InfiniBand is a lossless, RDMA-native network technology long used to tie HPC clusters together. Ultra Ethernet is the industry effort, standardized through the Ultra Ethernet Consortium, that brings RDMA-class transport, lossless behavior and multi-tenant scale to standard Ethernet so it can carry AI training and inference. The question for anyone building AI infrastructure is which fabric to standardize on, and the answer increasingly shapes a decade of switch, NIC and optics purchasing.
Ethernet or InfiniBand for AI: Which Wins?
It depends on what you are building and what you already run. If you operate an existing InfiniBand estate with HPC workloads, it still fits, and rip-and-replace is rarely the first move. If you are building new AI factories, multi-tenant clouds, or clusters that must scale beyond a single vendor’s silicon, we project Ethernet wins, because it inherits a vast ecosystem of merchant silicon, optics, tools and operators. Our published view is that AI-first Ethernet is the default fabric of the AI era, and the market data below is why.
What Makes Ethernet ‘AI-First’ Now?
AI-first Ethernet eliminates the bottlenecks general-purpose networks create for AI: we examined the architecture shifts with NVIDIA SVP of Networking Gilad Shainer, including zero-jitter operation and 95% effective bandwidth in AI tuning (see Ethernet Switching Is the Heart of AI Factories). The transport behavior that made InfiniBand special, RDMA, is itself a key enabler of AI/ML deployments and now runs on Ethernet at scale (see RDMA Networks Are a Key Enabler to AI/ML Deployments).
What Do the Revenue Numbers Show?
Ethernet AI networking revenue surged over 150% Y/Y in 2024 with bandwidth up over 200% Y/Y (see our published results). Data center AI networking overall surged to nearly $20B in 2025 as of our June 2024 forecast, with the Ethernet share climbing (see the press release). As of October 2026, we project AI scale-out Ethernet revenue past $100B by 2030, and we published that Ethernet is set to surge in scale-out while ramping in scale-up (see In the AI Era, Ethernet Set to Surge in Scale Out and Ramp in Scale Up).
Where Does InfiniBand Still Fit?
InfiniBand remains deeply installed in HPC and national labs, and it continues to ship into AI clusters where a single-vendor, tightly-integrated fabric is acceptable. The migration dynamics matter more than the religion: buyers weigh ecosystem breadth, multi-sourcing, cloud operability and staffing against the maturity of InfiniBand’s lossless transport. The NIC layer tells the same story, with Ethernet NICs stepping up in step with switch speeds; our SmartNIC and DPU Vendor Guide maps the device landscape.
How Was This Calculated?
The figures in this Explainer come from 650 Group’s published quarterly research on AI networking and data center switching, which tracks fabric revenue, port shipments and vendor programs each quarter. Each figure links to its publication, and the full catalog is on our research reports page.
What to read next: The Widening Chasm: XPU Data Hunger Outpaces Network Infrastructure for why the fabric matters at all, and What Is Merchant Silicon? for the silicon economics behind Ethernet’s ecosystem advantage. Quarterly updates land in our AI infrastructure category.