUs inside a rack was covered in our explainer on optical interconnects and SerDes.

2. How it differs from an electrical switch — reading versus connecting

Fig. 1 Conceptual diagram (vector drawing). The statements on delay, rate and wavelength independence, and switching time follow Google's Apollo paper [Ref. 1]. The side-by-side layout of the two approaches and the boxes are this article's schematic.

The Apollo paper sets out the advantages of an OCS as followsSourced.

Independent of data rate and wavelength: because the optical path is broadband and passive, the same OCS can be used as data rates and the number of multiplexed wavelengths rise. The optical fabric becomes part of the building, like power and cooling, and its cost can be amortised over decades rather than years
Low power: with no per-packet processing, energy per bit can be orders of magnitude lower than an electrical switch. Electrically, MEMS mirrors are capacitive loads; their drive voltage is high, from 1 V to several hundred volts, but the power needed to hold position is extremely small. A well-designed high-voltage driver draws on the order of tens of mW per mirror or port
Low latency: delay is set by the propagation time of light, about 5 ns/m in fibre and about 3.3 ns/m in free space. An electrical switch of equivalent throughput adds tens to hundreds of ns per hop

3. The approaches compared — MEMS, liquid crystal, robotic, piezoelectric

There are many ways to change the direction of light. The Apollo paper compares OCS technologies as followsSourced.

Approach	Relative cost	Port count (example)	Switching time	Insertion loss	Drive voltage	Holds state without power
MEMS	Medium	320×320	ms	<3 dB	Hundreds of V	No
Robotic	Medium	1008×1008	Minutes (per connection)	<1 dB	—	Yes
Piezoelectric	High	384×384	ms	<2.5 dB	Tens of V	No
Guided wave	Low	16×16	ms	<6 dB	A few V	No
Wavelength switching	Not yet known	100×100	ns	<6 dB	0	Yes

All Sourced (Google Apollo paper, Table 1 [Ref. 1]). Cost and insertion loss are for the port count shown, and insertion loss includes connector loss. "—" marks a cell not given in the table. The liquid-crystal approach is not included in this table.

The paper judges the MEMS approach the most promising for delivering the large port counts data centres need at acceptable cost, noting that connections beyond 1000×1000 had been achievedSourced. It sums up the others as follows: the robotic approach handles large port counts and any fibre but switches slowly; the piezoelectric approach is costly because of assembly complexity; and guided-wave switches are cheap and compact but small in scale and high in lossSourced.

Liquid crystal and MEMS — what each company claims

Item	Coherent's announcement (liquid crystal)	Lumentum's announcement (MEMS)
Product and port count	A 300×300-port OCS. Announced March 2024 and demonstrated at OFC 2024	R300, 300×300 ports. As of March 2025, several customers were evaluating samples
Technology	Applies the digital liquid-crystal technology used for more than 18 years in wavelength-selective switches (WSS)	MEMS mirrors with decades of track record in telecom applications
Advantages the company claims	Liquid-crystal cell drive voltage below 10 V. Competing electromechanical approaches tend to be less reliable because of the very high voltages inside	Unlike liquid-crystal approaches, operates across the O, C and L bands. No closed-loop control needed. Over a trillion cumulative mirror hours in the field
Shipment outlook	Volume shipments expected to begin in 2025 (at the time of the announcement)	General availability expected in the second half of 2025 (at the time of the announcement)

All are Sourced from the companies' official announcements, but the advantages row consists of each company's claims, which this article has not verified (Coherent [Ref. 3], Lumentum [Ref. 4]). Shipment outlooks are plans as of the announcements.

What is interesting is that each company points to a weakness in the other's approach. Coherent cites the high voltage of electromechanical approaches such as MEMS; Lumentum cites the wavelength range of liquid crystal. The Apollo paper also says that in the MEMS approach, high-voltage operation can become a reliability constraint where quality control is inadequateSourced. For materials engineers, the point worth noticing is that both weaknesses stem from material properties (Section 7).

4. Inside a MEMS OCS — two mirror arrays and camera-based control

Fig. 2 Conceptual diagram (vector drawing). The 2D fibre collimator array, the two 2D MEMS mirror arrays and the control by 850 nm probe light and cameras follow the configuration of the Palomar OCS described in Google's Apollo paper [Ref. 1]. Mirror layout, angles and light paths are schematic and do not show the real optical design. A dashed path shows light sent to a different exit when a mirror is tilted.

The optical core of the Palomar OCS that Google developed in-house is described as followsSourced.

Signal light enters from a 2D fibre collimator array (an N×N fibre array plus a 2D lens array), reflects twice off two 2D MEMS mirror arrays, and enters the output collimator
To couple light optimally from any input to any output, the two MEMS mirror packages provide four degrees of freedom
850 nm monitoring light is injected into each MEMS array, the reflection is picked up by a camera, and the mirrors are controlled from the camera image to minimise loss. This made control far simpler than the earlier approach of placing a photodetector for each mirror
The result is a 136×136 non-blocking OCS with switching time on the order of milliseconds, worst-case insertion loss of 2 dB and return loss of −38 dB
Maximum power for the whole unit is 108 W, a small fraction of an electrical switch of the same capacity. Tens of thousands of these 136×136 OCSes (including 8 spare ports) have been built and deployed over the past decade

5. How data centres use it — Apollo and TPU v4

(1) The core layer of the data-centre network (Apollo)

The Apollo paper presents Apollo as "the world's first large-scale production deployment of optical circuit switches (OCSes) for datacenter networking", to the company's knowledgeSourced (Google's claim). It says that using OCSes removed the electrical switches and optical interfaces that had implemented the conventional spine layer, reducing cost and powerSourced. It also says that when introducing OCS, a budget of 15% of network cost was set aside for optical switchingSourced.

(2) Interconnecting AI computers (TPU v4)

Google's TPU v4 paper (ISCA 2023) reports a supercomputer that links 4,096 TPU v4 chips with OCSesSourced.

64 chips (4×4×4) are linked within a rack by passive electrical cables, and 64 of these blocks are linked by 48 OCSes
Each block has 16 links on each of its six faces, 96 optical links in total. Links on opposite faces go to the same OCS, so each block connects to 6 × 16 / 2 = 48 OCSes
The OCSes dynamically reconfigure the topology to suit the workload, improving scale, availability, utilisation, power, performance and more
The cost of the OCSes and optical components (the whole optical fabric, including optical modules, fibre and OCS equipment) is less than 5% of total system capital cost and less than 3% of system power

Note the split between copper and light

TPU v4 uses passive electrical cables inside the rack and light plus OCSes between racksSourced. It is the same pattern as the "copper inside the rack" of GB200 NVL72 seen in our explainer on optical interconnects and SerDes (our commentary).

6. Our calculation: counting ports, and mirror yield

Fig. 3 Drawing that includes our calculation (vector drawing). The underlying figures come from the TPU v4 paper [Ref. 2] and the Apollo paper [Ref. 1]. 6,144, about 77% and about 63% are this article's calculations, not published values. The 77% does not indicate a mirror defect rate.

Our calculation: other rules of thumb

Power per port: 108 W / 136 ports = about 0.8 WOur calculation. Compared with the 800G to 1.6T optical modules in our explainer on optical transceivers (specified ceilings of over 30 to 40 W), it is in a different league. But the OCS itself neither sends nor receives signals, so the optical modules at each end still need their own power
The effect of circulators: with bidirectional communication over a single fibre, the paper says, the OCS ports and fibres needed are halvedSourced. That works out to a 136-port OCS effectively handling 272 one-way connectionsOur calculation

7. A materials engineer's view (1): an OCS is a bundle of optical materials

Fig. 4 Conceptual diagram (vector drawing). Each element follows the description in Google's Apollo paper [Ref. 1]. The cross-section shape, thicknesses, angles and mirror count are schematic and do not show the real package structure.

Why this matters for materials engineers: mirrors, windows, seals — materials share a 2 dB loss budget

The Apollo paper describes Palomar's MEMS mirror package as followsSourced.

Inside a ceramic package is a large MEMS die with 176 individually controllable mirrors, of which 136 are finally used after calibration
A window placed over the package seals the MEMS die, protecting the micromirrors from dust and moisture over the product's life
The window is tilted to prevent coupling in the zero-voltage state and back-reflection into the collimators
The window coating is designed to pass both the 850 nm monitoring light and the telecom wavelengths with minimal loss. This is critical because each data path passes through window coatings eight times in all (top and bottom surfaces)
Each micromirror has a highly reflective gold coating to minimise optical loss. The mirrors are made in a DRIE (deep reactive-ion etching)-based process as large, flat, highly reflective micromirrors

The paper wants insertion loss ideally below 2 dB, calling the optical link budget "a very precious commodity"Sourced. That 2 dB has to cover fibre-to-lens coupling, two reflections off gold mirrors and eight passes through window coatings (our framing).

Our calculation: the loss a window coating is allowed

Assumption (ours): suppose that, of the 2 dB, 0.1 dB in total can be allotted to the window coatings.
0.1 dB / 8 passes = 0.0125 dB per surface. In power terms that is 1 − 10^(−0.00125) = about 0.29% lossOur calculation.
And that has to be achieved simultaneously at two widely separated wavelengths, 850 nm and the telecom band (such as the O band). The 0.1 dB allocation is an assumption for illustration and does not appear in the paper.

What this shows is that the performance of an OCS is not set by the MEMS mechanical design alone. Multilayer coatings with low reflection in two separate wavelength bands, metal reflective films whose reflectance does not fall over time, sealing structures that keep out moisture, and materials around the electrodes and wiring whose insulation does not degrade under hundreds of volts applied for years — the Apollo paper, too, calls the high-voltage ICs and operation one of the largest reliability challenges for MEMS optical switchesSourced. The clash of company claims in Section 3, "liquid crystal is low-voltage, MEMS is broadband", ultimately comes down to a choice of materials: the wavelength characteristics of liquid-crystal materials versus the reliability of materials that withstand the high voltages of MEMS (our commentary).

8. A materials engineer's view (2): circulators and garnet — the crystal that doubles the ports

Fig. 5 Conceptual diagram (vector drawing). The function and materials of the circulator follow Google's Apollo paper [Ref. 1]. The internal optical layout is not drawn, and the shape of the figure is this article's schematic.

Why this matters for materials engineers: a magneto-optic crystal doubles the switch's port count

According to the Apollo paper, an optical circulator is a three-port non-reciprocal device: light into port 1 goes to port 2, and light into port 2 goes to port 3. It is built from a birefringent crystal, a magneto-optic Faraday rotator (typically made of garnet) and polarisersSourced. Combining transmit and receive on one fibre halves the number of OCS ports and fibres neededSourced.

The paper goes on to say that before large-scale data-centre use, circulators were used in limited volumes, mainly in erbium-doped fibre amplifiers for the C band, and that they were extended to the O band (CWDM4) through appropriate optical coatings and a rework of the optical designSourced. It also stresses that reducing reflections and ensuring directivity (suppressing leakage from port 1 to port 3) is critical, since leaked light is effectively the same as a reflection within the linkSourced.

The materials story is concentrated right here.

Faraday rotator: a magneto-optic crystal that rotates polarisation in a magnetic field. The amount of rotation and the transparency vary with wavelength, so a change of wavelength band means a change of material and thickness design
Coatings: here too, anti-reflection multilayers were the key to making a C-band component usable in the O band
A change in volume: a component once made for amplifiers in submarine cables and long-haul links has become a mass-produced item along with data-centre OCS

Specialised crystal components for telecom are now wanted in data-centre quantities — for optical communication materials suppliers, that means a new outlet for existing crystal-growth and thin-film technologies (our commentary).

9. What is still hard

(1) Switching speed

The Apollo paper says the switching time of commercial OCSes is typically 10 to 20 ms, limited by control software and mirror settling time, and that receiver initialisation in optical transceivers also stands in the way of fast switchingSourced. It adds that milliseconds are enough for today's uses, but that large OCSes able to switch in microseconds represent a major future opportunityNot yet confirmed.

(2) The three-way trade-off between port count, switching time and loss

The same paper says current OCS designs involve a fixed trade-off between port count, switching time and insertion loss, and that exploring other approaches will remain an important area of researchSourced. The figures in the table in Section 3 show exactly that three-way bind.

(3) Handling a wider wavelength range

The Apollo paper points out that future data centres may use many wavelengths spanning the O, S, C and L bands, and that keeping reflections low over such a wide range becomes difficultSourced. The window-coating problem in Section 7 gets harder as the band widens (our commentary).

(4) Commercial uptake

At the time of their respective announcements, Coherent's and Lumentum's 300×300-port products were expected to reach volume shipment or general availability in 2025Sourced. Actual shipments and the scale of adoption since then could not be confirmed from primary sources within the scope of this article's researchNot yet confirmed. Market-size estimates for OCS are not used, in line with this series' policy.

The article in summary

An OCS reconnects an input fibre to an output fibre without turning the light back into electricitySourced
It is independent of data rate and wavelength, low-power and low-latency. Its weakness is a switching time of around 10 to 20 msSourced
Google's Palomar is a 136×136 MEMS OCS with 2 dB worst-case loss and 108 W maximum power, deployed in the tens of thousandsSourced
In TPU v4, 48 OCSes link 4,096 chips, and the whole optical fabric accounts for less than 5% of capital cost and less than 3% of powerSourced
Liquid crystal (Coherent: low voltage) and MEMS (Lumentum: broadband) each point to a materials-rooted weakness in the other (the companies' claims)
Inside are gold mirrors, window coatings passed eight times, seals and garnet Faraday rotators — a bundle of optical materials (our commentary)

10. Glossary

Optical circuit switch (OCS): A switch that physically reconnects input fibres to output fibres without converting light into electricity.

Electrical packet switch (EPS): An ordinary network switch that turns light back into electricity and routes packets by reading their destination.

OEO conversion: Optical-to-electrical-to-optical conversion. It happens every time a signal passes an electrical switch.

MEMS: Micro-electro-mechanical systems. Movable structures made by silicon micromachining; in an OCS, tiny mirrors.

DRIE: Deep reactive-ion etching. A process that cuts deep, vertical features into silicon.

Collimator: A lens component that turns light leaving a fibre into a parallel beam.

Non-blocking: Able to connect any input to any output without obstructing other connections.

Insertion loss: The amount of light lost in passing through a component (dB). At 2 dB, about 63% remains.

Return loss: How little of the incident light comes back as reflection (dB).

Liquid-crystal approach: Switching by using a voltage to change the orientation of liquid crystal and so control the polarisation or direction of light.

Optical circulator: A non-reciprocal three-port device that sends light only to the next port in a fixed order.

Faraday rotator: A magneto-optic element that rotates the plane of polarisation of light in a magnetic field. Garnet and similar materials are used.

WDM: Wavelength-division multiplexing. Carrying several wavelengths on one fibre.

11. References (primary sources)

Urata R. et al. (Google, arXiv, 2022) "Mission Apollo: Landing Optical Circuit Switching at Datacenter Scale", arXiv:2208.10041 — arxiv.org
Jouppi N. P. et al. (Google, ISCA 2023) "TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings", arXiv:2304.01433 — arxiv.org
Coherent "Coherent Announces Optical Circuit Switch for Data Centers", 25 March 2024 — coherent.com
Lumentum "Lumentum Optical Circuit Switch to Improve Next-Generation AI Data Center Scalability", 26 March 2025 (PDF) — q4cdn.com (Lumentum investor relations)

12. Claim-to-source audit

Claim in the text	Basis	Label
That a MEMS OCS deflects light from the input port to the desired output port. That it is independent of data rate and wavelength and reusable across generations, with the optical fabric becoming part of the building and amortised over decades. That with no packet processing it is low-power; that mirrors are capacitive loads driven at 1 V to several hundred volts; and that drivers draw on the order of tens of mW per mirror or port. Latency of about 5 ns/m in fibre and about 3.3 ns/m in free space, against tens to hundreds of ns per hop for an EPS. That commercial OCS switching time is typically 10 to 20 ms, and that large microsecond-class OCSes are a future opportunity. Insertion loss ideally below 2 dB, with the optical link budget "a very precious commodity". That keeping reflections low across the O, S, C and L bands becomes difficult. The fixed trade-off between port count, switching time and insertion loss. The technology comparison in Table 1 and the judgement that MEMS is the most promising (with 1000×1000-plus achieved), the pros and cons of each approach, and that high voltage can become a reliability constraint. Palomar's optical core (2D collimator arrays, two 2D MEMS mirror arrays, four degrees of freedom, 850 nm monitoring light and camera control), 136×136 non-blocking, millisecond switching, 2 dB worst-case insertion loss and −38 dB return loss, 108 W maximum, and tens of thousands built and deployed (8 spare ports). That 136 of 176 mirrors in the ceramic package are used; the window seal (dust and moisture); the tilted window; coatings passing both 850 nm and telecom wavelengths, passed eight times per path; gold coating; DRIE; four lines per mirror; and high voltage as one of the largest reliability challenges. That Apollo is positioned as the world's first large-scale production OCS deployment, that the spine-layer electrical switches were removed, and the 15% network-cost budget. The circulator's function (three-port, non-reciprocal; birefringent crystal, garnet Faraday rotator, polarisers), halving ports and fibres, extension from EDFA use to the O band, and the importance of reflection and directivity	Google Apollo paper (arXiv:2208.10041)Reference 1 https://arxiv.org/abs/2208.10041	Sourced
That 4,096 TPU v4 chips are linked by OCS. That blocks of 64 chips (4×4×4) are linked within a rack by passive electrical cables. That each block has 96 optical links and connects to 48 OCSes. That Palomar is 136×136 (128 + 8 spare). That the OCSes reconfigure the topology dynamically. That the cost of the whole optical fabric (optical modules, fibre and OCS equipment) is under 5% of capital cost and under 3% of power	Google TPU v4 paper (ISCA 2023, arXiv:2304.01433)Reference 2 https://arxiv.org/abs/2304.01433	Sourced
That Coherent announced a 300×300-port OCS based on digital liquid-crystal technology on 25 March 2024 and demonstrated it at OFC 2024. That it has used liquid-crystal technology in WSS for more than 18 years and shipped more than 160,000 units. The claims that liquid-crystal cell drive voltage is below 10 V and that electromechanical approaches tend to be less reliable because of high voltage. Volume shipments expected to begin in 2025 (as of the announcement)	Coherent press release (25 March 2024)Reference 3 https://www.coherent.com/news/press-releases/optical-circuit-switch-for-data-centers-live-demo-at-ofc-2024-based-on-ultrareliable-dlx-technology	Sourced
That Lumentum's R300 (300×300 ports, MEMS) was sampling to several customers as of March 2025, with general availability expected in the second half of 2025. The claims that, unlike liquid-crystal approaches, it operates across the O, C and L bands, needs no closed-loop control, and has over a trillion cumulative mirror hours in the field	Lumentum news release (26 March 2025)Reference 4 https://s21.q4cdn.com/377324469/files/doc_news/Lumentum-Optical-Circuit-Switch-to-Improve-Next-Generation-AI-Data-Center-Scalability-2025.pdf	Sourced
The match 64 × 96 = 6,144 and 48 × 128 = 6,144. 136 / 176 = 77%. 10^(−0.2) = 0.63. 108 W / 136 = 0.8 W per port. 272 effective one-way connections for a 136-port OCS. 0.0125 dB per surface (about 0.29%) under the assumption that 0.1 dB is allotted to the window coatings	Our calculation. The 0.1 dB allocation is this article's assumption and does not appear in the paper. 77% is not a mirror defect rate	Our calculation
Realisation of large microsecond-class OCSes. Actual shipments and adoption of the Coherent and Lumentum products	The paper's future outlook and the companies' plans as of their announcements; no track record could be confirmed within the scope of this article's research	Not yet confirmed
The framing of the EPS/OCS contrast. The interpretation of the liquid-crystal versus MEMS dispute as a choice of materials. The framing that mirrors, windows and coupling share the 2 dB budget. The reading that multilayer coatings, metal reflective films, sealing and high-voltage insulation materials determine performance. The reading that Faraday rotator material design changes with wavelength band and that telecom crystal components are now wanted in data-centre quantities. The framing of TPU v4 and NVL72 sharing "electrical inside the rack"	This article's own framing and commentary based on public material. Not views expressed by the companies or researchers	Commentary
Market-size estimates for OCS (Lumentum's announcement cites a research firm's forecast)	Research-firm estimates are not used, in line with this series' policy	Commentary
That Figs. 1, 2, 4 and 5 are explanatory drawings, that Fig. 3 is a drawing that includes our calculation, and that the hero image is AI-generated	A note by this article	Commentary

Last updated 26 September 2026. Sources are limited to primary material (two Google papers and official announcements from Coherent and Lumentum). Statements of superiority in company announcements are treated as each company's claims and have not been verified here. Because the article includes structural readings and materials-design interpretations, those are marked as Commentary and kept separate from sourced fact. Commercial OCS shipments, scale of adoption and market size are not covered. All figures are for explanation. Figs. 1, 2, 4 and 5 are vector drawings, Fig. 3 is a vector drawing that includes our calculation, and the hero image is AI-generated; none of them shows the structure, light paths or dimensions of real equipment.