InfiniBand, Ultra Ethernet, and the Fight to Connect Accelerators
The scale-out fabric that connects accelerators across a cluster is the site of an intense and instructive commercial contest. It pits two networking philosophies against each other, and the outcome will shape the economics of AI infrastructure for years. The two contenders are InfiniBand and Ethernet, and their competition rewards a closer look because it exposes what an AI network actually needs.
InfiniBand was designed for high-performance computing, the world of scientific supercomputers, and it carries the assumptions of that world. It is engineered for low latency and for a property called losslessness: the network is built so that it very rarely discards data under congestion, using flow control to make senders pause rather than overflow. This matters enormously for AI, because the collective operations that synchronize accelerators are exquisitely sensitive to any link that stalls or drops. When one participant in a collective is delayed, all of them wait. InfiniBand's design minimizes those stalls, which is why it became the default fabric for large training clusters and why one vendor, having acquired the leading InfiniBand supplier, holds a powerful position in the layer.
Ethernet is the opposite in origin. It is the universal language of ordinary networking, produced at vast scale by many vendors, cheaper and more familiar than InfiniBand, and historically more tolerant of dropping data because the general internet was built to recover from loss rather than prevent it. That tolerance, acceptable for web traffic, is a liability for AI collectives. The industry's response has been to adapt Ethernet specifically for AI workloads, adding the flow control, congestion management, and low-latency behavior that training demands. This effort, pursued by a broad consortium of networking and chip companies under a banner commonly described as Ultra Ethernet, aims to give Ethernet the properties that made InfiniBand attractive while retaining Ethernet's openness, multi-vendor supply, and cost advantages.
Why does this contest matter beyond the engineering? Because it is a fight over who controls the fabric layer and how much it costs. A cluster's back-end fabric represents a large and growing share of its total cost, and the choice between a single-vendor InfiniBand ecosystem and a multi-vendor Ethernet ecosystem has consequences for price, for supply resilience, and for the freedom to mix components from different makers. Large operators, wary of depending on any single supplier for a layer this critical, have strong incentives to nurture the Ethernet alternative even where InfiniBand currently performs well. The pull toward an open, competitive fabric is as much strategic as technical.
The requirements that both technologies must satisfy tell you what an AI network really is. It must deliver very high bandwidth, because collectives move large volumes of data. It must deliver low and, crucially, consistent latency, because the slowest link paces the whole operation. It must avoid dropping data under load, because retransmission is fatal to tightly-coupled synchronization. And it must scale to thousands of endpoints without any of these properties degrading. These are demanding and partly conflicting requirements, and meeting them is why the fabric is engineered with such care and commands such attention.
There is a topological dimension as well. The switches that make up the fabric are arranged in structured patterns, commonly a multi-tier arrangement designed so that any accelerator can reach any other in a small, predictable number of hops. The design of this topology, how many switches, how they interconnect, how much bandwidth runs between tiers, determines both the cost of the fabric and its worst-case behavior. Two clusters using the same switching technology can perform differently because their topologies allocate bandwidth differently. The topology is where abstract networking choices become concrete performance.
The strategic reading is that the accelerator interconnect is the most contested layer in AI infrastructure precisely because it is the layer that most directly governs usable performance and represents a rising share of cost. Whether the future belongs to InfiniBand, to a matured Ultra Ethernet, or to a durable coexistence of both, the contest itself confirms the course's thesis: the network, not the chip, is where the decisive and under-tracked competition now lives.