The interconnect that wires GPU clusters together — roughly six per GPU. 800G is running well below demand, and Nvidia spent $4B locking up the laser supply.
Tightness gaugeCritical
Tightness
74/ 100Critical
+5 this quarter — Nvidia's $4B EML laser lock-up pushed rivals' supply past 2027 as 800G demand ran 40–60% ahead of capacity.
AI training clusters need roughly six high-speed transceivers per GPU to sustain bandwidth across the fabric — far more than conventional servers. The binding chokepoint is the 200G-per-lane EML laser chip, made by only a handful of firms, so optics quietly gate how large a usable GPU cluster can actually be wired together.
Through 20291.6T supply seen 30–40% short of demand
What would loosen it
EML capacity expansions (Coherent, Lumentum, ELASER), silicon-photonics substitution, and the shift to co-packaged optics. But McKinsey sees 800G short through 2027 and 1.6T through 2029.
Latest developments
2026-05-27
McKinsey: 800G production 40–60% below demand through 2027; 1.6T shortfall 30–40% through 2029