Thin-film lithium niobate (TFLN) is now doing real work in the optical transmitter. Packaged TFLN modulators are shipping, 400G-per-lane demonstrations are on record, and a full 3.2 Tbps experimental link has been published. That progress raises a practical question for anyone who owns a fiber plant: does 3.2T mean the cabling has to change?
- There is no such thing as "TFLN fiber." TFLN is a modulator material. A TFLN transmitter launches light into ordinary single-mode fiber.
- Complete, interoperable 3.2T pluggable modules are not yet a purchasable commodity. Components and reference designs exist; a finished IEEE standard does not.
- Installed OS2 fiber is often reusable - but that is decided by the module's interface specification, connector count and channel loss, never by the "OS2" label alone.
- What drives your cabling decision is DR8 versus FR8, not TFLN versus silicon photonics. The modulator material barely touches the passive plant.
Where 3.2T Actually Stands in 2026
Most confusion about 3.2T comes from mixing technology layers that are at very different stages of maturity.
| Layer | Status | Evidence type |
|---|---|---|
| TFLN modulators and PICs | Commercially available | Manufacturer specification |
| 400G-per-lane device demonstrations | Demonstrated | Company demonstration / conference paper |
| 1.6T TFLN reference transceiver designs | Demonstrated | Manufacturer announcement |
| Net 3.2 Tbps eight-lane transmission | Demonstrated in the lab | Post-deadline conference paper |
| Interoperable, volume 3.2T pluggable modules | Early / not yet a market | Absence of standard and of shipping product |
| IEEE 3.2T Ethernet standard | Not complete | Study group formed March 2026 |
That last row matters more than any product announcement. On 13 March 2026 the IEEE 802.3 working group chartered the 400 Gb/s per lane signaling study group to develop a project authorization request covering electrical interconnects and single-mode optical interconnects out to roughly 500 metres. A study group writes the case for a project; it does not publish a PHY. The Ethernet Alliance's own commentary describes the effort as deliberately scope-limited so the highest-pressure AI applications can be addressed first.
So when a roadmap slide shows 3.2T as eight lanes of 400G, read it as an architectural expectation with strong industry momentum - not as a finished interface you can specify in a tender document today.
What TFLN Is, and What It Is Not
Lithium niobate has been used in optical communications for decades for its strong electro-optic effect, low loss and excellent modulation linearity. The limitation was always physical size: bulk lithium niobate modulators are too large for dense pluggable transceivers.
TFLN bonds a very thin lithium-niobate layer onto another substrate - typically silicon or silicon dioxide - so the light can be confined in tightly patterned waveguides. The result is a photonic integrated circuit that keeps the electro-optic behaviour and loses most of the bulk. In practice that buys high electro-optic bandwidth, low drive voltage, high linearity, and support for both intensity modulation and coherent formats.
What it does not buy is anything on the fiber side. TFLN sits inside the transmitter. The photons that leave the module are the same photons any other O-band transmitter produces, and they travel through the same G.652.D single-mode fiber you already have installed. Keep that boundary clear and most "does TFLN change my cabling" questions answer themselves.
Why 400G per Lane Is Genuinely Hard
A 1.6T module is typically eight lanes of about 200 Gbps. Getting to 3.2T on the same lane count means doubling each lane, and the difficulty does not scale linearly.
The bandwidth requirement follows directly from the symbol rate. As a rule of thumb, the analog bandwidth needed for a PAM4 signal sits somewhere around 0.45 to 0.55 of the baud rate. A 400G-class PAM4 lane runs at roughly 225 GBaud, which puts the required device bandwidth in the region of 100 to 125 GHz - for the modulator, the driver, the photodetector, the TIA and the DSP alike. That is why the number "about 100 GHz" keeps appearing; it is not a marketing figure, it falls out of the baud rate.
Two vocabulary traps are worth clearing up here, because they make otherwise consistent sources look contradictory:
- 400 Gbps per lane normally refers to the net payload rate.
- 448 Gbps per lane normally refers to the physical line rate including FEC and coding overhead.
They describe the same lane. Higher symbol rates also tighten sensitivity to chromatic dispersion, laser linewidth, electrical interconnect loss, reflections and jitter - which is precisely why the modulator platform became a competitive battleground.
Has TFLN Already Carried 3.2T?
At the experimental level, yes, and the evidence is specific enough to check.
A team from Ciena, HyperLight and McGill University reported an eight-lane O-band IM/DD system running 225 GBaud PAM4 through TFLN modulators driven by a 3 nm CMOS SerDes, in both FR8 and DR8 configurations, over 2 km of fiber, below the hard-decision FEC threshold - a net 3.2 Tbps, with 4.2 Tbps reached using PAM8. The work was presented as a post-deadline paper at OFC 2025 and is catalogued by Optica as paper Th4B.1; Ciena's accompanying announcement adds the test-setup detail, including the packaged 125 GHz reference modulators used for FR8 and an integrated eight-modulator DR8 chip.
Separately, HyperLight released a 145 GHz packaged intensity modulator in March 2026 aimed at 448 Gbps-per-lane IM-DD and 260 GBaud coherent work.
Read those two facts at the right resolution. The first is a peer-reviewed system result: the physics and the device bandwidth are not the obstacle. The second is a vendor product release: real, purchasable, and still a component rather than a qualified transceiver. Between a working experiment and a module you can buy in volume sit yield, thermal performance, power per bit, packaging, long-term reliability, DSP and switch interoperability, standards compliance and supply chain - and none of those is a small step.
Can Existing OS2 Cabling Support 3.2T Optics?
Often, yes - but "OS2" is a fiber category, not a compatibility guarantee. Compatibility is determined by whether the installed channel meets the loss, reflectance, reach and connector requirements of the specific interface, as defined against ITU-T G.652 fiber attributes plus the module's own PMD specification.
Here is where a number makes the point better than a checklist. Consider an illustrative 100-metre DR8 channel inside a hall, O-band, with four mated MPO pairs between the two transceivers:
- Fiber attenuation: 0.35 dB/km × 0.1 km ≈ 0.035 dB
- Connector loss: 4 mated pairs × 0.35 dB ≈ 1.4 dB
- Channel total ≈ 1.44 dB, of which the fiber contributes about 2%.
This is a worked example, not a measurement - substitute your own connector grades and it still lands in the same place. In short-reach data center links, connector count and connector condition dominate the loss budget by an order of magnitude over fiber length and fiber grade. The practical consequence: when you audit for 3.2T readiness, count mated pairs before you argue about cable. A channel with two extra patch points is a bigger risk than a channel that is 40 metres longer.
The exception is reach. Past a few hundred metres, or wherever the interface leans on a dispersion-limited window, fiber attributes stop being a rounding error and the full channel - length, wavelength, splices, reflectance, environment - has to be evaluated against the module datasheet.
DR8 or FR8: The Choice That Actually Drives Your Cabling
The modulator material is invisible to your patch panel. The lane architecture is not. Both DR8 and FR8 appeared in the 3.2T demonstration, and they impose very different physical plants.
| DR8 (parallel single-mode) | FR8 (wavelength multiplexed) | |
|---|---|---|
| Fibers per link | 16 (8 Tx + 8 Rx) | 2 (duplex) |
| Connector | MPO/MTP multi-fiber | Duplex LC or equivalent |
| Wavelengths | 1 | 8 |
| Typical reach target | Shorter, in-row / in-hall | Longer, cross-hall |
| Cabling complexity | High fiber count, polarity management critical | Low fiber count, optical complexity moves into the module |
| Migration cost driver | Trunk capacity and panel density | Module cost and wavelength stability |
Neither is better in the abstract. What matters is that they pull your infrastructure in opposite directions: DR8 pushes cost into MPO/MTP trunks and cassettes and into disciplined polarity documentation, while FR8 keeps the fiber plant thin and pushes cost into the optics. If you are deciding today what to install for a network you will populate in three years, this - not the modulator roadmap - is the decision that binds you.
If you have already been through an 800G build, the fiber-count arithmetic will look familiar; the same tradeoffs are laid out in more depth in this 800G MPO guide for AI data centers.
Which Requirements Come From TFLN, and Which Do Not?
This table is the single most useful thing to keep on hand when a vendor tells you a new modulator technology requires a cabling change.
| Question | Decided by TFLN? | Decided by module / interface architecture? |
|---|---|---|
| Whether OS2 fiber can be reused | No | Yes |
| Whether MPO/MTP is required | No | Yes |
| Fiber count per link | No | Yes |
| Modulator bandwidth and linearity | Yes | Partly |
| Drive voltage and transmitter power | Yes | Partly |
| Achievable link distance | Indirectly | Yes |
| Optical loss budget | Indirectly | Yes |
Read down the first column: TFLN determines almost nothing you can touch with your hands. That is not a criticism of TFLN - it is the reason a cabling plant designed for a range of interfaces will outlive several generations of transmitter technology.
An Audit Sequence, in Priority Order
Most operators do not need new cabling. They need to avoid installing something that blocks migration, and they need to know where they actually stand. Work in this order - the sequence matters, because each step can make the next one unnecessary.
- Get the target module specification first. Everything downstream - loss budget, connector type, reach, wavelength plan - is derived from it. Auditing a plant before you know which interface you are auditing against produces numbers you cannot interpret.
- Count mated pairs per channel. As shown above, this is usually the dominant term. If a channel carries five or six patch points for a 100-metre run, that is the finding.
- Measure insertion loss and reflectance on real channels, not on components in isolation, and compare against the module's stated budget. Return loss deserves particular attention at high symbol rates.
- Verify polarity documentation. Parallel-optics migration fails on undocumented polarity far more often than it fails on optical performance. If your trunk and cassette records are inconsistent, fix that before anything else - it is cheap now and expensive during a cutover.
- Inspect and clean end faces to a defined standard. IEC 61300-3-35 gives quantitative pass/fail criteria for debris, scratches and defects, which is what turns "the connectors look clean" into a repeatable process. Contamination on multi-fiber ferrules is a recurring source of intermittent link errors; this guide to MPO end-face contamination covers the failure modes in detail.
- Then, and only then, look at pathway and panel capacity. Higher-capacity switches usually mean more links, not fewer - faster optics rarely reduce total fiber count in an AI fabric.
Steps 1 to 3 tell you whether you have a problem. Steps 4 to 6 are what you fix. Reversing that order is how organisations end up replacing cable that was never the constraint. If you are planning a fabric rather than auditing one, our data center connectivity solutions cover the trunk, cassette and panel architecture side of the same problem.
Will TFLN Win?
TFLN is a leading candidate, not a settled outcome. Its bandwidth, linearity and low drive voltage are genuinely well matched to 400G-per-lane requirements and to the power budgets that come with them.
But modulator performance has never been the only criterion. Manufacturing yield, packaging reliability, module power, supply stability, interoperability and cost per bit decide which platform ships in volume. Silicon photonics, indium phosphide and differential EML approaches are all being pushed toward the same lane rate, and the contributions submitted to the IEEE study group treat several of these as viable for 400G-per-lane PAM4. Hybrid optical engines combining materials are also plausible.
For cabling planners this competition is, usefully, almost irrelevant - and that is the point of the decision table above. A 400G silicon photonics link and a 400G TFLN link impose broadly the same demands on your passive plant, because those demands come from the lane rate and the interface architecture, not from the material inside the transmitter.
FAQ
Q: Does A TFLN Transceiver Require Special Optical Fiber?
A: No. TFLN describes the modulator inside the transmitter. The link runs over whatever fiber the interface specification calls for - normally single-mode.
Q: Are 3.2T TFLN Modules Available To Buy?
A: Not as interoperable pluggables. TFLN modulators, PICs and reference transceiver designs are commercially available, and 3.2T-class transmission has been demonstrated in the lab. Complete 3.2T modules remain an emerging category, and the corresponding IEEE project has not produced a standard. Anything quoted as a "3.2T module" today should be checked for which interface it claims to implement and whether that interface exists in draft form.
Q: Can Existing OS2 Fiber Support 3.2T Transmission?
A: Frequently, yes. The determining factors are the module's loss budget, connector type and count, operating wavelength, reach and the measured condition of the channel - not the fiber category label. A 100-metre channel with four clean mated pairs is a very different proposition from a 100-metre channel with six dirty ones, even though both are "OS2."
Q: Will 3.2T Modules Use MPO/MTP Connectors?
A: Parallel implementations such as DR8 will. Wavelength-multiplexed implementations such as FR8 can run over duplex fiber. Since the interface is not standardised yet, the honest answer is that a flexible plant should be able to serve both - which in practice means modular cassettes and documented polarity rather than committing to a single connector strategy.
Q: Does Moving To 3.2T Reduce The Amount Of Fiber We Need?
A: Usually the opposite. Per-transceiver efficiency improves, but AI cluster scale-out adds links between accelerators, leaf and spine faster than module capacity absorbs them. Plan pathway and panel capacity for growth in link count, not for a reduction.
Q: What Should We Fix First If The Budget Is Limited?
A: Polarity documentation and end-face cleanliness. Both are low-cost, both are the most common causes of failed parallel-optics migrations, and neither becomes cheaper if you defer it.
Conclusion
TFLN has moved from research to a production-oriented photonic platform, and the published 3.2 Tbps result shows the modulator is not what stands between the industry and 400G-per-lane links. What still stands in the way is standardisation, qualification and volume manufacturing - and none of that changes the physical medium.
For the fiber plant, the useful framing is this: the modulator material sets what is possible inside the module; the interface architecture sets what you have to build outside it. Audit against the interface, count your connectors, document your polarity, and keep the plant flexible enough to serve both parallel and multiplexed optics. Those decisions will still be right whichever modulator technology wins.