EPIC Photonics for AI Data Centers 2026 | Lumentum | External Laser Sources for Deployable CPO-Based AI Networks
1. Why Now: AI Clusters Can't Afford a Single Link Flap
EPIC, the European Photonics Industry Consortium, hosted an online technology meeting on Photonics for AI Data Centers, with six speakers spanning lasers, PCB-embedded waveguides, hollow-core fiber, photodiodes, photonic computing and optical switching. Lumentum's Senior Director of Product and Strategic Marketing, Christian Urricariet, opened by saying he would not talk about technology itself, but about deployability.
The reason is practical. Pluggable optics have been the industry's mainstay since the late 1990s: if one fails, you pull it, swap it, or upgrade it. CPO puts the optical engine on the same substrate as the ASIC, which gives the best energy efficiency but sharply reduces serviceability. NPO sits in between: the optical engine leaves the faceplate and is socketed or board-mounted next to the ASIC, shortening the electrical path, but these modules are not field-replaceable.
AI networks are also less forgiving than cloud networks: Ethernet tolerates link flaps, AI workloads do not. An idle GPU burns money, and without a module to pull, fault containment and recovery have to be redesigned.
Urricariet's view is that these factors will determine who adopts CPO and NPO, and at what scale. Beyond the few hyperscalers that understand these issues well, others will be more cautious.
2. Where Does the Laser Go: ILS, ELS or ELSFP
Optical engine placement (pluggable, NPO, CPO) is one axis. Laser placement is a separate, orthogonal axis. The closer the optical engine gets to the GPU, the harder thermal management becomes, and the laser is the most heat-sensitive part of the link.
Laser placement | Approach | Strengths | Costs |
ILS (internal laser source) | Laser co-packaged with the optical engine, closest to a classic transceiver | Short optical path, fewer optical interfaces | Tightest thermal coupling; a laser failure means replacing the host card or daughter card |
ELS (external laser source) | Laser placed elsewhere on the host board | Better thermal isolation; laser can be optimized and multi-sourced independently | Generally not faceplate-serviceable; optical power still has to travel |
ELSFP (pluggable external laser source) | Looks like a transceiver, emits light only, sits in the faceplate | Faceplate-serviceable, thermally isolated, sparable, monitorable, pay as you grow | Requires a PM-fiber shuffle or FAU inside the host |

▲ Laser placement options: optical power delivery complexity versus thermal isolation (Image source: Lumentum | EPIC Online Technology Meeting 2026 presentation)
Lumentum offered two market data points. The only CPO in public volume deployment so far, NVIDIA's Spectrum-X and Quantum-X switches, uses an external laser source. And Meta's optical backplane mockup at the OCP Global Summit had its faceplate packed with ELSFPs. We covered the thermal interface for rack-ready ELSFPs in Ayar Labs and Wiwynn: Rack-Ready High-Power ELSFPs, and Lumentum's message matches what we saw there: ELSFP has moved from MSA documents onto rack design drawings.
Notably, Urricariet did not frame this as the final answer. He said integrating the laser with the optical engine is coming and will be an important development for later generations. ELSFP is the most deployable option for the first few generations of CPO and NPO, not the end state.
3. The Cost of ELSFP: Power Budget Decides How Many Engines One Module Can Feed
ELSFP solves the reliability problem by moving the optical problem into the host board. With the laser far from the optical engine, every connector in between eats optical power:
A plain CW laser in an ELSFP leaves less power to split, so each module supports fewer optical engines.
Ultra-high-power (UHP) lasers deliver roughly 24–25 dBm, or about 350–400 mW ex-facet, enough to cover splitting and connector insertion loss and let a single ELSFP feed the maximum number of optical engines.

▲ ELSFP architecture: the laser stays at the faceplate and a PM-fiber network on the host board feeds the optical engines around the GPU (Image source: Lumentum | EPIC Online Technology Meeting 2026 presentation)
Lumentum therefore proposed a set of CPO and NPO deployment metrics beyond pJ/bit:
Optical power delivery margin: CW versus UHP lasers sets the margin of the distribution network.
Thermal envelope: how close the laser sits to the ASIC and how it is cooled.
System-level reliability: not just the laser or module, but the whole board and system.
Mean time to recovery (MTTR) and workload impact: how fast pluggable, non-pluggable and factory-replaceable NPO options can each recover.
The other direction is DWDM external light sources. The hyperscaler-led OCI MSA aims to raise fiber bandwidth density in scale-up with multiple wavelengths, and Lumentum has shown high-power DWDM external sources at OFC and ECOC. Urricariet recalled the first OCI iteration at about 53 Gb/s per channel. This points the same way as our earlier analysis, Laser Demand Is Counted in Wavelengths, Not Units: once DWDM becomes mainstream in scale-up, laser value shifts from units shipped to how many stable, high-power wavelengths you can deliver.
4. Slow and Wide: Lumentum Brings Face-Recognition VCSELs into the Rack
In Q&A, Urricariet was asked where the biggest opportunity lies, and his answer was more informative than the keynote itself.
The industry is starting to define a layer inside scale-up called scale-in: switchless optical connectivity between GPUs within a single tray. Should this layer copy scale-out and go fast and narrow at 200G per channel? Lumentum is betting on the other side, slow and wide: many channels at a lower per-channel rate.
Lumentum's confidence comes from consumer electronics. Over the past decade it has shipped roughly 2.5–3 billion VCSEL emitters for face recognition, and that manufacturing and reliability experience carries over to VCSEL arrays for short-reach interconnect. He also stressed that InP will continue to dominate scale-across, scale-out and most of scale-up. VCSELs target the shortest reach, effectively replacing copper even deeper inside the system than scale-up.
The cost of slow and wide surfaced in the same Q&A. An attendee from Focuslight cited estimates from a US event that a single rack could eventually need up to one million fiber cores, whether single-mode, multimode or multicore. The problem is no longer the laser or the photonic chip, but implementation. Urricariet agreed that multicore fiber (MCF) will be needed at that scale, and noted that Lumentum's demo partner Corning is working on high-density shuffle solutions.
This is consistent with our ECOC series: It Wasn't Physics That Killed Slow-and-Wide, It Was 60,000 Fibers. The number has gone from 60,000 fibers to a million cores; the scope differs (one system's fiber count versus a whole-rack core estimate), but the conclusion is the same: VCSELs are already efficient enough, and how far slow and wide goes depends on whether fiber and connectors can keep up.
5. The Other Five Speakers: Every Bottleneck Points to Getting Light In and Out
Company | What it does | Key numbers | What it asked the community for |
vario-optics (Switzerland) | Polymer multimode waveguides embedded in PCBs | 5–10 µm alignment tolerance; AT&S demo board with 4 embedded channels of about 100G | Optical engine and connector partners, PCB design challenges |
Albis Optoelectronics (Switzerland) | High-speed InP and GaAs photodiodes | 800G = 4×200G; 3.2T ultimately 8×400G | Form-factor requirements from TIA makers |
iPronics (Spain) | Silicon photonics MZI-mesh optical circuit switch (OCS) | 32×32 to 64×64 at almost the same die size; in-house amplifiers recover 8–10 dB loss | Large fiber array attach |
GLOphotonics (France) | Hollow-core fiber (HCF) | 31% lower latency; dispersion about 2.5 vs 17 ps/nm/km | Larger preforms and draw towers |
Lucore Technology (Denmark) | Frequency-domain photonic neural networks | Edge and defense first, data centers in about 3–4 years | Packaging and use-case partners |
vario-optics: the competitor is copper in the PCB, not fiber. It uses lithography to build multimode waveguides as a planar layer laminated into the PCB, so channel count scales like traces without assembling fibers one by one. The key is a 5–10 µm alignment tolerance, far looser than single-mode CPO, which allows vision-based or passive alignment through a standard electronics assembly flow. CTO Nikolaus Flöry was direct: the target is in-rack links still held by copper, such as optical PCIe.
Albis: photodiode innovation moves from packaging back into the device. At 50G, photodiodes were top-illuminated and wire-bonded. 100G introduced a backside integrated lens and flip-chip mounting. At 200G, wire-bond parasitics forced a move to flip-chip on an interposer or directly on the TIA. At 400G the interface stays the same, and new epitaxial designs cut parasitic capacitance and inductance. Two new paths follow: 2D detector arrays for fast and wide, and mass transfer printing, which transfers hundreds to thousands of waveguide UTC photodiodes onto a silicon photonics wafer in one step and is seen as a viable route to 400G photodiodes in volume.

▲ Photodiode evolution: from wire bonding and backside-lens flip-chip to mounting directly on the TIA (Image source: Albis Optoelectronics | EPIC Online Technology Meeting 2026 presentation)
iPronics: silicon photonics OCS closes the loss gap. It controls MZIs with thermo-optic phase shifters arranged in a mesh, and because silicon photonics is lossy, it developed its own O-band optical amplifiers to recover roughly 8–10 dB from input to output. Tested with 1.6T (200G per lane) transceivers from Coherent and Lumentum, the BER penalty was under one decade with Lumentum's module and essentially none with Coherent's. MEMS switches use less power, but carry 2–3 dB of loss they cannot compensate and a larger footprint. On switching speed, the phase shifter itself reconfigures in 100–300 µs, but at system level the server-to-switch communication dominates; using DMA inside the FPGA, iPronics has reached about 2 ms. Further reading: iPronics Uses Graph Theory to Benchmark Silicon Photonics OCS.
GLOphotonics and Lucore: long-term options. Hollow-core fiber loss records have reached 0.04–0.05 dB/km, below the roughly 0.14 dB/km limit of standard single-mode fiber, but it still costs 5–20 times as much, splicing to standard fiber loses about 0.5 dB, and drawing 100 km in one run is still a challenge; its battlefield is between data centers. Lucore's photonic neural networks are 3–4 years away from data centers. Neither changes the near-term scale-up supply chain picture.
6. What It Means for Taiwan's Supply Chain
Stacking all six talks together, value is moving toward three places:
1. High-power, multi-wavelength CW lasers. The ELSFP spec battle is about output power and wavelength count, not unit price. In LandMark Optoelectronics (3081) 2026 Q2 earnings, multi-wavelength silicon photonics lasers were already positioned as a moat, while Elaser's (3450) external light source is still on the eve of volume production. The timing of the OCI MSA's DWDM source specification is the key signal for this line.
2. Fiber distribution, arrays and connectors. ELSFP needs a PM-fiber shuffle inside the host, slow and wide needs fiber management at the million-core scale, and iPronics wants larger fiber arrays. This segment is more fragmented than lasers and more dependent on volume yield, which makes it the most natural entry point for Taiwan's precision assembly and passive component makers.
3. Optical PCBs and passive alignment. vario-optics' 5–10 µm tolerance proves that with enough tolerance, optics can follow an electronics assembly flow. Taiwan's PCB and substrate makers have volume scale, but in Europe AT&S has already partnered with a waveguide maker. Whoever brings the optical waveguide layer into mainstream PCB processes first holds the first ticket to slow and wide.
7. Conclusion
The most important signal from this meeting is not any single component spec, but that the industry is asking a different question: from how much power each bit saves to how fast a failure can be fixed and how much fiber a rack needs.
STT's view: for CPO generations before 2028, the laser stays on the front panel. ELSFP is the most pragmatic trade-off between deployability and efficiency today, and laser-engine integration will wait until enough reliability data has accumulated. Whether slow and wide replaces copper in the scale-in layer depends on fiber and coupling, not on the VCSEL.
Three things to watch next:
Whether ELSFP reaches a second volume platform beyond NVIDIA, for example Meta's optical backplane moving from mockup to production racks.
The OCI MSA's DWDM light source specification, as laser demand shifts from unit count to wavelength count.
Which of multicore fiber, PCB waveguides and mass transfer printing crosses the volume threshold first, since together they set the ceiling for slow and wide.
This article is for technology and industry trend analysis only and does not constitute investment advice.



Comments