In 2025 the two Chinese module assemblers posted the best returns in this chain: InnoLight +468% and Eoptolink +369%. The two split roughly 60% of NVIDIA's 800G volume, and the GB200 transition raised the optical content per server, so the move looked natural. In 2026 the ranking flipped. The assemblers slowed to +69-75% while the layer above them pulled ahead: Marvell in DSPs at +187%, Lumentum in lasers at +113%, Ciena at +98%.
The bottleneck moved
The reason sits in supply structure. Moving from 800G to 1.6T makes EML lasers and 3nm DSPs the cost axis of a module. EMLs are an oligopoly of five (Lumentum, Coherent, Mitsubishi, Broadcom, Sumitomo) and DSPs a duopoly (Broadcom, Marvell). Lumentum's backlog is effectively sold out through 2028, and NVIDIA invested $2 billion directly to lock up capacity. Assembly can scale; lasers cannot, not quickly. The bottleneck climbed from assembly to components.
If 2025's assembler surge repriced volume, 2026's upstream rally reprices scarcity. Excess returns migrating one layer up the same chain matches this site's repeated observation: where the bottleneck sits decides that year's winner.
What has not moved yet
Fabrinet, the assembly outsourcer, is the chain's laggard at +6% in 2026. One reading says its +107% in 2025 already priced the cycle; another says 1.6T volume has not yet flowed down to outsourced assembly. By chain logic, when EML supply loosens, assembly volume rises again and Fabrinet catches the late rally. The risk to this thesis is CPO: if co-packaged optics arrive faster than expected, module assembly itself shrinks, and that is where this argument breaks.
Demand timing still belongs to NVIDIA, yet NVIDIA itself is near the bottom of this chain at +8.7% in 2026. This is a year when the scarcity of what goes into the chip is priced higher than the company selling the chip.