The Register reported Nvidia networking revenue above $17.5 billion in Q2, compared with Cisco revenue of $17.3 billion for the same period — a milestone whose foundations Nvidia laid with its 2019 Mellanox acquisition. For UK buyers specifying AI clusters, that history now dictates procurement decisions as much as GPU benchmarks do.
View the data behind this chart
| Nvidia Networking | Cisco Total | |
|---|---|---|
| Q2 revenue | $bn17.5 | $bn17.3 |
The numbers that changed the procurement conversation
The Register reported Nvidia networking revenue above $17.5 billion in Q2, compared with Cisco revenue across all products in the comparable period of $17.3 billion. That is not a GPU maker dabbling in switches; it is a vendor whose interconnect sales now rival a dedicated networking incumbent's entire output.
The roots of that position go back to 2019, when Nvidia acquired Mellanox for $6.9 billion, a company known mainly for ConnectX NICs and InfiniBand switches used in HPC and limited hyperscale deployments, including a major Microsoft deployment. UK buyers who still mentally file Nvidia under 'GPU supplier' are working from an outdated org chart.
Why Nvidia bought a networking company, not a chip company
Dell'Oro Group analyst Sameh Boujelbene told The Register that customers choose Nvidia for its tightly integrated compute and networking architecture, optimised for AI workloads — and that integration, not raw silicon performance, is the actual product being sold.
Recursion CTO Ben Mabey's account bears this out: the drug-discovery firm moved from Ethernet-linked gaming GPUs to Nvidia's Quantum InfiniBand and ConnectX NICs specifically because buying compute and interconnect from one vendor simplified optimisation and troubleshooting into a single-system problem, rather than two separate ones. For a UK buyer evaluating InfiniBand versus Ethernet fabrics, that single-vendor accountability is the real value proposition, not just port speed.
Three fabrics, three buying scenarios
Nvidia's portfolio now spans a layered stack rather than a single product: NVLink and NVSwitch for in-rack scale-up, Quantum InfiniBand and Spectrum-X Ethernet for scale-out clusters, and Spectrum-XGS for linking sites separated by distance. For UK infrastructure teams, this means fabric choice is now a layered procurement decision rather than a single line item. Nvidia lists the SN6600 at 128 ports of 800G on its Spectrum-6 ASIC, specified at 102.4 Tb/s; its SN6800 5U configurations should be quoted with the exact port and ASIC configuration, as Nvidia's published results describe differing variants (including a 512-port 800G configuration and a separate four-ASIC, 409.6 Tb/s, 2,048-port 200G system).
Nvidia claims Spectrum-X can sustain 95% network efficiency in deployments exceeding 100,000 GPUs, delivering up to 1.6x the performance of off-the-shelf Ethernet in specified AI workloads. Buyers evaluating the Nvidia Spectrum-X networking platform against commodity Ethernet need to weigh that vendor claim against their own workload profile — the uplift is workload-specific, not universal.
NVSwitch: the revenue line most buyers miss
For the NVL72 rack systems described by The Register, Nvidia sold nine NVSwitch trays, enabling the 72 GPUs to operate as a tightly coupled system. Dell'Oro analyst Sameh Boujelbene estimates NVSwitch now accounts for roughly one-third of Nvidia's networking revenue, up from almost nothing less than two years ago.
The procurement implication is blunt: UK buyers specifying NVL72-class racks are buying networking hardware whether or not they request it separately. That should be factored into capex modelling and into any comparison with 800G AI networking alternatives from Broadcom-based or Cisco silicon vendors, since the scale-up fabric cost is effectively bundled into the compute order.
Adapters, optics and the hidden line items
Fabric choice now extends down to the adapter: Nvidia identifies BlueField-3 for Hopper-era deployments, ConnectX-8 for Blackwell systems at up to 800 Gb/s aggregate throughput, and ConnectX-9 for Vera Rubin systems at up to 1,600 Gb/s per GPU. Nvidia's Rubin reference architecture specifies BlueField-4 DPUs, ConnectX-9 SuperNICs, InfiniBand and Mission Control software as one design, reinforcing that Nvidia is selling a cluster blueprint rather than discrete parts.
Optics are part of the same budget line. Nvidia's LinkX product documentation covers both InfiniBand and Spectrum-X Ethernet transceivers, with rated link speeds up to 1.6 Tb/s depending on the exact product and configuration, meaning transceivers, cabling, power and cooling all need quoting alongside switch and GPU list prices. Teams speccing high-speed network adapters for AI pods should treat this as a total cost of fabric exercise, not a single SKU comparison.
Managing lock-in before signing
Deeper integration across GPU-to-GPU links, NICs, switch silicon, optics and fabric software simplifies validation and support, but it also concentrates dependency on one supplier. UK buyers should require documented RoCE or InfiniBand interoperability evidence, open telemetry and automation interfaces, clear component-replacement and software licensing terms, and a tested migration path before committing at scale — the same due diligence Servnet applies when assessing a Cisco SmartNet alternative for support continuity.
Nvidia positions Spectrum-XGS as a way to connect sites separated by hundreds of kilometres for a distributed AI environment. For UK operators, that raises latency, optical capacity and data-residency questions earlier in the design process than a single-site cluster would — worth cross-checking against the AI networking adoption dataset before scoping a multi-site build.
- 01The Register — Backbone of a beast: charting Nvidia's rise to networking titan · 31 October 2026
- 02Nvidia — Spectrum-X networking platform · 31 October 2026
- 03Nvidia — Ethernet switching products · 31 October 2026
- 04Nvidia — Rubin reference architecture press release · 31 October 2026
- 05Nvidia — Interconnect technologies overview · 31 October 2026
- 06nvidia.com
- 07perspectives.nvidia.com
- 08assets.theregister.com