InfiniBand
InfiniBand is a high-bandwidth, low-latency network interconnect technology used to link thousands of GPUs into a single usable cluster; it is the dominant fabric in large-scale AI training installations.
Last refreshed: 26 May 2026 · Appears in 1 active topic
Why is NVIDIA's networking revenue growing three times faster than its chip revenue?
Timeline for InfiniBand
served as the primary driver of NVIDIA networking revenue growth in Q1 FY2027
Data Centres: Boom and Backlash: NVIDIA networking up 199%, chips up 77%- What is InfiniBand and why does it matter for AI?
- InfiniBand is a high-speed network interconnect that links thousands of GPUs together for AI training. It delivers 200 Gb/s bandwidth and sub-microsecond latency, enabling GPU clusters to act as one coherent computer. Without it, large AI training runs cannot scale efficiently.Source: Lowdown data-centres update 4
- Why is NVIDIA's networking revenue growing faster than its chip revenue?
- Each Nvidia GPU that ships requires proportionally more InfiniBand fabric to connect it to a cluster. As GPU deployments scale, networking demand compounds faster than compute demand — Nvidia's Q1 FY2027 results showed networking up 199% versus chips up 77%, growing 2.6 times faster.Source: Lowdown data-centres update 4
- Who owns InfiniBand?
- InfiniBand was originally developed by an industry consortium in 1999. Nvidia took ownership of the technology through its $7 billion acquisition of Mellanox Technologies in 2020. Nvidia now develops and sells InfiniBand under the Mellanox brand.Source: Lowdown data-centres update 4
- What is the difference between InfiniBand and Ethernet for AI clusters?
- InfiniBand provides higher bandwidth (200 Gb/s per port in HDR) and lower latency (<1 microsecond) than standard Ethernet, making it better suited for the tight synchronisation required in large AI training runs. Standard Ethernet has higher latency and was designed for general networking rather than tightly-coupled parallel compute.Source: Lowdown data-centres update 4
Background
InfiniBand is a high-bandwidth, low-latency network interconnect technology used to link thousands of GPUs into a single usable cluster for AI training. Originally developed by a consortium in 1999 and now principally owned and developed by Nvidia following its $7 billion acquisition of Mellanox Technologies in 2020, InfiniBand is the dominant fabric in large-scale AI training installations. In Nvidia's Q1 FY2027 results (reported 20 May 2026), data-centre networking revenue — InfiniBand and Ethernet — reached $14.8 billion, up 199 per cent year-on-year, growing 2.6 times faster than compute revenue.
InfiniBand's technical advantages over conventional Ethernet are throughput and latency: HDR InfiniBand delivers 200 Gb/s per port and round-trip latencies below 1 microsecond, which matters when thousands of GPUs must synchronise gradient updates during a single training step. Without the fabric, GPU clusters cannot be scaled efficiently; the interconnect is the connective tissue that makes a warehouse of accelerators into a coherent computer. This explains why networking revenue is growing faster than chip revenue — each GPU that ships needs proportionally more fabric to be useful.
The 199 per cent networking growth figure signals a structural shift in where AI infrastructure spending is going. The GPU cluster bottleneck is moving from chip availability to network fabric capacity. Nvidia's control of both Mellanox InfiniBand and its own Spectrum Ethernet switching gives it end-to-end margin across the full cluster stack — a position no competitor currently matches at scale.