The backbone of modern AI supercomputers. A $19.2B market accelerating at 24.1% CAGR — from InfiniBand XDR and 800GbE / Ultra Ethernet to NVLink 5, the newly released UALink 2.0 specification, and co-packaged optics entering commercial ramp as the fabric of 100,000-GPU clusters.
Interconnect infrastructure has evolved from a supporting role into a primary determinant of AI cluster performance. As GPU clusters scale toward hundreds of thousands of accelerators, the fabric connecting them — and its latency, bandwidth, and topology — dictates the speed of every training run. The market is experiencing a structural bifurcation between scale-up networks (NVLink, UALink) and scale-out fabrics (InfiniBand XDR, 800GbE with Ultra Ethernet). The UALink Consortium released its 2.0 specification in April 2026, and co-packaged optics (CPO) is transitioning from advanced prototypes to volume commercial deployment this year.
The most consequential market tension remains InfiniBand vs. Ethernet for AI scale-out. InfiniBand maintains a 3–5× latency advantage for collective operations critical in distributed training. However, 800GbE with the Ultra Ethernet Consortium (UEC) 1.0 stack — released June 2025 — is narrowing the gap, and Ethernet surpassed InfiniBand in AI back-end share during 2025. Hyperscalers (Meta, Google, Microsoft) increasingly choose open Ethernet stacks for vendor flexibility, and 2026 marks the first volume year for 1.6 Tbps switches, reinforcing a durable split.
| Technology | Generation | Bandwidth | Latency | Use Case | Status |
|---|---|---|---|---|---|
| NVLink 5 | Blackwell B200 / GB200 | 1.8 TB/s / GPU | ~1 µs | Scale-up: intra-pod | Production |
| InfiniBand XDR | Quantum-X800 (800G) | 800 Gb/s / port | ~100 ns | Scale-out: AI clusters | Production 2025+ |
| InfiniBand NDR | Quantum-2 (400G) | 400 Gb/s / port | ~100 ns | Scale-out: AI clusters | Deployed 2023+ |
| 800GbE / Ultra Ethernet | UEC 1.0 · IEEE 802.3df | 800 Gb/s / port | ~1–5 µs | Scale-out: hyperscale AI | Volume 2025-26 |
| UALink 200G (1.0 / 2.0) | Open Consortium | 200 GT/s / lane | ~500 ns | Scale-up: multi-vendor | Silicon H2 2026 |
| HPE Slingshot-11 | Rosetta / Cray EX | 200 Gb/s / port | ~200 ns | HPC: national labs | Deployed |
| CPO Photonics Switches | Quantum-X / Spectrum-X / Bailly | 102.4 Tb/s aggregate | ~100 ns | Scale-out: 800G+ hyperscale | Ramping 2026 |
The full Module 06 report delivers 60+ pages of primary research: vendor competitive intelligence, topology benchmarks, RFP selection frameworks, CPO transition playbooks, and buy-side procurement guidance for AI interconnect infrastructure.