The real AI “tolls the rack” war: Ethernet-scale NVLink’s moat can’t stop fabric spend from compounding

Most AI “network” headlines focus on link speed. The investment-relevant question is different: in a 100k-GPU build, how much spend is forced by fabric scale-out rather than by GPU compute alone?

NVLink/NVSwitch primarily serves a scale-up domain: moving data quickly inside the rack so collective operations finish fast. Ethernet (especially upgraded for AI/HPC) is what connects those racks together at lower unit cost and with broader ecosystem reach. When hyperscalers redesign clusters, they’re implicitly choosing where the economic toll gets collected—by GPU interconnect vendors, by switch/NIC silicon and systems, or by hyperscalers’ own commodity fabric stacks.

Ultra Ethernet is positioned as a full-stack Ethernet evolution for AI and HPC workloads, built to close performance gaps at scale using networking behaviors designed for modern cluster collectives. The publicly available Ultra Ethernet specification documentation emphasizes the goal of transforming Ethernet for AI/HPC at scale and explicitly frames Ultra Ethernet as an interoperable standards effort.

Separately, NVIDIA’s NVLink materials emphasize a rack-scale, low-latency, all-to-all scale-up fabric, including large all-to-all systems (Vera Rubin NVL72). Taken together, the picture is consistent: NVLink protects the rack fast path, while Ethernet is the scale-out fabric that must be engineered to behave well under heavy collective traffic.

Here’s the causal mechanism investors can use.

1) Hyperscalers pick a cluster topology that maximizes training throughput per watt and minimizes time-to-train. That choice creates a traffic pattern: some collectives are served in-rack, others require cross-rack communication.

2) Ultra Ethernet’s goal is to make Ethernet behave well under AI/HPC traffic at scale, which reduces the performance penalty of cross-rack collectives.

3) When the cross-rack penalty falls, the architecture can rely more on Ethernet scale-out capacity and less on expanding NVLink-favored scale-up islands.

4) That shifts the revenue and margins toward Ethernet switch/NIC silicon ecosystems and toward systems integrators that standardize deployments.

NVLink’s all-to-all framing means it remains a powerful tool for in-rack collectives, but the economic question becomes whether hyperscalers can “afford” to push more work to the Ethernet fabric rather than paying the NVLink premium for larger scale-up domains.

Over 1–3 years, the “winner” isn’t the fabric with higher peak bandwidth on paper. It’s the fabric that enables repeatable oversubscription strategies while still meeting training timelines.

Ultra Ethernet’s standards framing is relevant because it targets interoperability and scale-out operability for AI/HPC traffic. NVLink’s value is relevant because it defines how much expensive high-bandwidth scale-up capability you need to achieve acceptable collective completion times.

Net: Ethernet’s advantage compounds if it keeps collectives efficient when pushed across racks; NVLink’s advantage compounds if hyperscalers repeatedly design bigger in-rack domains.