Making terabit optics: What 400G per lane really demands from the photonic engine

As AI infrastructure pushes toward 400G per lane and beyond, the photonics industry faces a critical transition in materials, modulation formats and manufacturing capacity. Thin-film lithium niobate is emerging as a strong contender for next-generation optical engines, but its success will depend as much on foundry readiness and supply-chain maturity as on device performance.

By Frédéric Loizeau Co-Founder & CRO, Lightium

The bottleneck driving everything
As GPU clusters have grown from thousands to hundreds of thousands of accelerators, the optical fabric connecting them has become the limiting variable for continued scaling. More bandwidth per fibre, more reach per lane, more density per rack: interconnects have become the bottleneck of AI infrastructure, and pressure to deliver more comes from every direction.

Three distinct connectivity layers define how AI infrastructure actually moves data (Figure 1). Scale-Up is the tight interconnect between accelerators within a single compute domain, where latency and bandwidth density are essential. Scale-Out is the intra-datacentre fabric, the Spine/Leaf networks connecting racks and buildings, increasingly spanning distances that approach 2 km as datacentre footprints grow. Scale-Across links connect physical facilities on a shared campus, operating over distances of 2 to 20 km. All three layers are expanding in reach and throughput simultaneously, and all three are converging on the same bottleneck: per-lane optical speed.

Figure 1: AI data center interconnect schematics. Scale-up: intra-rack links below 500 m. Scale-out: rack-to-rack and rack-to-building links up to 2 km. Scale-across: inter-data center connectivity from 2 to 20 km.

Optical modules are inherently parallel: Total throughput equals lane speed multiplied by lane count. Doubling the per-lane speed halves the required number of lasers, modulators, and fibres at a given aggregate bandwidth, reducing cost, improving packaging density, and lowering system power. The industry is now transitioning from 200G to 400G-per-lane. That shift will simultaneously reshape modulation schemes, material platforms, and manufacturing capacity.

What won at 200G, and why
Understanding the transition requires treating two dimensions separately: the modulation scheme and the material platform.

On the modulation side, the 200G generation in datacentres is built almost entirely on IMDD (intensity modulation with direct detection), using pulse amplitude modulation (PAM4) at 116 GBaud. It encodes data in optical intensity; the receiver measures optical power directly, yielding a structurally simple, low-cost, power-efficient link. Coherent modulation, on the other hand, simultaneously encodes data in amplitude, phase, and polarization, thereby requiring a local-oscillator laser at the receiver. Spectral efficiency is approximately four times that of IMDD, but cost and power are substantially higher. In the 200G era, coherent is confined to long-haul DCI and telecom. It has no presence within the data centre’s short reach.

On the material side, four platforms are relevant (Table 1):
InP electro-absorption modulated lasers (EML) integrate a DFB laser and an electro-absorption modulator on a single die. They are the 200G workhorse: at 116 GBaud, the modulator operates comfortably within its bandwidth envelope, and the monolithic laser integration is mastered at high volume, making the device compact and mature.

Table 1: Photonic platform capabilities and application fit at 400G per lane

InP Mach-Zehnder modulators (MZMs), either monolithically integrated with the laser, SOA, and detectors or combined with silicon photonics for the receiver part, enable full coherent modulation, but the cost and integration complexity exceed the 200G IMDD requirements.

Silicon photonics (SiPh) uses CMOS-process waveguides on silicon wafers with an external InP laser source. The availability of germanium detectors makes silicon photonics very attractive for the receiver function, and recent progress on SiGe-based electro-absorption modulators (EAMs) opens a path towards 400G-per-lane IMDD. The 300 mm wafer is the other key manufacturing advantage, while the MZI modulator bandwidth is adequate for 200G. While transceivers were commonly introduced by the more costly InP-based platforms, silicon photonics was used to significantly reduce costs at 200G.

Finally, TFLN leverages the exceptionally strong electro-optic Pockels effect; modulation bandwidth exceeds 100 GHz, and drive voltage is sub-1V, but industrialization was not mature enough to catch the 200G deployment window. It also lacks detectors natively, so TFLN is currently confined to the transmitter side of the photonic engine.

In the 200G landscape, InP EML and SiPh capture virtually the entire datacentre short-reach IMDD market. InP modulators serve long-haul coherent while TFLN was not mature enough to enter the market. The dominant platforms won because 200G IMDD made relatively modest material demands: 50 GHz of electro-optic bandwidth, manageable 2-3 V drive voltages, and forgiving link budgets for sub-2 km reach. At 400G, that room disappears entirely.

Why 400G is a fundamentally different problem
The 200G-to-400G transition significantly raises the bar for device bandwidth, link budget, and modulation efficiency. The core physical problem is dispersion. IMDD at 400G PAM4 requires roughly 226 GBaud, approximately double the 200G rate, and chromatic dispersion tolerance scales inversely with the square of the baud rate. Hence, IMDD reach at 400G per lane collapses to under one kilometre at moderate link loss.

The Scale-Across segment (2 to 20 km) that IMDD covered confidently at 200G is now stranded, and coherent modulation must be considered. However, deploying a tunable laser, a dual-polarisation IQ modulator, and a full four-dimensional DSP for a 2-20 km datacentre campus link solves the wrong problem. The cost, power, and complexity overhead of conventional coherent links are unjustified. The correct architecture is Coherent Lite: Single-polarisation IQ modulation at 113 GBaud, the same baud rate as today’s 200G IMDD PAM4, with a fixed DFB replacing the tunable laser and a simplified receiver architecture. No breakthrough in device bandwidth required. But IQ modulation is now the requirement, and that single change rewrites the material selection logic entirely.

What 400G will be built on
InP EML is the natural home for 400G IMDD. Demonstrations have progressed rapidly, from the first 400G EML at OFC 2025 to a 99 GHz electro-optic bandwidth result presented at OFC 2026. The material boundary, however, is fixed by physics. The electro-absorption mechanism modulates only intensity. There is no IQ modulation path from the InP EML; therefore, there is no Coherent or Coherent Lite path.

Figure 2: Wafer-scale electro-optical characterization of TFLN dies on a 200 mm wafer. Measuring every die in-situ before dicing enables known-good-die selection and provides the statistical yield data required by a PDK-based design flow.

InP MZM modulators can perform IQ modulation. Monolithic integration of the laser, IQ modulator, and detector on a single InP die provides the functionality Coherent Lite requires. The constraint is supply. InP PICs are significantly larger than EMLs, roughly 100-150x larger. InP wafers are predominantly produced in 3- to 4-inch diameters, with 6-inch being on the horizon. Small wafer and large die: the volume ceiling is hard, and is amplified by geopolitical influences and export restrictions.

The supply constraint deepens when you consider that InP EML and InP PIC draw from the same substrate supply. As 400G IMDD deployment accelerates, EML demand will grow substantially. This is not a device performance question. It is a wafer economics question, and it creates a structural opening for alternative modulator platforms in the Coherent Lite segment that InP PICs will find difficult to fill at volume.

Figure 3: A 200 mm LNOI wafer. The transition from 150 mm to 200 mm provides roughly 2.3× more dies per wafer and opens access to the global installed base of high-volume silicon foundries.

Silicon photonics faces its own material transition at 400G. The plasma dispersion effect that underpins silicon MZI modulators reaches fundamental bandwidth and drive-voltage limits well before 226 GBaud. For Coherent Lite, silicon photonics is best suited to the receiver side thanks to the germanium photodetectors. On the transmitter side, it requires an external laser and a high-performance IQ modulator made of another material. The silicon photonics community has responded with heterogeneous integration, using GeSi or III-V absorbers on the silicon platform to recover bandwidth that plasma-dispersion silicon cannot reach. IMEC’s demonstration at OFC 2026 of a >110 GHz GeSi electro-absorption modulator on a 300 mm silicon photonics platform is a credible result on this path to date. The architectural boundary, however, is the same as for InP EML: a GeSi EAM modulates only intensity. It addresses 400G IMDD on silicon but has no path to IQ modulation and therefore no path to Coherent Lite. SiPh at 400G is evolving into an integration substrate, with the high-performance modulation function delegated either to GeSi absorbers for IMDD or to alternative material platforms for coherent applications.

TFLN is in a different category entirely. The electro-optic Pockels effect in lithium niobate creates a direct linear relationship between applied electric field and refractive index shift, with no charge carriers involved. The response is instantaneous and temperature independent. The result is a set of modulator performance parameters that no other platform simultaneously achieves: greater than 100 GHz electro-optic bandwidth with no roll-off from carrier dynamics; drive voltage below 1V, reducing thermal load and relaxing driver IC specifications; waveguide propagation loss below 0.5 dB/cm, preserving link budget in multi-channel configurations where loss accumulates; and intrinsic linearity from DC to millimetre-wave frequencies, critical for the signal fidelity that 16QAM demands (Figure 2).

These parameters are achieved together, in a single material, without the trade-offs that constrain every alternative. The architecture consequence is direct: a TFLN Mach-Zehnder modulator can function as a single-arm intensity modulator for PAM4 IMDD or be configured as an IQ modulator for SP-16QAM Coherent Lite, using the same waveguide platform and fabrication process. No other modulator material covers both 400G pathways without a fundamental compromise.

Maturing TFLN supply chain enables volume production
TFLN missed the 200G window. Not because of what the material could deliver, but because of what the foundry ecosystem could not. Substrate supply came from a small number of vendors, wafer sizes were limited to 150 mm, and the fabrication toolsets available at that diameter were research-grade: capable of producing excellent individual devices, but unable to deliver the process uniformity, throughput, and statistical characterization that production qualification demands. A module manufacturer qualifying a new component needs wafer-scale yield data, lot-to-lot consistency, and a PDK grounded in manufacturing statistics rather than nominal values from a handful of devices. None of that was available from 100- or 150-mm TFLN research lines at universities or RTOs.

Two structural shifts are changing this: First, 200 mm TFLN wafers are now available and compatible with production-grade semiconductor toolsets (Figure 3). The area advantage matters for economics, but the more important change is process control. Production equipment qualified on 200 mm substrates delivers within-wafer and wafer-to-wafer uniformity that is quantitatively superior to that of 150 mm research toolsets. Electro-optic bandwidth histograms, Vpi distributions, and propagation loss maps measured across a full production wafer provide the statistical foundation a reliable PDK requires and the qualification evidence a procurement team needs before committing volume to a new platform. Second, the smart cut TFLN wafer supply base is broadening. Multiple vendors are now producing wafers at commercially relevant quality levels and at quantities to satisfy the present and future demand of the industry. In photonics supply chains, single-vendor substrate dependency is a qualification risk that no serious customer accepts. A multi-vendor wafer market resolves that exposure and signals a level of ecosystem maturity that 150 mm TFLN never reached.

Co-design is not optional at 100+ GHz
The third shift is structural to the industry: The emergence of pure-play open-access TFLN foundries. At 400G-per-lane and 100 GHz modulation bandwidth, photonic circuit design cannot be decoupled from system architecture. Drive voltage, differential versus single-ended topology, RF electrode geometry, packaging approach, and co-packaging strategy all interact directly with modulator circuit performance. The coupling between the photonic circuit and its electrical and thermal environment is tight enough that a single reference modulator design, however well engineered, will not be optimal for every system architecture that deploys it.

Module manufacturers need to design their own photonic engine, tailored to their specific system context, and own the resulting IP. A foundry model that makes this possible fully separates fabrication from product design: The foundry provides a stable, well-characterized process, a statistically validated PDK, and a reliable manufacturing service. What the customer builds on that foundation belongs entirely to them. That combination of process access, design freedom, and IP ownership is not a commercial differentiator at 400G. It is a prerequisite for extracting the performance that TFLN physics makes available.

Figure 4: A 100+ GHz assembly-design-kit-compatible test package, co-developed by PHIX BV and Lightium AG.

Conclusion
TFLN missed the 200G window because the infrastructure was not yet in place. The 400G window is opening under different conditions: System-level performance is validated, 200 mm production capability is coming online, the wafer supply base is maturing across multiple vendors, and pure-play open-access foundry services are available for the first time. The InP supply chain will be pulled strongly toward EML as 400G IMDD demand scales, creating a structural opening in the Coherent Lite segment that InP PIC will find difficult to fill at volume. TFLN is the only modulator platform that serves both 400G pathways from the same material system, without compromise. The question is no longer whether the material works. It is whether the foundry ecosystem scales in time to meet the opportunity window.