AI training clusters and large-language-model inference workloads have pushed data centre networking into 800G as the new default. Arista 7060X6, Cisco Nexus 9332D-H2R, and Juniper QFX5240 are the three platforms most UK AI / hyperscale-tier enterprise buyers shortlist. All three are 32×800G capable. The differentiators are buffering architecture, congestion control, telemetry depth, and which switch OS your team operates.
Why 800G now
NVIDIA H100, DGX B200 and GB200 NVL72 deployments saturate 400G links during distributed training all-reduce phases. 800G doubles per-port capacity and halves the number of fabric links needed for equivalent throughput.
For AI training in particular, network utilisation can hit 80%+ sustained for hours during training runs. Sub-optimal buffering or congestion control adds days to training timelines — and at GPU rental costs, that's £100k+ per delay.
Arista 7060X6 — when to pick it
Arista 7060X6 uses Broadcom Tomahawk 5 (TH5) silicon — 51.2 Tbps, 32×800G or 64×400G. Industry-standard non-blocking architecture.
EOS — Extensible Operating System. Linux-based, fully programmable, mature CloudVision management. The configuration paradigm that hyperscalers (Meta, Microsoft Azure) standardised on.
Best-in-class telemetry. EOS streaming telemetry (gNMI / gRPC) feeds Arista CloudVision for real-time fabric visibility — essential for diagnosing training-job network hot-spots.
Cisco Nexus 9332D-H2R — when to pick it
Cisco's 800G entry runs on Silicon One G200 or Broadcom Tomahawk 5 (depending on SKU). Tighter integration with Cisco UCS compute, Cisco NDFC fabric automation, and existing Cisco TAC support relationships.
NX-OS — familiar to teams running existing Cisco Nexus estates. Configuration translation from current Nexus 9300 / 9500 deployments is straightforward.
See Cisco UK partner + Nexus catalogue.
Juniper QFX5240 — when to pick it
QFX5240 also uses Broadcom Tomahawk 5. Junos OS provides the same commit / rollback / structured config semantics as Juniper's routing + firewall platforms — operational consistency across MX / SRX / QFX.
Juniper Apstra (fabric intent-based automation) is best-in-class for greenfield AI fabric deployments. Define the intent ("32-node leaf-spine, 1:1 oversubscription, RoCEv2 ready"), Apstra generates the configs.
See Juniper UK partner.
The deciding question — which OS + which fabric model
If your network team operates Cisco Nexus today: Cisco 9332D-H2R is the lowest-friction choice. NX-OS familiarity + Cisco support relationship + UCS integration outweigh any feature delta.
If you're building greenfield + want hyperscaler-grade automation: Arista 7060X6 + CloudVision. The EOS configuration model + streaming telemetry are operationally different.
If you run Juniper MX / SRX elsewhere: QFX5240 + Apstra gives you Junos consistency + the best intent-based automation in the category.
For pure £/800G-port: pricing within ~10% across all three at competitive bids. Not the decision driver.