Deep|OFC 2026 Preview: 400G per Lane — The Next Major Inflection in Optical Interconnects
We have been steadfast supporters of Optics since a year and a half ago, publishing in-depth Optics reports almost every month. Regarding this OFC and future Optics technology trends, we have a lot to discuss — especially today, when there is so much debate and speculation.
We released the GTC Preview last week — you can read them together.
AI clusters have scaled from hundreds to hundreds of thousands of accelerators, making optical interconnects a defining infrastructure bottleneck for continued cluster scaling. In 2026, the industry stands in front of the inflection point of the 200G-to-400G per lane transition — a doubling of lane speed that will simultaneously reshape modulation schemes, material platforms, and manufacturing capacity. Short-reach IMDD remains dominant, but the 1–20 km segment is opening a new market for Coherent Lite — one that was largely absent in the 200G era. This report traces the path from AI interconnect demand through the 200G/lane status quo and 400G/lane technology divergence, evaluates the competitive positioning of InP, SiPh, and TFLN material platforms in the new cycle, and maps the findings to six key industry themes and the associated value chain participants.
I.Optical Interconnects: The Core Bottleneck of AI Infrastructure
AI training clusters have scaled from hundreds to hundreds of thousands of accelerators, and optical interconnect demand has grown in lockstep. To understand this demand, it is useful to distinguish among three network dimensions.
Scale-Up refers to the high-bandwidth, low-latency interconnect between accelerators. Scale-Out covers intra-datacenter networking, spanning rack-to-rack through building-level Spine/Leaf fabrics. Scale-Across denotes inter-datacenter and cross-campus connectivity. IEEE 802.3 defines three standard reach tiers: DR (500 m), FR (2 km), and LR (10 km). DR and FR serve Scale-Out; LR maps to campus-level Scale-Across. As individual datacenters grow physically larger, Scale-Out reach is extending from hundreds of meters toward 2 km; meanwhile, multi-datacenter campus buildouts are driving rapid densification of the 2–20 km segment.
AI datacenter optical interconnect dimensions. Scale-Up: high-bandwidth accelerator-to-accelerator links. Scale-Out: rack-to-building Spine/Leaf networks. Scale-Across: inter-datacenter and cross-campus connectivity.
Multi-datacenter campus interconnect schematic. Per-building GPU counts are constrained by power delivery; 15–20 km optical links stitch multiple DCs into a unified AI factory.
Leading hyperscalers are already accelerating optical infrastructure deployment. Google uses Optical Circuit Switches (OCS) to address both Scale-Up and Scale-Out — leveraging OCS within TPU Superpods to extend accelerator interconnect scale, and on the network side to enable reconfigurable Scale-Out fabrics. OCS is transparent to the underlying optical transport technology, supporting mixed-generation transceivers on a single switching fabric; however, each OCS node introduces additional insertion loss, raising the link budget requirements for optical transceivers. OCS deployment is expanding rapidly, with adoption poised to move beyond Google into the broader datacenter ecosystem: Anthropic’s million-TPU agreement, Meta’s multi-year TPU lease, and the formation of an OCS subproject within OCP all point in this direction. NVIDIA is advancing on both Scale-Up and Scale-Out simultaneously: the NVLink domain expands from 72 GPUs (NVL72) to 576 GPUs in Rubin Ultra NVL576 (expected 2027), which will introduce optics into the Scale-Up network for the first time; GTC 2025 unveiled two silicon photonics CPO switches — Spectrum-X Photonics and Quantum-X Photonics — bringing co-packaged optical engines into the Scale-Out network, and GTC 2026 further detailed the CPO roadmap. To secure next-generation AI optical interconnect capacity, NVIDIA on March 2 invested a combined $4 billion in Coherent Corp and Lumentum and signed multiyear strategic agreements, explicitly identifying InP and SiPh as supply chain bottlenecks for next-generation optical interconnects.
All of these optical interconnect scenarios — Scale-Up, Scale-Out, and Scale-Across — are ultimately governed by the evolution of per-lane speed. Optical modules are inherently parallel: total module throughput = lane speed × lane count. Doubling the per-lane speed halves the required number of lasers, modulators, and fibers at a given aggregate bandwidth, reducing cost, improving packaging density, and lowering system power. The industry is now transitioning from 200G per lane to 400G per lane; this shift will simultaneously reshape modulation schemes, material platforms, and manufacturing capacity — the central focus of this report.
Lane Speed × Lane Width evolution. The industry is transitioning from 200G to 400G per lane, with module throughput scaling as a product of lane speed and lane count.
II. 200G per Lane: The Current Technology Landscape
2.1 Two Independent Dimensions: Modulation × Material Platform
Understanding optical interconnect technology requires decomposing the problem into two independent dimensions:
Modulation (how information is encoded onto light): IMDD vs. Coherent.
Material platform (what the devices are made of): InP EML / InP PIC / SiPh / TFLN.
These two dimensions combine freely. Modulation and material choice are not one-to-one — for example, SiPh can implement both IMDD (using MZI or MRR intensity modulation) and Coherent (using IQ modulators for phase-and-amplitude modulation). Conflating these two dimensions is a common analytical error in the market.
2.2 Modulation Schemes
IMDD (Intensity Modulation / Direct Detection): Encodes data in optical intensity; the receiver directly measures optical power. It is structurally simple, low-cost, and power-efficient. Virtually all 200G per lane deployments use IMDD + PAM4 (4-level pulse amplitude modulation, 2 bits/symbol, ~106 GBaud).
Coherent (Coherent Modulation / Detection): Simultaneously encodes data in optical amplitude, phase, and polarization; the receiver requires a local oscillator (LO) laser for coherent detection. Spectral efficiency is approximately 4× that of IMDD, but cost and power consumption are substantially higher. It has traditionally been used in long-haul telecom and DCI applications. In the table below, we compare IMDD v.s. Coherent across multiple system attributes.
Left: IMDD vs. Coherent receiver architecture. Top: IMDD — photodiode (PD) + transimpedance amplifier (TIA), minimal complexity. Bottom: Coherent — local oscillator laser (P_L), coherent mixer, and balanced photodetectors (BPD/TIA), substantially higher complexity. Right: IMDD vs. Coherent modulation efficiency. Left panel: IM-DD PAM-2 uses only intensity (real axis). Right panel: Coherent 4QAM utilizes both I/Q dimensions; detected signal strength is ~4× that of IMDD at equal peak power.
The core trade-off: IMDD is simple, low-cost, but reach-limited; Coherent offers superior spectral efficiency and reach but at higher cost. In the current 200G/lane generation, IMDD covers the full distance range from short-reach to campus 10 km (via chirp management); Coherent only becomes the default choice beyond 10 km. All of Google’s deployed datacenter and TPU interconnect optics use IMDD.
2.3 Material Platforms
InP EML: DFB laser and EA modulator integrated on a single InP die. The 200G/lane workhorse — mature and reliable. The EA modulator supports only intensity modulation and cannot perform Coherent. 200G EML 3 dB bandwidth is approximately 60–67 GHz, supporting 112 GBaud PAM4. In the 200G era, EML covers all three IEEE P802.3dj standard reach tiers:
Lumentum’s 200G EML is shipping at volume: “currently shipping 200 Gbps per lane EMLs at scale… the foundational component enabling the latest 800G & 1.6T transceivers.”
InP PIC: Monolithically integrates laser, IQ modulator, SOA, and other components on an InP substrate. Capable of full-dimensional coherent modulation — the core long-haul Coherent solution. But for 200G datacenter IMDD, its functionality and integration depth exceed application requirements, making it economically unattractive for 200G datacenter IMDD.
SiPh (Silicon Photonics): CMOS-process waveguides and modulators on silicon wafers; requires an external InP laser source (silicon cannot emit light). Supports both IMDD and Coherent. The 12-inch wafer is the key advantage. Three main modulator types: MZI (Mach-Zehnder interferometer) — mature, high-bandwidth, commercial mainstream; MRR (micro-ring resonator) — ultra-compact but temperature-sensitive; EAM (electro-absorption modulator) — via SiGe heterogeneous integration, fast, compact footprint. Two MZIs form an IQ modulator for coherent modulation.
TFLN (Thin-Film Lithium Niobate): Exceptionally strong electro-optic effect (Pockels effect), modulation bandwidth >100 GHz, ultra-low drive voltage. Capable of both IMDD and Coherent IQ modulation — a universal candidate material for all 400G per lane modulation formats. Performance is outstanding but the supply chain remains early-stage (detailed in §4.2).
2.4 The 200G Landscape: Modulation × Material Platform
The table below maps four material platforms (rows) against two modulation schemes (columns), showing which combinations were commercially viable at 200G per lane.
200G status quo: InP EML and SiPh IMDD capture virtually the entire datacenter short-reach market. Coherent serves only long-haul DCI and telecom, with no presence in datacenter short-reach. TFLN’s industrialization was not mature enough to catch the 200G deployment window. Both Coherent and TFLN represent white space in the datacenter short-reach segment.
III. 400G per Lane: IMDD’s Reach Limitation and the Emergence of Coherent Lite
The transition from 200G to 400G per lane significantly raises the bar for device bandwidth, link budget, and modulation efficiency, challenging the established 200G technology landscape. Google published a seminal paper in the Journal of Lightwave Technology in February 2026 — “Next-Gen Datacenter and Campus Optics: Technology Choices beyond 200 Gbps per Lane IM-DD”— systematically analyzing the reach limitation of IMDD at 400G, the cost barriers of full Coherent, and the core simplification thesis behind Coherent Lite. New modulation schemes and new material platforms will enter the stage simultaneously — this is the core opportunity window discussed in this report. (Terminology note: “Coherent Lite” is a technology descriptor and should not be confused with Coherent Corp (COHR) or Lumentum (LITE) as publicly traded entities. That said, both COHR and LITE may emerge as core Coherent Lite players.)
3.1 IMDD Remains Viable at 400G — but with Reach Constraints
400G IMDD is technically feasible — bandwidth advances in D-EML and TFLN modulators make 400G PAM4 possible. The challenge is reach. IMDD employs square-law detection, where chromatic dispersion tolerance scales inversely with the square of the symbol rate (∝ 1/B²). 200G PAM4 operates at ~106 GBaud and covers 10 km; 400G PAM4 requires ~226 GBaud — baud rate doubles, dispersion tolerance falls to one-quarter. IMDD at 400G is only viable for distances under ~1 km with moderate link loss budgets — both reach and link budget are binding constraints. Beyond 1 km, and especially beyond 2 km, coherent technology must be considered. OCS deployment in datacenter and TPU interconnect networks imposes even higher link loss budgets, further narrowing the viable IMDD envelope.
3.2 Why Coherent? Why Not Full Coherent?
Coherent detection offers two fundamental advantages over IMDD. First, superior receiver sensitivity: the coherent receiver’s electrical signal is the product of signal and LO optical fields — the LO effectively amplifies the received signal. Under IMDD square-law detection, each 1 dB of optical loss translates to ~2 dB of electrical SNR degradation; under coherent detection, only ~1 dB. Second, greater resilience to optical interference: in OCS deployments, Google uses single-fiber bidirectional (BiDi) transmission to increase OCS radix and reduce per-port cost, but BiDi dramatically worsens in-band optical crosstalk — a single reflection produces co-propagating interference, whereas duplex systems require double reflections. Coherent systems tolerate optical interference 6 dB better than ideal IMDD.
However, deploying full telecom-grade Coherent is not viable. Even when transmitting single-polarization signals, a conventional dual-polarization (DP) system still requires a complete dual-polarization receiver to resolve polarization state and carrier phase — hardware complexity cannot be meaningfully reduced. Three barriers make full Coherent impractical for the datacenter: Cost — tunable ITLA, dual-polarization IQ modulator, full 4D coherent receiver push per-module cost far above IMDD. Power — polarization demultiplexing, frequency offset estimation, and wideband chromatic dispersion compensation consume substantial DSP power budgets that datacenter short-reach scenarios simply do not require. Capability excess — datacenters need only ≤20 km reach; traditional Coherent’s multi-thousand-km capability provides zero incremental value.
3.3 Coherent Lite: Stripping Complexity the Datacenter Does Not Need
The answer is not full telecom-grade Coherent, but a simplified coherent architecture that preserves spectral-efficiency advantages while removing functions unnecessary for short-reach links. This is the 2D PoFo (Polarization Folding) Coherent Lite architecture.
Core simplifications: the tunable ITLA is replaced by a fixed-wavelength DFB, dramatically reducing cost; dual-polarization is reduced to single-polarization, halving TX/RX device count and DSP complexity; short-reach scenarios eliminate the need for chromatic dispersion compensation, polarization demultiplexing, and frequency offset estimation. Most critically — 400G SP-16QAM requires only ~113 GBaud, the same baud rate as current 200G IMDD PAM4, requiring no breakthrough in device bandwidth.
Left: 2D Coherent Lite (single-polarization) architecture. TX: single laser → IQ modulator (MZM I + MZM Q + π/2 phase shift) → single-polarization QAM signal. RX: Pol-Folding → 1× 90° hybrid → 2× BPD → 2× TIA/ADC → DSP. Only 2 ADC channels required. Right: Conventional 4D Coherent (dual-polarization) architecture. TX: laser → 4× MZM dual-pol IQ modulator → PBC. RX: PBS → 2× 90° hybrid → 4× BPD → 4× TIA/ADC → DSP. Requires 4 ADC channels — approximately 2× the complexity and power of Coherent Lite.
2D PoFo Coherent Lite vs. 4D Coherent receiver sensitivity comparison. (a) 400G SP-16QAM @ 113 GBaud: 2D PoFo achieves sensitivity superior to 4D Rx at best polarization, with only ~1 dB penalty at worst polarization. (b) 400G SP-QPSK @ 226 GBaud: consistent trend. Conclusion: the Coherent Lite 2D simplified architecture does not sacrifice performance relative to the full 4D coherent receiver.
Target applications:
First, intra-datacenter Scale-Out. At 400G, IMDD covers only sub-1 km, moderate link-budget scenarios. In the FR (2 km) segment, Coherent Lite has a clear opening and may coexist with IMDD. As datacenter footprints expand, 10 km intra-DC distances are emerging — a primary incremental opportunity for Coherent Lite.
Second, campus-level Scale-Across (2–20 km). The 2–10 km segment was previously served by IMDD; at 400G, Coherent Lite may become the more viable option, creating incremental TAM. The 10–20 km segment was previously served by full coherent (ZR/ZR+); Coherent Lite can replace it at lower cost and power.
Third, OCS-paired interconnects. OCS serves both Scale-Up and Scale-Out, with deployment scaling rapidly. Each OCS node introduces additional insertion loss, raising link budget requirements for the paired transceivers and further expanding the addressable range for Coherent Lite.
Capital signal: Alphabet’s CapitalG led a $140 million funding round for Celero Communications in November 2025. Celero is focused exclusively on Coherent Lite DSP, with a founding team drawn from the Inphi/Marvell coherent DSP core group. Google is simultaneously the largest OCS industry driver, the author of the JLT paper proposing the Coherent Lite architecture, and — through CapitalG — backing Coherent Lite’s core silicon. Three signals, one direction.
3.4 400G per Lane Technology Roadmap Summary
The table below compares three optical transmission architectures—IMDD, Coherent Lite, and full Coherent—across key system characteristics, highlighting how modulation format, spectral efficiency, baud rate, reach, and system complexity evolve across different optical interconnect regimes.
IV. Material Platform Assessment: What Will 400G per Lane Be Built On?
400G per lane splits into IMDD (<1 km) and Coherent Lite (FR/campus/OCS). Four candidate platforms compete for these two pathways: InP EML, InP PIC, SiPh, and TFLN. None covers all scenarios.
4.1 Multi-Dimensional Assessment Matrix
The table below compares four photonic material platforms—InP EML, InP PIC, SiPh, and TFLN—across key technology and manufacturing dimensions, highlighting their relative maturity and suitability for 400G IMDD and emerging Coherent Lite architectures.
(★ = 1 point, ☆ = 0.5 point)
InP EML and SiPh lead on maturity and capacity certainty, but their performance sweet spots concentrate in different segments (EML skews IMDD; SiPh functions as an integration platform adaptable to various materials). The platforms with high performance fit across both 400G IMDD and Coherent Lite are InP PIC and TFLN — but their key uncertainty is manufacturing capacity: whichever scales production first and establishes stable volume manufacturing is more likely to capture early 400G/lane deployment share.
4.2 Pathway-by-Pathway Analysis
1) InP EML
