Skip to main content

InfiniBand vs Ethernet — Products

Updated 6/19/2026

Engine-synthesised product landscape for InfiniBand vs Ethernet, ranked by trend signal across hiring, capital, orders, and discussion axes.

Last refresh: 2026-06-18.


Ayar Labs Optical I/O Chiplet / Co-Packaged Optics for Ethernet/InfiniBand — co-packaged optical I/O chiplet

Trend: ↗ rising

Opportunity: Both InfiniBand and Ethernet roadmaps need optical I/O to break the SerDes copper-reach wall at 800G+. Ayar's protocol-agnostic chiplet (Ethernet/IB) bridges both camps.

Hot potential infrastructure layer — all three accelerator vendors are co-investors. Heavy ASIC + packaging hiring confirms silicon ramp; missing public product_launch/order means opportunity stage.

Companies committing: Ayar Labs, Intel.


Cornelis Networks Omni-Path Express CN5000 — HPC interconnect fabric

Trend: ↗ rising

Opportunity: Revived Omni-Path positioning itself as the only non-NVIDIA, non-Ethernet alternative for sovereign/lab HPC; AI Networking SA hire signals attempt to break out of DOE silo.

Niche third option with strong R&D + GTM hiring momentum but no hyperscaler logo. A 'potential opportunity' bet — credibility hinges on first AI cluster reference outside DOE.

Companies committing: Cornelis Networks.


Arista 7800R Ethernet AI Fabric (RoCEv2 for hyperscale GPU clusters) — Ethernet switch fabric

Trend: ↗ rising

Opportunity: Hyperscalers want an open, vendor-diverse alternative to InfiniBand; Meta's 24K H100 deployment validates RoCEv2-over-Ethernet at scale.

Hot — Arista's hyperscale Ethernet switch is the proven InfiniBand-alternative at Meta-scale and the public/investor narrative for Ethernet is accelerating.

Companies committing: Arista Networks, Meta Platforms.


Ultra Ethernet (UEC) Transport / Specification — open consortium spec/fabric standard

Trend: ↗ rising

Opportunity: 7 of the top demand queries in the sample are direct 'ultra ethernet' lookups — the market is actively trying to learn this standard before buying.

Hottest 'wind direction' signal in topic. Massive question-stage demand + flagship HPE and Meta product alignment. The standard, not a SKU — opportunity for early NIC/switch vendors who certify first.

Companies committing: HPE, Meta Platforms.


800G AI Ethernet RoCEv2 Fabric (51.2T switch for B300/Blackwell era) — 800G Ethernet switch fabric / transceivers

Trend: ↗ rising

Opportunity: Buyers want 51.2T 800G RoCEv2 switches for Blackwell B300 clusters; existing H100 RoCE deployments suffer ~130 idle hours/month, evidencing tuning + product pain.

Wind direction toward 800G RoCEv2 Spine-Leaf is unmistakable — demand queries + Meta's fabric-engineer hires confirm. Opportunity for both switch ASIC vendors and the optics ecosystem.

Companies committing: Enfabrica, Meta Platforms.


NVIDIA Quantum-2 InfiniBand NDR 400G Fabric — InfiniBand switch fabric

Trend: ↗ rising

Opportunity: Hyperscale GPT-class training requires lossless 400G fabric with mature RDMA; Azure Eagle is the showcase deployment.

Incumbent default for top-tier GPU training fabrics. Strong customer-win + active demand for matching 400G optics. The bar competitors must clear.

Companies committing: NVIDIA.


HPE Cray Slingshot 11 (400Gbps Ethernet-class interconnect) — high-speed Ethernet-class HPC fabric

Trend: → steady

Opportunity: DOE labs + UEC-aligned customers need a non-NVIDIA 400G fabric; Juniper acquisition gives HPE Ethernet switching depth to scale Slingshot into AI clusters.

Strongest credible Ethernet-class challenger to InfiniBand at the exascale tier. Hot product backed by acquisition capital and a dedicated GPU-fabric performance hiring.

Companies committing: HPE.


Meta Disaggregated Scheduled Fabric (DSF) over RoCE — hyperscaler-engineered Ethernet/RoCE fabric architecture

Trend: → steady

Opportunity: Largest existence proof that hyperscale GPU training can run on RoCE/Ethernet rather than InfiniBand. Multi-GW Prometheus/Hyperion buildouts will define the reference design for AI fabrics.

Hot reference architecture — Meta's DSF is shaping switch/NIC procurement industry-wide. The pain (RoCE tuning, idle GPUs) is exactly the opportunity surface for switch ASIC, NIC, and observability vendors.

Companies committing: Meta Platforms.


Intel Gaudi 3 with on-package 24x 200GbE RDMA Ethernet — AI accelerator with integrated Ethernet NIC

Trend: → steady

Opportunity: Eliminates separate NIC SKU and InfiniBand dependency in mid-market AI servers (IBM Cloud / Dell / Supermicro) — Ethernet-native path to scale-out.

Hot product: integrated 200GbE RDMA on accelerator is a structural Ethernet bet against NVLink+InfiniBand. Design wins exist; ecosystem driver investment ongoing.

Companies committing: Intel.


Enfabrica SuperNIC (high-radix NIC + fabric switch board) — SuperNIC / smart NIC board

Trend: → steady

Opportunity: PCIe Gen6 + 800G Ethernet SuperNICs are the disaggregation play against NVIDIA BlueField — open to challengers willing to support both InfiniBand-class workloads and RoCE.

Potential opportunity, not yet hot — board design hiring confirms silicon-to-system push but pre-revenue with structural NVIDIA channel conflict.

Companies committing: Enfabrica.


Get this data as JSONLast updated: Jun 19, 2026