Silicon has no Pockels effect, and everything about modulator materials follows from that

A silence in the crystal

Some of the most important facts in engineering are refusals. Here is one. Send light through silicon, apply an electric field, and try to make the field change the light’s phase directly. Nothing happens.

The refusal has a clear geometric cause. The effect in question is the Pockels effect: a refractive index that changes linearly with an applied field. It can only exist in crystals that have no center of symmetry. Silicon’s diamond lattice has one. Every silicon atom sits in a neighborhood that looks identical when inverted, and that inversion symmetry cancels the linear response exactly. In unstrained silicon, the cancellation is not partial. Group theory makes it exact. Strain can break the symmetry and give a small effect, but the crystal itself has none.

Silicon won every other job in electronics, and humans built their best manufacturing system around it. But silicon cannot give photonics the one response it needs most. Every important question about modulator materials in 2026 starts from this one gap.

Speaking by brute force

Silicon photonics found a workaround. The field cannot change the light directly, so it moves charge instead. A PN junction across the waveguide injects or depletes carriers, and the carrier density changes the index. This mechanism is plasma dispersion, and it works. Every silicon photonic transceiver to date shipped with it.

But plasma dispersion is a workaround, and the mechanism brings three costs with it:

• Loss and chirp. Free carriers shift the phase, but they also absorb light. So every modulation step also attenuates the signal. This shows as insertion loss (reports give figures around 4.4 dB per millimeter of phase shifter). It also shows as chirp: a residual phase change on a signal that should change only in amplitude.
• A bandwidth ceiling. The PN junction’s capacitance and the series resistance of the doped silicon set an RC limit. Reported silicon modulators reach near 60 GHz of bandwidth.
• Nonlinearity. The carrier response is nonlinear, and this costs more as modulation formats get denser.

At 100G per lane, these costs were acceptable. At 200G, they become tight. At 400G per lane, the question is whether to continue to pay them. To answer it, you must know what the alternatives cost instead.

The currency of the argument

Every modulator in this essay is a Mach-Zehnder. It splits the light, shifts the phase of one arm against the other, and recombines the two arms. The phase difference then becomes an amplitude difference. The material controls one thing only: how much phase one volt gives per millimeter of electrode.

One figure of merit captures that quantity: the voltage-length product Vπ·L. It is the half-wave voltage multiplied by the length necessary to reach it. For a given material and design it is roughly constant. So the designer selects a point on it: a long device with a low voltage, or a short device with a high voltage.

This number is important because of a square. Driver power scales with the square of the voltage swing. So a material with half the Vπ·L is not two times better. At fixed length, it is four times better on drive power. Remember that multiplier for the rest of the essay, and try the trades here:

The heir, and its factories

Lithium niobate shows what a modulator material can do when the crystal has the right symmetry. It has no inversion center and a true Pockels coefficient near 30 pm/V. It gives pure phase modulation, with no carriers to absorb light, no chirp and no PN junction to limit the bandwidth. The thin-film version (TFLN) delivers what the physics predicts:

• voltage-length products near 1.3 V·cm
• devices demonstrated above 110 GHz, with research parts that extrapolate to 220 GHz while they carry 390 Gb/s of PAM8
• total insertion loss under 5 dB on good platforms.

On merit, it is the successor.

Until recently, the missing part was not physics but factories. The important news in 2026 is ordinary manufacturing news:

• HyperLight, UMC and Wavetek announced a partnership for volume TFLN production on 6-inch and 8-inch wafers.
• CCRAFT opened a production-ready foundry.
• The key enabling change is in lithography: from electron-beam writing to DUV steppers on 200 mm. This change turns a research process into a product line.

Two numbers now decide adoption, and neither is a coefficient. They are the commercial yield and the cost per die against silicon photonics.

One number drives the other. Cost per good die scales inversely with yield. So part of the cost gap is a yield gap, and wafer cost and process steps give the rest. This link contains a forecast: as yields improve, part of the gap closes too, with no new physics necessary.

The cousin built for heat

Lithium tantalate is the less famous relative of lithium niobate. It has almost the same electro-optic coefficient, and it is better suited to hard environments:

• seventeen times less birefringence
• much less photorefractive damage
• a reported tenfold lower RF loss tangent
• demonstrated DC-bias stability at around 120 °C, which may be the most important property.

That last number aims at a specific future. In co-packaged optics, the modulator sits inside an optical engine next to a switch ASIC. It does not get its own thermal budget, so it must operate uncooled.

Tantalate also comes with a supply chain that photonics does not have to build. Surface-acoustic-wave filters use the same crystal, and almost every phone contains one. So wafer capacity that photonics can use already exists at scale, mostly in Japan. For lithium niobate, the main obstacle is factories. A material that already has factories deserves more attention than its press coverage gives it.

The prodigy and the extremist

Two more materials complete the field, and each one teaches a different lesson.

Barium titanate (BTO) has the largest coefficient. Measurements of its Pockels coefficient give values from about 900 pm/V on silicon to about 1,270 pm/V in single-crystal films. That is thirty to forty times the coefficient of lithium niobate. Reported devices reach 0.09 V·cm, an order of magnitude better than any other material here. Through the square law, that is a possible hundredfold decrease in drive power. BTO also integrates monolithically on silicon-on-insulator.

The caution about BTO is one frequency-dependent fact. Careful measurements from megahertz to sub-terahertz show that the large coefficient decreases as frequency increases. Resonant BTO devices showed bandwidths as low as 11 GHz. A coefficient measured at low frequency is a promise, not a datasheet value. No measurement establishes the upper limit of the material yet, so it is a reason for both interest and patience.

Plasmonics goes to the extreme of speed. A metal slot filled with electro-optic polymer confines the light. Devices then shrink to tens of microns, bandwidth goes above 170 GHz, and laboratory results approach a terahertz. The cost is around 8 dB of insertion loss.

The asymmetry essay showed that a complete DR link budget is about 3 dB. So the conclusion is clear: plasmonics uses up the one budget that an intensity-modulated link does not have. Physics this good will find uses, in test equipment, in cryogenics, and possibly beyond IM-DD (intensity modulation with direct detection). The data hall is not one of those uses yet.

Nobody has to win

Put the contenders side by side, and a pattern is more useful than any ranking. Every material is the best on one axis and fails on another:

• Silicon has the manufacturing but cannot reach the bandwidth.
• Niobate has the bandwidth but is still building its factories.
• Tantalate has the thermal advantage.
• Titanate has the coefficient, if it holds at high frequency.
• Plasmonics has the speed but pays in loss, the wrong budget.

In a design space like this, the usual best move is not to select one material. It is to stop forcing one material to do every job. The field is moving in exactly this direction. Bond a strip of lithium niobate onto a silicon photonic platform, and let each crystal do what its symmetry permits.

Silicon supplies the waveguides, splitters, couplers and detectors, the full manufacturable set of parts that it does best. The niobate supplies the one thing that silicon cannot: a clean linear phase shift. It does this only in the few millimeters where modulation occurs. Demonstrated hybrids reach 110 GHz, and another hybrid device reports 1.8 dB of on-chip loss.

A folded design puts enough electrode length into a chip under nine millimeters long to reach a 1.24-volt half-wave voltage. Its voltage-length product is worse than monolithic TFLN, around 3.1 V·cm. That is the cost of keeping the rest of the die in the best factory system on earth. Manufacturing usually decides, and manufacturing prefers this version.

The whole interconnect roadmap follows one rule: give each particle the job that suits its statistics. Electrons compute and photons carry. The modulator is the component where those two meet, and it follows the same rule one level down. The result is not a war of materials but a division of labor between crystals. Each crystal does the job that its symmetry permits, and only that job.

What to take away

1. Silicon’s inversion symmetry forbids the Pockels effect. So silicon modulates by moving charge, and pays permanently in loss, chirp and a bandwidth limit near 60 GHz.
2. Compare materials on Vπ·L, and remember the square: half the voltage gives a quarter of the driver power.
3. Thin-film lithium niobate is the physics successor, and its real frontier is manufacturing: DUV on 200 mm, yields that must rise, and a cost gap to silicon photonics that must still close.
4. Watch lithium tantalate for co-packaged optics, for one main reason: bias stability near 120 °C, on a wafer supply chain that the phone industry already built.
5. The likely result is assignment, not victory: silicon everywhere it does well, and a Pockels crystal bonded in only where modulation occurs. Watch the hybrid demonstrations, not the material press releases.