InfiniBand vs Spectrum-X: Picking a Scale-Out Fabric

"InfiniBand or Ethernet?" was a religious war about three years ago. In 2026 it's a procurement and operations decision. Both options run 800 Gb/s per port, both can train frontier models, and NVIDIA sells both. What differs is where the intelligence lives, who on your team can run it, and how it fits into the rest of your network.

TL;DR

• Ethernet is winning the share battle. Dell'Oro says Ethernet was about two-thirds of AI back-end switch sales in 1Q26, and its sales more than doubled. InfiniBand sales more than tripled in the same quarter on Blackwell Ultra 800G upgrades (Dell'Oro, June 2026).
• InfiniBand is still the lowest-friction way to run one giant synchronous training job, especially if you use SHARP in-network reductions.
• Spectrum-X is the better default for multi-tenant, cloud-style, or mixed environments where Ethernet tooling, EVPN/VXLAN, and a deep vendor pool matter.

Same speed, different design philosophy

InfiniBand (Quantum-X800 / XDR). The Q3400 switch has 144 ports at 800 Gb/s across 72 OSFP cages, 115.2 Tb/s aggregate, with 200G-per-lane SerDes (NVIDIA XDR switch docs). Losslessness comes built in through credit-based link-level flow control: a sender doesn't transmit unless the receiver has buffer space for it. A central Subnet Manager computes routes. SHARP lets the switches perform reductions in the network, which cuts the data volume in all-reduce.

Spectrum-X (Ethernet). Spectrum-4/6 switches paired with SuperNICs. The idea is to make RoCEv2 behave in AI workloads without depending on perfect PFC tuning. The switch sprays packets across paths (per-packet adaptive routing). The NIC puts out-of-order packets directly into GPU memory and runs congestion control based on in-band telemetry. NVIDIA's headline proof point is xAI's Colossus, where it claimed ~95% data throughput with Spectrum-X versus ~60% for standard Ethernet because of flow collisions (NVIDIA Newsroom).

So both get to "lossless-ish, high utilization." InfiniBand does it at the link layer by design. Spectrum-X does it with NIC + switch co-design on top of Ethernet.

The market has picked a lane (mostly)

The Dell'Oro data is worth reading closely:

• Ethernet's AI back-end vendor ranking in 1Q26 was Celestica first, NVIDIA a close second, Arista third, and Arista would have been near the top if you counted its deferred AI revenue. Cisco gained the most share (Dell'Oro). Celestica on top tells you hyperscalers are running white-box Ethernet at huge scale.
• Dell'Oro thinks much of the InfiniBand rebound came from brownfield upgrades of the installed base, not new greenfield wins. That fits what you hear in practice: teams that already run IB keep running IB.
• 1.6 Tb/s switches were sampling, with a ramp expected in 2H26.

Meta is the best public case study of Ethernet working at scale. Its SIGCOMM 2024 paper describes running RoCE for distributed training across tens of thousands of GPUs, including the routing and congestion-control work needed to make it hold up (Meta, SIGCOMM 2024). The Ultra Ethernet Consortium is pushing in the same direction with a standards-based transport built for AI/HPC.

How I'd actually decide

Question	Leans InfiniBand	Leans Spectrum-X
One tenant, one giant training job?	✅
Multi-tenant cloud, per-customer isolation?		✅ (EVPN/VXLAN, standard tooling)
Team skill set	HPC / IB veterans	Network engineers who live in Ethernet
SHARP-sensitive collectives (large all-reduce)?	✅	Partial
Want multiple switch vendors / white box?		✅
Already have IB installed base?	✅ (brownfield)
Front-end and back-end on one operational model?		✅

Some nuance the table leaves out:

1. SHARP's value depends on the workload. It helps most for large all-reduce in data-parallel-heavy training. MoE training is often bound by all-to-all, where SHARP helps less. Benchmark your own model before paying extra for it.
2. Ops cost beats capex. Switch list prices aren't public and get negotiated heavily, and optics are a big chunk of the fabric bill either way, since every port-end needs a transceiver. What really separates the two over five years is how many engineers can debug the fabric at 3 AM.
3. Supply matters. Lead times on 800G optics and switches have been tight. Being able to source from several Ethernet vendors is a real hedge, and Dell'Oro has called the market supply-constrained rather than demand-constrained.
4. Scale-up is a separate decision. Inside the NVL72 rack, NVLink handles GPU-to-GPU. This choice only covers the scale-out fabric between racks.

Bottom line

If you're building a single-purpose training cluster and your team already knows IB, Quantum-X800 is still the lowest-risk path, and the 1Q26 rebound says plenty of buyers agree. If you're building something that looks like a cloud, Spectrum-X (or another well-tuned RoCE fabric) is the default now, and the share data backs that up. Either way, nobody's getting cooked by picking the "wrong" one at 800G. They get cooked by under-provisioning optics, skipping congestion-control tuning, or hiring a team for the other fabric.