Server and networking hardware (InfiniBand, high-speed interconnect)

Server and networking hardware, including InfiniBand and other high-speed interconnect, is the cluster fabric that links thousands of accelerators into a single training system: the servers, switches, network cards, and chip-to-chip links without which individual AI chips cannot function as frontier compute. It is the connective layer of the compute chokepoint, and export controls have recognised it as such by defining controlled chips partly by their interconnect performance.

Function

Training a frontier model requires accelerators to exchange data constantly; the interconnect, not the chip, sets the ceiling on effective cluster size. Within a server, proprietary links such as Nvidia's NVLink couple GPUs at bandwidths far beyond ordinary networking; between servers, InfiniBand, a market Mellanox came to dominate and which was folded into Nvidia by its acquisition of the firm completed in 2020, competes with high-performance Ethernet as the cluster fabric. Control of both the accelerator and the fabric gives Nvidia a full-stack position: chips, links, switches, and software.

Strategic significance and control history

The United States advanced-computing and semiconductor controls on China (2022-present) regulate specified chips and systems through technical thresholds that include computing performance and interconnect capability. That legal treatment does not make every network card, switch or Ethernet fabric a controlled item. On 13 January 2026, the Bureau of Industry and Security moved applications for Nvidia H200, AMD MI325X and comparable products meeting stated conditions to case-by-case review. It did not create a general authorisation for those products, associated servers or cluster fabrics.

The fabric layer also shapes enforcement because accelerators may move in servers assembled by system integrators. Datacentre construction, power draw and cooling can expose cluster-scale activity, but those observations do not establish prohibited end use or state intent without further evidence. Domestic interconnect and switch development is a substitution response; its performance and scale remain product-specific questions rather than a conclusion that controls have succeeded or failed.

See also

AI accelerators and GPUs (Nvidia H100, A100, and the export-tuned H20) · Compute clusters and AI datacentres (sovereign AI compute) · United States advanced-computing and semiconductor controls on China (2022-present) · High-bandwidth memory (HBM) · Economic statecraft

Sources

• BIS January 2026 semiconductor licensing policy
• 2022 advanced-computing rule
• Nvidia InfiniBand product architecture
• Nvidia InfiniBand cluster procedure