How the switch works.

Patented technology from the University of Arizona. Three steps, all of them optical; nothing is detected, buffered, or re-transmitted. Then the engineering that is still ahead, stated plainly.

The fabric is where the power goes.

A data center network is a tree of switches. Every time a packet crosses one, the light that carried it is detected by a photodiode, handled as electrons by a switch chip, and re-emitted by a laser in a transceiver. Each of those conversions costs energy: about 40 picojoules per bit to cross one 51.2 Tb/s switch with its optics, measured at the box.

Operators compensate for unpredictable traffic by over-provisioning, wiring more paths than they need so that traffic always has somewhere to go. Every spare path is another switch drawing power and making heat.

Optical circuit switches remove the conversion: they steer the light itself from one fiber to another. Google has run them in production since 2022, and they cut the power of its network fabric by 41%. But the switches that ship today tilt small mirrors, and a mirror takes ten to twenty-five milliseconds to settle. That is fine for rewiring a cluster between jobs and useless for following traffic inside one.

Two cautions, from the inventor. This is a circuit switch, not a faster packet switch: an electronic switch forwards each packet in nanoseconds and keeps that job, while a circuit switch holds a path open for the elephant flows that carry most of a cluster's bytes. And the thousand-fold speed claim is against the mirror-tilt optical switches it replaces, never against electronics. What a circuit switch gives in return is independence from data rate and protocol: it steers light without reading it, so the fabric built today carries the transceivers of 2030.

Steering by diffraction.

FIG. 2. The pattern on array 22 is a computed hologram of the output layout. The signal is spread across thousands of mirrors, so the loss of a few barely dims it: 2.5% dead mirrors cost 0.25 dB.

Light arrives
Data comes in on a fiber (20) and fans out as a collimated beam onto a chip the size of a postage stamp.
A pattern steers it
Mirrors far smaller than the beam form a diffraction pattern (22). The pattern's spacing sets the exit angle; its orientation sets the direction. Rewriting it takes microseconds.
It lands on the right fiber
The beam couples into the chosen output (24) and continues at whatever data rate and format it arrived with. 25 Gbaud QPSK has been passed with no degradation.

Try the hologram.

The left pane is a 128 by 128 micromirror array showing a real binary pattern. The right pane is its actual Fourier transform, computed in your browser. Pick a port and the pattern rewrites row by row; the bright spot is where the light lands, and the faint spots are the orders a real switch absorbs.

Micromirror array dark cells tilt toward the output

Output fiber plane far field, the Fourier transform of the array

Send to

Rewrite 12 µs

The chip, today and next.

The prototypes used a Texas Instruments DLP micromirror chip, the same part found in projectors. Each of its mirrors tilts twelve degrees one way or the other, so the pattern it can show is binary: a mirror is either sending light toward the output or away from it. Physics caps a binary amplitude hologram at 10% efficiency, which is a 10 dB loss before the optics around it add about 6 dB more.

That loss is the engineering that is still ahead, and the patent already describes the answer. In its phase-modulating embodiment each mirror does not tilt; it moves straight up or down by up to 775 nm, half a wavelength at 1550 nm, under an individually addressed electrode. The pattern becomes a grayscale hologram that shapes the light instead of masking it. With only four height levels the diffraction efficiency exceeds 80%; with ten to twenty it approaches 100%, and the loss falls to 0.3 dB. Such arrays exist in research labs at 310 kHz; PQT's next step is a commercial-grade one, built with a MEMS fabrication partner.

With loss at 3 dB, the switch fits inside a normal transceiver's link budget with no amplifier, and the energy case closes. The energy page shows exactly how.

What the patent provides beyond speed.

US 9,453,970 and US 9,726,827 describe a complete switching system, not just a fast mirror. These are the parts a network architect will care about.

Independent regions, hitless changes

Each input fiber illuminates its own region of the array, and each region is reset on its own. Switching one path causes no loss on any other. Patterns are precomputed and held in a look-up table, so a change is a memory load and a reset pulse, not a calculation.

No feedback loop

A mirror-tilt switch spends most of its milliseconds in a servo loop locating each mirror. Here the mirrors land on physical stops. Nothing has to be measured to know where the light went, which is why the switch is faster, simpler, and draws less power.

Achromatic operation for wavelength-multiplexed traffic

Diffraction angle depends on wavelength. The patent's two-array configuration diffracts in equal and opposite directions so the net deviation is zero across the whole C-band, and the angle at every output fiber stays the same whatever the path. A single dispersion-compensating element gives up to sixteen-fold more throughput at some wavelengths.

Isolation and uniformity

The hologram is optimized with an iterative Fourier algorithm against crosstalk. The prototype measured isolation above 40 dB and 0.3 dB of variation across every output, independent of how many mirrors had failed.

Multicast, monitoring, attenuation

A pattern can form several spots at once, so one input can reach several outputs. A set fraction of a signal can be diverted to a detector. The pattern sets per-port attenuation. These are software features of the same chip.

Scale

Each input needs about 200 by 200 mirrors. At today's 13.68 µm mirror pitch, three tiled chips make a 144×144 switch; smaller mirrors deflect light further and raise the count into the hundreds. Polarization-dependent loss is minimal because the switch works in reflection.

Measured, then designed.

Table 1 Diffractive switch prototypes against shipping optical circuit switches
	Bench prototypes, measured	Next-generation chip, designed	Shipping OCS⁵
Ports per switch	7×7 and 32×32 built^{1, 2}	thousands⁴	136 to 576
Reconfiguration	12 µs¹	about 10 µs⁴	10 to 25 ms
Loss of light during rewrite	7 µs²	a few µs	n/a
Insertion loss	6.5 dB optics plus 10 dB diffraction²	3 dB⁴	2 dB or less
Crosstalk	below −33 dB at 100 µm port spacing²	same	below −50 dB
Switch electronics	about 12 W, summed from part datasheets⁶	about 12 W	45 to 140 W
Data rate	Independent of rate and format; 25 Gbaud QPSK passed with no degradation²		Independent
Tested in	The NSF CIAN network testbed and a three-rack testbed at Microsoft Research³

B. Lynn et al., "Design and preliminary implementation of an N×N diffractive all-optical fiber optic switch," IEEE/OSA Journal of Lightwave Technology 31(24), 2013.
P.-A. Blanche et al., "Fast non-blocking N×N optical switch using diffractive MOEMS," OFC 2015; P.-A. Blanche, L. LaComb, Y. Wang, M. C. Wu, "Diffraction-based optical switching with MEMS," Applied Sciences 7(4), 411, 2017.
M. Ghobadi et al., "ProjecToR: agile reconfigurable data center interconnect," ACM SIGCOMM 2016, with Microsoft Research.
Post-Quantum Tek design targets, January 2026. Not yet measured.
R. Urata et al., "Mission Apollo," Google, 2022; Calient S320 and Polatis 7000n datasheets.
Texas Instruments DLP7000 4.4 W, DLPC410 2.7 W, DLPA200 2.4 W, plus an application FPGA.

Where it goes from here.

Bench-proven, 2013 to 2017
Seven- and 32-port switches on DLP chips, 12 µs, full C-band, tested in the NSF CIAN testbed and at Microsoft Research. Patented. Done.
Commercial prototype
A hardened switch with an engineering partner: fiber arrays, an FPGA control plane with an open management interface, and the first phase-modulating mirror array. The target is 3 dB and hundreds of ports.
Pilot
A research network or test facility where microsecond reconfiguration has immediate value and the remaining loss can be amplified away, followed by a hyperscaler qualification.

Words we use.

Optical circuit switch: A switch that connects one fiber to another by steering the light itself, without reading the data. It holds a path until told to change it.
Elephant and mouse flows: Network shorthand for the few large, long-lived flows that carry most of the bytes and the many short ones that do not. A circuit switch is built for elephants; a packet switch handles the mice.
Packet switch: The electronic switch in every data center today. It reads each packet's address and forwards it, which means converting light to electricity and back at every hop.
OEO and OOO: Optical to electrical to optical: what a packet switch does to the signal. Optical to optical to optical: what this switch does. The middle letter is where the energy goes.
Diffraction: What light does when it meets a pattern of features close to its own wavelength. Instead of reflecting as a mirror would, it spreads into directions set by the pattern's spacing. A rainbow on the back of a disc is diffraction.
Hologram: A pattern computed so that light diffracting from it forms a chosen picture somewhere else. Here the picture is a dot on the chosen output fiber.
Micromirror array: A chip carrying up to two million mirrors a few microns wide, each one moved electrically. The projector in a conference room has one.
Insertion loss: How much light the switch wastes, in decibels. 3 dB is half; 10 dB is nine-tenths. Transceivers have a limited budget for it, which is why loss decides whether an amplifier is needed.
Reconfiguration time: How long from the command to the light arriving on the new fiber. Milliseconds for mirror-tilt switches; twelve microseconds here.
Spine and leaf: The two layers of switches in a data center network. Leaves connect to servers; spines connect leaves. Optical circuit switches replace spines first.
PUE and WUE: Power usage effectiveness: facility electricity divided by IT electricity. Water usage effectiveness: liters of cooling water per IT kilowatt-hour. Both multiply any watt saved inside the hall.

Questions we get.

Is this a packet switch?

No. It is a circuit switch: it sets up a light path between two fibers and holds it until told to change. A scheduler or software-defined controller decides when. In network terms it is an elephant switch, not a mouse switch: it carries the large, long-lived flows that dominate a cluster's traffic, and the short bursty ones stay on electronic packet switches. What is new is that the change takes microseconds instead of milliseconds, so the controller can follow traffic rather than only rewire between jobs.

Is it faster than an electronic switch?

No, and no optical switch is. An electronic packet switch makes a forwarding decision in nanoseconds. The thousand-fold figure on this site compares the PQT-HOS with the mirror-tilt optical circuit switches shipping today, which take ten to twenty-five milliseconds. The two kinds of switch do different jobs and sit in different places in the fabric.

Why does the data rate not matter?

The switch only steers light; it never reads it. A 100 Gb/s signal and a 1.6 Tb/s signal pass through the same way, in any modulation format or protocol. That is true of every optical circuit switch, and it is what makes one future-proof: the fabric does not need replacing when transceivers get faster.

What is "directionless and contentionless"?

Any input can reach any output, and two signals never block each other on the way. Diffraction makes this natural: each input lights its own region of the array, and each region aims independently.

How much energy does it save?

Crossing an optical circuit switch costs about one picojoule per bit; crossing an electrical switch with its optics costs about forty. Replacing the spine of a fabric with optical switching cut network power 41% in Google's production network. At the facility level that is 2 to 3% of IT load. The energy page has the full model.

What about the 16 dB of loss?

It is the main engineering task, and it is an energy problem as much as an optical one: today's prototypes need an amplifier on every fiber. The phase-modulating mirror array described and claimed in the patent is designed to bring total loss to 3 dB, inside a standard link budget, with no amplifier.

Can it multicast?

Yes. The hologram that aims one input at one output can just as well aim it at several. One signal reaching many fibers at once is exactly the shape of a broadcast or all-reduce step in AI training, and it needs no extra hardware.

What happens when mirrors fail?

Very little. Each path is carried by tens of thousands of mirrors, so losing 2.5% of them costs 0.25 dB and the pattern can be recomputed around them. In a mirror-tilt switch, one failed mirror is one dead port.

Who has verified it?

The results are in eight peer-reviewed papers. The Microsoft Research work is the ProjecToR paper at SIGCOMM 2016; the UC Berkeley work is the 2017 Applied Sciences paper with Ming Wu's group; Texas Instruments supplied the DLP chips. The science page links every one.

What is the data center switch market?

Optical circuit switching went from a Google-only practice to a merchant market in 2025. Cignal AI forecasts it above $8 billion a year by 2030. Lumentum, Coherent, and Huber+Suhner are shipping millisecond switches; the microsecond tier is where PQT and a handful of funded startups are heading.