(undated)
1.6T Optics: The AI Interconnect Bottleneck Breaker (Q3 2026 Field Guide)
Why Q3 2026 is the real inflection
- 1.6T module shipments: 1.2M units by Q4 2026 (LightCounting)
- 1.6T DR8 ASP Q3 2026: $1,800-$2,200
- Power per bit: 800G DR8 ~15 pJ/bit -> 1.6T DR8 ~9 pJ/bit
- OSFP-XD module power: 25-30W
- Broadcom Tomahawk 6 capacity: 102.4 Tb/s
- NVIDIA GB200 NVL72 rack: 120 kW, 72 GPUs, 14.4 Tb/s scale-out
- Coherent + Innolight hyperscaler 800G share 2025: ~55%
- EML laser supply concentration: >80%
- 1024-GPU H200 cluster optics cost at 800G: ~$3.5M
- Same cluster at 1.6T: ~$4.1M (17% capex premium)
- Power savings at 1.6T: ~50 kW per cluster annually
- Photonic AI startup funding through Q1 2026: $1.5B+
- CPO power savings vs pluggable: 30-40%
- LPO BOM savings: ~$300/module, ~5W
- 1.6T volume Q3 2026; CPO volume late 2027
The 30-Second Answer
Hyperscale cloud providers are moving from 800G to 1.6T pluggables in Q3 2026, with first-volume shipments crossing 1.2M units by end of Q4 2026 (LightCounting). NVIDIA's GB200 NVL72 rack pulls ~120 kW and needs 14.4 Tb/s of scale-out fabric per rack. Power per bit falls from ~15 pJ/bit (800G DR8) to ~9 pJ/bit (1.6T DR8). ASP $1,800-$2,200, OSFP-XD power 25-30W, CPO volume not before late 2027.
Why Q3 2026
The shift is gated by the 200G-per-lane PAM4 SerDes. Broadcom's Tomahawk 6 (102.4T, native 200G SerDes) samples Q1 2026; Marvell Nova 2 DSP and Broadcom DSPs qualify at TSMC N5 in early 2026. Without these three pieces, 1.6T is just two 800G ports glued together.
The Vendor Map
Coherent and Innolight hold ~55% of hyperscaler 800G shipments in 2025; both committed to 1.6T DR8 volume in Q3 2026. Eoptolink and Accelink follow. Mitsubishi, Sumitomo, and Lumentum control >80% of qualified EML laser supply — the real bottleneck.
Three Architectures
A — 1.6T pluggable OSFP-XD: 25-30W, $2,000/module, the Q3 2026 safe bet. B — LPO: saves ~5W and ~$300/module, 2027 pilot. C — CPO: 30-40% power cut, volume late 2027-2028.
Cost and TCO
A 1024-GPU H200 cluster at 800G costs ~$3.5M in optics; at 1.6T ~$4.1M (17% premium), but cuts module count in half and saves ~50 kW/yr per cluster.
Reverse-Hype Watch
Photonic AI processors raised $1.5B+ through Q1 2026; production at scale is 2028-2029. The underrated story is optical test and measurement at 200G/lane PAM4.
Next 30 Days
Lock 1.6T roadmaps with Coherent and Innolight; pre-qualify switches for Tomahawk 6 or Cisco G200; map EML supply at Mitsubishi/Sumitomo.