1.6T Optics: The AI Bottleneck Breaker Arrives in Q3 2026

The 200-word answer

Hyperscalers (Microsoft, Google, Meta, AWS, Oracle) are deploying 1.6T pluggable optics in Q3 2026, roughly 18 months after the first 800G ramps in early 2025. The driver is not Ethernet roadmap pace, it is GPU scale-out: an NVIDIA GB200 NVL72 rack needs 72 GPUs × 9 NICs each at 800G-class lanes, and Rubin Ultra (2027) pushes that to 1.6T per port. Three numbers matter. First, 1.6T shipments forecast to cross 4 million ports in 2026 and 14 million in 2027. Second, CPO trials at TSMC, Broadcom, and NVIDIA are real but still <3% of port volume in 2026. Third, Silicon Photonics 800G has reached ~$0.18/Gbps street price by May 2026.

1. What changed between 800G and 1.6T

800G optics shipped in volume from Q1 2025 using 8×100G PAM4 lanes. 1.6T uses 8×200G PAM4 lanes. Marvell's 1.6T DSP entered sampling in November 2025 at a reported 18W per module — close to the OSFP-XD thermal ceiling of 22W.

2. Who is deploying, and when

Microsoft Azure has publicly committed to 1.6T spine deployments in Q3 2026. Meta's MTIA v3 cluster, lit up in Prineville in October 2026, uses 1.6T as the leaf-to-spine fabric. AWS Trainium3 racks ship with 1.6T OSFP-XD cages but populate them initially with 800G modules.

3. The CPO question

Broadcom's Tomahawk 5-Bailly (51.2T CPO switch) entered customer trials in February 2026. For 2026, CPO captures <3% of hyperscale ports.

4. Silicon Photonics economics

SiPh-based 800G-DR4 hit $0.18/Gbps street price by May 2026, below the $0.22/Gbps for EML-based equivalents.

5. Supply chain pressure points

200G/lane EMLs — Lumentum and Coherent control ~70% of supply, lead times stretched to 38 weeks in Q1 2026.

6. Reverse-hype watch

Overhyped: fully integrated photonic AI processors. Underrated: optical test and measurement. Keysight's M8050A 1.6T BERT lists at ~$1.2M per seat.