The deal let Nvidia sell clusters rather than components. Its datacenter customers, the reasoning went, did not want individual GPUs or servers but whole systems, and moving data between those systems required interconnect. Mellanox's InfiniBand fit that bill, and Nvidia has since pushed a co-design approach, tuning compute and networking together for AI workloads. Dell'Oro analyst Sameh Boujelbene told The Register that customers choose Nvidia for exactly this tightly integrated architecture, which simplifies deployments and speeds time to market.

The results show in the revenue curve. Networking income has climbed steadily since the acquisition, with two sharp jumps: one after ChatGPT kicked off the AI buildout in calendar 2023, and another across 2024 and 2025 as Blackwell-generation GPUs shipped in volume. Boujelbene points to the Blackwell ramp, the arrival of NVLink switches and accelerating adoption of Spectrum-X Ethernet for AI back-end networks as the inflection points. Big deployments followed, including the first half of Elon Musk's 200,000-chip Colossus cluster, built on H100 and H200 chips, which came online in 2024 and was at the time among the largest clusters deployed using Nvidia's Spectrum-X Ethernet switches. Meta has also made extensive use of Spectrum switches in its training clusters, including at its million-GPU Prometheus site in New Albany, Ohio.

Spectrum-X has done well partly because Nvidia reproduced InfiniBand's attractive qualities over Ethernet while gaining higher radix, meaning more ports per switch, and better interoperability. For very large clusters that matters: fewer switches are needed for a given port speed, which often makes Ethernet more economical beyond a certain GPU count. The Register cautions this is sometimes misconstrued as InfiniBand being unable to scale, which is not necessarily the case.