Ethernet displaced InfiniBand as the default AI back-end fabric between mid-2025 and Q1 2026, with the data center Ethernet switch market reaching $15.4 B in Q1 2026, up 39.8% year-over-year, and roughly $80 B in cumulative switch sales projected over the five-year window through 2030 [S3][S4].
InfiniBand commanded roughly 80% of AI training cluster share in 2023, but Ethernet overtook it as the lead AI back-end network by mid-2025, driven by Ultra Ethernet Consortium (UEC) 1.0 specification maturity, hyperscaler RoCE validation at Meta, Microsoft, and Oracle, and the OCP Ethernet for Scale-Up Networking (ESUN) production track that lowered per-port bill of materials [S1][S2][S3].
2026 market share and revenue by fabric
The data center Ethernet switch segment reached $15.4 B in Q1 2026, growing 39.8% year-over-year, and the bulk of new AI back-end switch ports are expected to ship at 800 Gbps in 2025, 1600 Gbps by 2027, and 3200 Gbps by 2030 [S3][S4]. Nvidia's Ethernet switch share jumped from under 4% two years ago to 21.5% in Q1 2026, on $2.1 B in switch revenue, up 193% year-over-year, while its combined InfiniBand plus Spectrum-X Ethernet networking business is now running at roughly $60 B annualized [S4]. Arista's 2026 revenue forecast is above $10 B (about +20% YoY) [S4]. InfiniBand's AI back-end share has fallen to the high teens to roughly a quarter of new cluster wins, concentrated in single-vendor Nvidia-stack builds and the most latency-sensitive training jobs [S2][S4].
Criterion-by-criterion fabric comparison
On latency, InfiniBand NDR delivers roughly 1 microsecond small-message latency with predictable tail behavior, while well-tuned RoCE Ethernet can approach that figure for the majority of AI training workloads when PFC and ECN are configured correctly; the older "5 to 10x slower" Ethernet narrative reflected misconfigured networks, not the technology's ceiling [S2]. On cost, tier 2 and tier 3 buyers face 1.5x to 2.5x higher per-port costs with InfiniBand once switches, NICs, and specialized staff are tallied, an increment that, on a 512-GPU cluster, is large enough to fund an additional 128 GPUs [S2]. On ecosystem, Ethernet now benefits from a multi-vendor UEC track, merchant silicon from Broadcom (Tomahawk 5, 51.2 Tbps AI leaf switches), and OCP ESUN reference designs, whereas InfiniBand remains a single-supplier fabric tied to the Nvidia full-stack bundle [S1][S4]. On roadmap risk, the UEC 1.0 specification released on June 11, 2025 standardized congestion signaling, transport, and telemetry for AI and HPC workloads, removing the multi-vendor interoperability fragility that hurt earlier RoCE deployments [S2].
Vendor ranking inside the Ethernet wave
When InfiniBand still wins the spec
InfiniBand remains the right pick for the smallest-latency, single-vendor Nvidia-only training builds, where the full-stack GPU-plus-NIC-plus-switch-plus-Spectrum-X bundle is already qualified and tail-latency variance is the dominant variable [S2][S4]. The fabric also holds the lead in narrow inference clusters under 256 GPUs that are all-Nvidia and price-insensitive, where the 1.5x to 2.5x per-port premium is absorbed by operational simplicity [S2]. Buyers building on merchant silicon, multi-vendor Ethernet, or any non-Nvidia GPU stack (AMD Instinct, custom ASICs) should treat InfiniBand as out of scope, because there is no equivalent single-vendor path [S1][S4].
Use-case selection rules for 2026
Pick Ethernet + RoCE (with UEC 1.0 silicon) for any cluster between 256 and 2,048 GPUs where TCO, multi-vendor sourcing, and a non-Nvidia GPU escape path matter, and where Meta, Microsoft, or Oracle-style RoCE-at-scale playbooks are referenceable [S2][S3]. Pick InfiniBand only for the narrowest latency-critical, all-Nvidia training builds, with the explicit assumption of a 1.5x to 2.5x per-port cost premium and lock-in to the Nvidia networking roadmap [S2]. For GPU-as-a-Service providers such as CoreWeave, Lambda Labs, and Vultr, Dell'Oro projects growth to outpace Tier 1 cloud providers over the next five years, which tilts the fabric decision toward Ethernet plus UEC silicon, since GPUaaS buyers mix GPU vendors and want commodity optics [S3]. On the optical side, 400G and 800G transceivers (DR4, FR4, SR8, plus breakout DAC and AOC) are functionally interchangeable between the two fabrics, so optical sourcing does not change the fabric call [S2].
Engineering risks and open watch items
The main Ethernet-AI deployment risk is configuration drift: PFC storms, ECN mis-tuning, and inconsistent congestion signaling still cause GPU idle time when RoCE is deployed without UEC 1.0 silicon, which is why UEC 1.0 compliance, not raw Ethernet support, should gate vendor selection [S2]. Co-packaged optics adoption in AI clusters is expected to start materializing in the next few years, led by Nvidia, and that shift will reshape per-port power and density budgets for both fabrics [S3]. Two trackable signals for the next 90 days: Nvidia's combined InfiniBand plus Spectrum-X revenue disclosure in its next quarterly, and the first UEC 1.0-compliant switch SKUs shipping in volume from Arista, Broadcom ODM partners, and Cisco, which will set the floor for AI-fabric Ethernet pricing through year-end [S4]. Buyers comparing fabric options for 2026 builds should weigh Ethernet-side TCO and supply-chain breadth first, and treat InfiniBand as a single-vendor performance tier, not a default. For context on how this fabric decision sits inside broader AI infrastructure capex, see the 2026 AI capex overbuild vs underbuild map, and for the analog instrumentation side that feeds the same control rooms, compare the adjustable vs fixed deadband pressure switch spec map.
Spec-level background on the components involved: pressure transmitter, flow meter, and industrial valve.