Optical components — lasers and EMLs

Why the concentration exists

The supply picture is made tighter by the architecture choices already locked into next-generation transceivers. A 1.6T DR4 OSFP module, for example, is built around four 400G differential EML lasers, and an 800G EML module requires eight 100G EML lasers at roughly ten dollars each, which puts the laser subsystem alone at eighty to over one hundred dollars per module. Each module therefore consumes multiple EML dice, multiplying the bottleneck from a per-laser problem into a per-transceiver problem across millions of ports. Because every additional 800G or 1.6T link carries several EMLs, a small shortfall in EML wafer output cascades into a large shortfall in finished optics.[11][12][21]

What the evidence shows

Who supplies it

The EML market is dominated by three firms, with Lumentum the clear leader. Lumentum is reported to hold between fifty and sixty percent of the global EML market, with one Bank of America estimate cited through LightCounting and Dell'Oro placing the figure at roughly fifty percent of high-speed EMLs. The top three EML suppliers — Lumentum, Broadcom, and Mitsubishi Electric — collectively account for about seventy-two percent of the market. Mitsubishi and Sumitomo Electric are significant Japanese players in EMLs, while Coherent (II-VI), Broadcom, EMCORE Corporation, Source Photonics, and HiSilicon are listed as additional participants across different segments of the laser chip market.[14][20][10][22][23][9]

What it depends on, and what depends on it

Inside the supply chain, EMLs sit between the indium phosphide wafer foundry and the optical transceiver or co-packaged optics module, and they are the upstream input that limits how many 800G and 1.6T ports can be built. A typical 1.6T DR4 OSFP module draws on four 400G differential EMLs from Lumentum, and an 800G EML module draws on eight 100G EMLs. Co-packaged optics implementations change the picture only partially: Nvidia's CPO switches still depend on an external continuous-wave laser source, and Lumentum's ultra-high-power CW lasers are estimated to represent nearly fifty percent of the external light source module's bill of materials.[11][21][15][16]

Downstream, optical components are consumed by transceiver and switch OEMs serving hyperscale AI and cloud data centers, where each additional rack of accelerators pulls in additional pluggable optics. The same EML and InP ecosystem also feeds co-packaged platforms: Nvidia unveiled its Quantum-X and Spectrum-X Photonics CPO platforms in March 2025, built with TSMC, Coherent, Lumentum, Corning, and Foxconn. Broadcom's Gen 3 TH6-Davisson CPO switch, which began shipping in October 2025, delivers 102.4 Tbps at 200G/lane using sixteen 6.4T Davisson DR optical engines on TSMC COUPE Gen 2. These switches and transceivers in turn connect GPU racks inside the AI clusters operated by the largest cloud customers.[19][17]

EMLs are also positioned against alternative light-source architectures that compete for the same role. Continuous-wave lasers without integrated modulation are the main alternative and are easier to manufacture because the chip is simpler, but capacity expansion is slow because equipment lead times are long. Directly modulated lasers represent another architectural option for lower-speed links, and silicon photonics entrants are pushing SiPh as an alternative to InP lasers for certain reaches. Indium phosphide EMLs nonetheless remain the preferred choice for 1.6T transceivers because of their superior performance at high temperatures, which is precisely the operating regime of densely packed AI switch racks.[2][8][4][6]

Where it sits in the stack

What would break it

The first fragility is that Lumentum is undershipping demand and is effectively sold out. Lumentum reports it is meeting roughly seventy percent of customer EML demand, and all EML capacity is locked under long-term agreements through calendar 2027. Bank of America research cited through LightCounting and Dell'Oro indicates Lumentum is sold out through 2028. The pricing that falls out of this allocation is unusually concentrated: Jefferies notes 200G EML pricing at nearly twice the average selling price of 100G EMLs at only a fifteen percent increase in cost of goods, so 200G EMLs contribute roughly ten percent of data center laser revenue from only about five percent of unit volume. Lumentum's revenue reached 665 million dollars in fiscal Q2 2026, up sixty-five percent year-over-year, with operating margins rising from 7.5 percent to 25.2 percent in twelve months.[14][10][13]

The second fragility is that the supply of EMLs is now shaped by a single customer's commitments more than by market prices. Nvidia committed four billion dollars in March 2026 to Lumentum and Coherent to secure priority access to EML production, which pushed delivery lead times for non-Nvidia buyers beyond 2027 and is described as risking a worldwide shortage. Multiple reports describe Nvidia as having pre-allocated massive capacity from Lumentum and Coherent, leaving rival optical module makers waiting for components beyond 2027. The structure means that any cut, delay, or renegotiation of Nvidia's commitment would change allocation for everyone else, while any expansion of Nvidia's orders would tighten the market for all other buyers.[1][3][5][7][21]

The third fragility is the narrow set of substitutes that can actually displace InP EMLs at these speeds. Silicon photonics is being pushed by several startups and incumbents as an alternative to Lumentum's InP lasers, and Intel has shipped over eight million photonic integrated circuits and thirty-two million on-chip lasers through its OCI chiplet, which integrates DWDM laser arrays and semiconductor optical amplifiers directly on chip without an external light source. CW lasers are the main near-term alternative but bring their own long equipment lead times, and Chinese suppliers such as HiSilicon are growing but face export restrictions that limit their access to cutting-edge InP fab technology. The result is a thin substitution layer in which the leading alternative architectures are themselves reliant on the same constrained InP ecosystem.[4][6][2][9][17]

What to watch

Several alternative architectures have dated milestones that could shift the supply picture. Broadcom's Gen 3 TH6-Davisson CPO switch began shipping in October 2025 and uses sixteen 6.4T Davisson DR optical engines on TSMC COUPE Gen 2, while Nvidia's Quantum-X and Spectrum-X Photonics CPO platforms were unveiled in March 2025. POET demonstrated its Blazar high-power multi-channel hybrid laser at OFC 2026 with claimed packaging cost reductions of up to seventy-five percent versus competing approaches. Nubis Communications, acquired by Ciena in 2025, demonstrated a 1.6 Tbps full-duplex optical engine at under 4 pJ/bit using two-dimensional surface coupling, and Intel's OCI chiplet has shipped over eight million photonic integrated circuits and thirty-two million on-chip lasers. Each of these dated demonstrations is a watch item for whether external light sources become partially redundant inside next-generation switches.[17][19][18][20]

Related nodes

Sources

1. techtimes.com · 2026-05-27T19:25:11
2. edgeoptic.com · 2026-03-02T14:56:56
3. eetasia.com · 2025-12-16T02:37:23
4. markets.financialcontent.com · 2026-03-26T14:16:56
5. trendforce.com · 2025-12-08T08:30:03
6. markets.chroniclejournal.com · 2026-03-26T14:16:56
7. sdxcentral.com · 2025-12-09T00:00:00
8. qsfptek.com
9. mlq.ai
10. terminal-x.ai
11. semiconductor-today.com · 18 March 2026
12. photoncap.net · 2026-03-16
13. benpouladian.com
14. chipstrat.com · Mar 02, 2026
15. chipstrat.com
16. jasonschips.ai
17. photoncap.net · October 2025
18. poet-technologies.com · 2026
19. exoswan.com · March 2025
20. semiconductor-today.com · 4 June 2026
21. photoncap.net
22. globalmarketstatistics.com · October 2022
23. verifiedmarketresearch.com