Artificial intelligence data centers trigger a wave of laser shortage

The latest research released by TrendForce indicates that in the context of the evolution of artificial intelligence data centers towards large-scale clusters, high-speed optical interconnect technology has become the core key to improving system performance and scalability. The report predicts that the global shipment of 800G and higher rate optical transceivers will reach 24 million units by 2025, and is expected to achieve approximately 2.6-fold growth by 2026, climbing to nearly 63 million units.

TrendForce reports that the surge in demand has caused a significant upstream bottleneck in laser light sources. Nvidia, motivated by strategic reasons, has secured capacity at key electro-absorption modulated laser (EML) suppliers, leading to extended lead times beyond 2027 and a worldwide shortage. Optical module manufacturers and CSPs are now actively searching for secondary suppliers and alternative designs, changing the competitive landscape within the laser industry.

Nvidia’s strategic monopoly on EMLs

Beyond VCSELs used in short-reach links, mid- to long-reach optical modules mainly depend on two laser types: EML and continuous wave (CW).
EMLs combine modulation functions on a single chip, which makes them highly complex and very challenging to produce. Only a few suppliers are available, such as Lumentum, Coherent (Finisar), Mitsubishi, Sumitomo, and Broadcom.

EMLs, known for their excellent reach and signal integrity, have become a critical bottleneck as hyperscale data centers extend their transmission distances. Nvidia’s silicon photonics and CPO development plans have advanced more slowly than anticipated, leading to ongoing dependence on pluggable modules for GPU cluster expansions. To ensure supply, NVIDIA pre-allocated a large portion of EML capacity, reducing availability for other regions.

CW lasers: The new favorite of CSPs—and the next capacity race

CW lasers offer a steady optical signal and are paired with silicon photonics chips produced at semiconductor foundries used as external modulators. Their simpler design stems from the absence of integrated modulation, which broadens supplier options. Consequently, CW lasers combined with silicon photonics has become the main alternative route for CSPs facing EML shortages.

However, CW production faces increasing constraints due to several factors: long equipment lead times restrict expansion, and strict reliability standards necessitate labor-intensive die-cutting and aging tests. Consequently, many vendors outsource these steps, which adds to downstream bottlenecks. This situation is causing the CW ecosystem to approach a capacity crunch, leading suppliers to hasten their expansion efforts.

High-speed PD demand surges; Taiwanese epitaxy vendors benefit

In addition to laser transmitters, optical modules need high-speed photodiodes (PDs) to receive signals. Leading vendors like Coherent, MACOM, Broadcom, and Lumentum are releasing 200G PDs to enable 200G-per-channel data transmission.

PDs are manufactured on indium phosphide (InP) epitaxial wafers, similar to EMLs and CW lasers. As laser manufacturers focus on expanding epitaxy capacity for laser production, many are outsourcing InP epitaxy to specialized foundries like IntelliEPI (iET) and VPEC, which presents a notable spillover opportunity for Taiwan’s epitaxy sector.

TrendForce forecasts that AI-driven demand is tightening not only memory supply but also the entire upstream laser ecosystem. Nvidia’s aggressive EML lock-in ensures its own supply security, but has inadvertently accelerated the shift toward CW-based and silicon-photonic solutions among non-NVIDIA players. Concurrently, the industry-wide race for capacity is restructuring supply-chain roles and fueling growth across compound-semiconductor epitaxy and processing vendors.

Source: TrendForce