A thousand gigahertz is not a product

One session, five materials, one finish line

The postdeadline session at OFC shows the industry what became possible in the previous twelve months. The conference accepts these papers days before it starts, so that it can show results too new for the normal deadline. In March 2026, the modulator papers in that session included almost every material that anyone ever proposed for the job:

• NTT put four channels of a membrane electro-absorption array on silicon. Each channel ran 448 gigabits per second at a drive swing of half a volt.
• A team from Michigan and Nokia Bell Labs ran thin-film lithium tantalate at 768 gigabits of line rate. Its half-wave voltage (the drive voltage for a half-cycle phase shift) was 1.35 volts.
• McGill, Ciena and Lumiphase ran barium titanate at 448 gigabits per lane, four lanes to a net 1.6 terabits. They integrated it monolithically on a commercial silicon photonics platform. A 3 nanometer SerDes (the serializer-deserializer circuit of a host chip) drove it.
• Thin-film lithium niobate appeared three separate times, once with gallium arsenide bonded on top of it.
• Coherent sent the one all-silicon paper. Its abstract started with the admission that silicon’s scalability to 400 gigabits per lane “is uncertain.” Then the paper demonstrated it anyway, with a commercial silicon-germanium driver.

A few days later, Tower Semiconductor and Coherent went further. They announced a silicon Mach-Zehnder modulator (an interferometer that changes phase into intensity) at 400 gigabits per lane in a production-ready process. The eye was clear and open at 420. Five materials reached the same finish line within one week.

For a decade, one question organized this field: which material can reach 400 gigabits on a single lane? That question now has an answer that does not help: several of them can. So the interesting question moved, and it moved to the rest of the datasheet.

The number everyone quotes

Start with the highest number in the field, because it shows the gap most clearly.

In 2025, an ETH Zurich team measured a plasmonic modulator with a 3 decibel electro-optic bandwidth of 997 gigahertz. This is not a typo and not a stunt. A companion result showed a response flat within three decibels from 75 megahertz to 500 gigahertz. Plasmonics puts light into a metal slot a few hundred nanometers wide. In that slot the electric field is very large, and the device can be ten microns long instead of several millimeters. On the bandwidth axis no other route is close, because plasmonics works in a different regime.

Now read the datasheet of the product from the same technology. Polariton, the ETH spinout that commercialized it, published a packaged Mach-Zehnder part with these values:

• bandwidth “greater than 145 gigahertz”
• insertion loss of 16 to 17 decibels
• maximum optical input of 3 decibel-milliwatts
• operating temperature given as a single point, at 25 Celsius.

Its reliability page lists Telcordia GR-468 (the reliability standard for optoelectronic devices in telecom equipment) as a goal. The page marks high-temperature operating life data as coming soon. Polariton also claims a ring-resonator variant at 1.2 decibels and beyond 400 gigabits per lane. So the loss comes from the geometry, not from the material. But the shipping Mach-Zehnder is the part with a datasheet.

The loss is not a packaging artifact that engineering can remove. It follows from first principles: about half a decibel for each micron of plasmonic slot, plus roughly 1.1 decibels for each photonic-to-plasmonic converter. That is why the 2025 result that ran 400 gigabits over a kilometer reported 8 decibels on chip and 14 fiber to fiber. Silicon and thin-film lithium niobate modulators have a loss between 2 and 6. A link budget starts with about 3 decibels of channel allowance. In that budget, ten decibels is not a specification detail but the whole argument.

In April 2026 Marvell acquired the company. Two months later, two of its founders wrote about plasmonics as Marvell engineers. They said that it offers roughly a terahertz of bandwidth in a ten micron device. Then they wrote that the technology is “still in the early days of development for commercialization.” They named reliability, performance and manufacturing scalability as the work ahead. The bandwidth argument was won, and it did not settle anything.

That is the pattern of this entire race.

Five gates, and the last three are nearly empty

If bandwidth no longer separates these routes, something else must. A useful test asks what a buyer needs before signing. Each gate must be binary, and a public document must answer it. This is better than an invented readiness score. Five gates do the work:

• A published device at or above 100 gigahertz.
• A demonstrated 400 gigabit lane.
• Availability in a production PDK, or in an open multi-project wafer run (a shared run where many customers split one wafer) that anyone can buy.
• A reliability qualification published with an actual test matrix.
• Revenue from datacom. Coherent transport and test instruments do not count, although several of these materials have real businesses there.

Count the field of nine against these gates, and the result is clear:

• Six routes have a 100 gigahertz device.
• Seven showed a 400 gigabit lane.
• Four are in some foundry’s PDK.
• Two published a qualification.
• Two earn datacom revenue.

Most routes pass the gates that a conference measures. Few routes pass the gates that a purchase order measures.

Most people underestimate the foundry gate. EUROPRACTICE is the service through which most of the world buys prototype silicon. Its 2026 photonics catalog offers silicon-on-insulator (a thin silicon layer on oxide on a silicon wafer), silicon nitride and GlobalFoundries’ 45 nanometer platform. It lists no thin-film lithium niobate, no barium titanate, no polymer and no bulk lithium niobate. Zero of four.

The exotic routes that did reach foundries got there as a back-end module on the silicon flow of another company. Polariton’s plasmonics reached imec’s 200 millimeter platform this way. Polymer is moving into Tower’s kit the same way.

Lithium niobate is the exception, and it shows the cost. Lithium contaminates a CMOS line. So a fab must bond lithium niobate at the last step, or a separate foundry industry built for it must process it. The largest published open run in that industry has a minimum die of about 10 by 5 millimeters.

The qualification gate is even thinner, and the contrast is useful. OpenLight published a Telcordia GR-468 qualification for every active component in its PDK on Tower’s platform. It named the test list. Two lots went past 5,000 hours and one went past 15,000, with zero failures. That is what a written qualification looks like.

HyperLight, the most visible thin-film lithium niobate company, describes “rigorous Telcordia qualification” in prose. It gives no standard number, no test matrix, no hours and no lot count. At the time of writing, no vendor, foundry or trade outlet published a completed GR-468 qualification for a thin-film lithium niobate modulator.

The qualification gate itself has a detail worth noticing. The Telcordia document that actually governs optics for information-handling equipment is GR-3013. Its title contains the words “Short-Life,” and Telcordia issued it in December 1999. GR-468 is its longer-life sibling. Its last revision was in September 2004, and a copy costs $795.

So the qualification of a 3.2 terabit module in a liquid-cooled rack uses paperwork written for carrier equipment before the first iPhone. Nobody in this race says that the standard is wrong, and that fact tells you something. An unrevised standard is a stable target, and an incumbent wants a stable target.

Nine routes to 400G on one lane

Sort by the number a press release quotes, by the number a production process states, or by how many of the five commercial gates a route has actually cleared. The order changes almost completely.

Sort by

Lab bandwidthProduction bandwidthGates cleared

Inspect route

SiliconIndiumThin-filmLithiumBariumPlasmonic-organicElectro-opticFerroelectricGraphene

• Plasmonic-organic hybridPolariton997 GHz400G per lane
• Thin-film lithium niobateHyperLight145 GHz448G per lane
• Silicon Mach-ZehnderTower110 GHz420G per lane
• Lithium tantalateLuxtelligence110 GHz768G per lane
• Electro-optic polymerLightwave Logic110 GHz240G per lane
• Indium phosphide EAM and EMLLumentum106 GHz448G per lane
• Barium titanate on siliconLumiphase70 GHz448G per lane
• Ferroelectric nematic glassPolaris Electro-Optics67 GHz400G per lane
• GrapheneCamGraPhIC67 GHz80G per lane

Plasmonic-organic hybrid

Polariton, acquired by Marvell · Product or platform published

Lab bandwidth

997 GHz

best published device

Production bandwidth

145 GHz

PDK or datasheet

Best lane rate

400G

ECOC 2025, over 1 km of fiber

Insertion loss

16 dB

shipping datasheet, C-band MZM

• ✓ 100 GHz in a lab
• ✓ 400G on one lane
• · In a foundry PDK
• · Qualification published
• · Datacom revenue

Metal absorbs about half a decibel per micron of slot, so the loss grows with exactly the length the drive voltage needs.

Lab figures are peer-reviewed or conference results; production figures come from a foundry PDK or a product datasheet, and the two are never mixed. The gates are binary and each is answerable from a public document: a published device at or above 100 GHz, a demonstrated 400 Gbps lane, availability in a production PDK or open multi-project wafer run, a reliability qualification published with a test matrix, and revenue from datacom rather than coherent or instruments.

The wafer decides, and lithium tantalate proves it

Below the gates is an economic argument, and the physics literature is now unusually direct about it. In a 2024 Nature paper, Kippenberg’s group wrote that industrial use of thin-film lithium niobate “is hindered by the high cost per wafer and the limited wafer size.” Then the paper named the reason. The high cost “results from the lack of existing high-volume applications in other domains of the sort that accelerated the adoption of silicon-on-insulator photonics.”

Suppliers make more than three million silicon-on-insulator wafers every year, at diameters up to 300 millimeters. Lithium niobate on insulator has no equivalent customer to pay for the same climb down the cost curve.

The same paper gives a control experiment, and it is close to perfect. Lithium tantalate has an electro-optic (Pockels) coefficient of 30.5 picometers per volt. Lithium niobate has 30.9. For a modulator designer, they are nearly the same crystal. But lithium tantalate has two advantages:

• The 5G filter industry uses it in volume. The paper cites a projected capacity of about 750,000 lithium tantalate on insulator wafers a year.
• It is easier to cleave. Its smart-cut process (implant ions, then split a thin film off the crystal) favors hydrogen implanted at 100 kiloelectron volts. Lithium niobate needs helium above 200, at implant currents ten times lower.

So a completely unrelated industry already buys the cheaper, larger, more available wafer. Optical grade is not acoustic grade. A cofounder of one lithium tantalate company says so plainly, together with the good news. But the direction holds: filter volume buys a cheaper boule and a shorter climb down the cost curve.

This is the main point of this essay. Two materials have the same Pockels coefficient. The one with a consumer volume driver arrived at 150 millimeters and a published foundry PDK. The other still negotiates for wafer supply. The material did not change. The volume application did.

The audited number

Every route in this race has a slide that claims a path to volume. One route files audited financial statements, and they are worth more than all the slides.

Advanced Fiber Resources is a listed Chinese optics maker. It committed 540 million yuan to a lithium niobate high-speed modulator chip program and spent about 71 percent of it. Its FY2025 annual report gives these results for the program:

• realized benefit for the year: negative 7.94 million yuan
• cumulative benefit: negative 27.98 million
• the answer to the standard disclosure question, “did the project meet its expected benefit?”: a plain no.

The filing states the cause in its own words. Downstream customer orders fell short of expectations, and cumulative capacity utilization stayed below the planned ramp. Its datacom part is a PAM4 (four-level pulse amplitude modulation) thin-film lithium niobate chip, for the AI links that this whole race is about. The report describes it as in sampling, with delays in customer development and product validation.

This is not an argument that the technology fails. The demonstrations are real, and they get better every March. The report is an auditor’s measurement of the distance between a demonstration and an order book.

The precedents are similar:

• In 2010, GigOptix announced that it was moving polymer modulators “to production.” Later company material claimed stable operation for 25 years at 85 Celsius. Eight years after the 2010 announcement, the entire commercial polymer modulator IP from that lineage, fifteen patents, changed owners for $350,000.
• Lightwave Logic, founded in 1991, booked $236,855 of net sales in fiscal 2025 against a $20.3 million net loss. It holds about $96 million in cash and guides volume production to 2027.
• Indium phosphide is clearly a better modulator material than silicon. It spent two decades in vertically integrated companies. Its story ended as a $2.3 billion absorption of Infinera into Nokia.

So far, an early start in a better material cost money and did not make it.

The fourth answer: change the system, not the material

One route does not appear in the materials debate at all, and its existence says something about that debate.

Aloe Semiconductor builds silicon photonic ICs for dual-polarization intensity modulation with direct detection. Light has two polarizations. An ordinary intensity-modulated link uses one and discards the other. A link that uses both carries twice the data on one fiber. It needs no higher symbol rate and no additional lasers.

That is how the company offers a 400 gigabit lane built from 112 gigabaud parts. This is the same baud rate that the 200 gigabit generation already uses. The silicon platform stays compatible with existing SerDes and with linear-drive optics.

An honest engineering check is important here. The dual-polarization work has peer-reviewed postdeadline papers and a journal article behind it. The company’s device-level modulator claims rest on booth eye diagrams, not on a published bandwidth measurement. Judge the two claims separately.

Microrings (small ring resonators) are a second form of the same idea. They trade a resonance for the length and drive voltage of a Mach-Zehnder. TSMC made that trade in the modulator inside its co-packaged optics platform. Higher-order modulation with heavier forward error correction is a third form.

These routes exist because of one number. Serial 400 gigabit PAM4 needs a modulator in the neighborhood of 120 gigahertz. Compare what each route delivers:

• Silicon in production has a specification of 40 to 65.
• Indium phosphide ships at 60 to 67 and reaches 106 in a lab.
• Thin-film lithium niobate packages at 110 and claims 145.

No route meets the requirement with real margin. Under exactly that condition, engineers stop trying to meet a requirement and start to change it. Aloe’s approach removes the bandwidth requirement instead of making small gains against it. That is also why it has no standards home and no second adopter yet.

What to take away

1. The material question has an answer, and the answer is “several.” Seven routes crossed 400 gigabits per lane within one OFC cycle. So a bandwidth number in a press release no longer makes anyone different.
2. Read the datasheet, not the paper. A 997 gigahertz laboratory device and a 145 gigahertz product with 17 decibels of loss are the same technology. You can buy only the second one.
3. Five binary gates sort the field: 100 gigahertz published, a 400 gigabit lane, a foundry PDK, a published qualification and datacom revenue. Out of nine routes, the counts are six, seven, four, two and two.
4. Wafer economics beat material physics. Lithium tantalate is the control experiment. It has the same Pockels coefficient as lithium niobate, and a completely unrelated 5G filter industry pays for its wafers.
5. Two sets of audited numbers put a price on the gap between a demonstration and an order book. One comes from a listed maker’s own filing, and one from three decades of polymer company financials. The gap is wide, and the real race happens in it.

For the rest of the decade, the pattern to watch is not which material wins. It is what happens to the materials that do not win:

• Marvell bought the plasmonics company.
• AMD bought a silicon photonics design house.
• GlobalFoundries bought a silicon photonics foundry.
• Nokia bought indium phosphide.

Other companies do not replace the exotic routes. They absorb them. Their physics then arrives inside the product of another company, two generations later than the demonstration suggested.

Three specific signals would show that this cycle is breaking:

• the first published Telcordia qualification of an exotic modulator, with its test matrix attached
• the first appearance of one of these materials in a mainstream multi-project wafer catalog
• the day when suppliers stop listing an 8-inch lithium niobate wafer as a product under development.

Until then, the fastest modulator in the world stays what it was for a decade: a paper, waiting for a supply chain.