Definition
InfiniBand is an open-standard, high-speed, low-latency interconnect technology used to network servers, storage, and accelerators in high-performance computing and large AI clusters. Unlike conventional Ethernet, it uses a switched-fabric architecture and offloads much of the communication work from the CPU, which keeps latency minimal even when thousands of nodes exchange data constantly. It is the dominant fabric connecting GPU servers in the biggest training systems.
How it differs from ordinary networking
InfiniBand's defining feature is native Remote Direct Memory Access (RDMA): one node can read or write another node's memory without involving either CPU in the data path, which cuts both overhead and latency. It carries clustering, storage, and management traffic over a single link. Speeds are described in tiers, with current NDR generations reaching 400 Gbps and 800 Gbps per port at 4x link width.
Why it appears in sovereignty discussions
InfiniBand mostly matters at data-center scale, but it is relevant to anyone reasoning about where large-model training actually happens. The fact that frontier training depends on specialized, expensive fabrics like InfiniBand is part of why running the very largest models remains centralized. For self-hosted inference on a single machine you will not need it, but understanding it clarifies the gap between home setups and industrial clusters.
InfiniBand typically carries traffic generated by NCCL and works alongside GPUDirect RDMA to move data straight between GPU memories across nodes.
In Simple Terms
InfiniBand is an open-standard, high-speed, low-latency interconnect technology used to network servers, storage, and accelerators in high-performance computing and large AI clusters. Unlike conventional Ethernet,…