NVIDIA–TSMC CPO Push — Optics Inside the Package for AI Factories
NVIDIA and TSMC advancing co-packaged optics (CPO) via TSMC COUPE for next-gen AI accelerators: silicon photonics at package level, targets like 200Gbps+/λ and ~4 Tbps/mm bandwidth density, lower energy vs copper/pluggables. 2026–2027 ramp ambition; NVIDIA optics supply bets (Coherent, Lumentum, Marvell) — roadmap, not shipping SKU day.
Copper is the quiet ceiling on AI cluster scale. The proprietary NVIDIA–TSMC frame: put the optics in the package (CPO on COUPE) so hyperscale training/inference stops paying pluggable tax in power and bandwidth.
NVIDIA and TSMC are integrating co-packaged optical interconnects into joint AI chip designs using TSMC’s COUPE platform. Optical I/O at package level via silicon photonics and advanced packaging — coverage cites 200Gbps+ per wavelength and bandwidth-density targets around 4 Tbps/mm, with energy-efficiency gains vs traditional copper or pluggable optics. Part of broader co-design for next-gen accelerators; production ramps discussed for 2026–2027. Partnership includes standards work (OFC-era) and supply-chain investments; NVIDIA committing billions to photonics partners including Coherent, Lumentum, and Marvell (industry reporting).
Claims vs checks
Collaboration direction and COUPE/CPO framing are industry/NVIDIA–TSMC roadmap coverage (blogs, OFC, March 2026 reports). Exact density/power deltas and $4B+ optics spend figures are secondary reporting — treat as directional, not audited CAPEX line items.
Limits
- CPO yield and field reliability still industry risk factors.
- 2026–2027 targets can slip with packaging/supply constraints.
- Overlaps with same-day co-designed 2027 accelerator story — optics is the interconnect slice.
Sources
- NVIDIA blog and announcements on co-packaged optics and CPO (2025–2026).
- Industry coverage of NVIDIA–TSMC COUPE collaboration, OFC 2026 standards, and supply chain (March 2026).
- Reports on NVIDIA optics investments and TSMC COUPE roadmap (March 2026).