AI Networking Index 2026: 800G Goes Mainstream

The 800G Inflection: What Actually Changed in 2026

Three years ago InfiniBand held roughly 80% of the data-centre switch market, the default fabric for GPU clusters chasing lowest latency. By January 2026, IEEE ComSoc's technology blog was reporting that Ethernet had overtaken InfiniBand in overall data-centre switch and server port counts — a reversal that vendor roadmaps had predicted but that only became visible in market data this year.

The speed axis moved in parallel. Cisco Live presentation material from mid-2025 forecast that the majority of switch ports in AI back-end networks would be running at 800Gbps in 2025, with that majority shifting to 1600Gbps by 2027. Read together, these are two separate but reinforcing trends: Ethernet winning the protocol argument, and 800G becoming the default port speed for new AI infrastructure rather than a premium tier.

• •InfiniBand share of the data-centre switch market in 2023: ~80% (IEEE ComSoc)
• •Ethernet overtook InfiniBand in switch/server port counts by January 2026 (IEEE ComSoc)
• •Cisco Live 2025 forecast: majority of AI back-end ports at 800Gbps in 2025, majority at 1600Gbps by 2027

Ethernet vs InfiniBand: The Share Reversal in Numbers

A January 2026 data-centre networking report circulated at the AI Infra Summit states plainly that Ethernet is now the dominant choice for new scale-out AI fabric builds, and that InfiniBand's market-share trajectory had already reversed by mid-2025. That's a distinct claim from the switch-port-count data above — this one is about which fabric buyers are specifying for new builds, not the installed base.

A separate Ethernet Alliance panel deck from December 2025 frames the split in bandwidth terms rather than port counts or new-build intent: Ethernet at 64% of AI bandwidth versus InfiniBand at 29%. IEEE ComSoc's own framing goes further, suggesting Ethernet's share of AI back-end switch ports is trending toward 70–80% as RoCE and Ultra Ethernet and UALink AI interconnects mature and scale. None of these three figures should be treated as interchangeable — they measure different things (new-build preference, AI bandwidth mix, and port-share trend) — but they all point the same direction.

Cost-per-Bit: What 800G Actually Costs Versus 400G in 2026

The clearest hard number in the 2026 market is switch-port economics. The Next Platform's Q1 2026 analysis put native 800Gbps switch ports at 35.8% of datacentre Ethernet sales, with native 200Gbps and 400Gbps ports still holding 34.1% — meaning the two speed tiers are running almost neck-and-neck in revenue share, not a clean 800G takeover.

On a cost-per-bit basis, the same analysis estimated native 800Gbps ports averaged $0.61 per Gb/s in Q1 2026, against $0.53 per Gb/s for the blended 200Gbps/400Gbps tier. That's a real premium for 800G at the switch-port level — but it's a switch-port figure only. The sourced material available for this study does not provide a reliable, segmented total-cost-of-ownership figure covering optics, cabling and cooling together, nor GBP-denominated pricing, so we're not going to manufacture a sterling number that isn't in the data. What the figures do tell UK buyers reliably: 800G isn't yet cheaper per bit than 400G, so it needs a workload reason — not just a roadmap reason — to justify the premium today.

The UK Reality Check: Where 800G Actually Fits

For most UK data centre estates, 400G remains the safer mainstream choice for brownfield upgrades and cost-sensitive fabrics — it's the tier where port economics are most favourable per the Q1 2026 cost-per-Gb/s data above, and it's the tier most existing infrastructure was designed around. 800G should be treated as the planning default specifically for new AI pods, dense spine layers, and hyperscale-style expansions, not a blanket refresh target for general enterprise or cloud estates.

The practical UK issue isn't only the port price — it's power, optics lead times and in-house skills. Because Ethernet's vendor ecosystem is broader than InfiniBand's, choosing Ethernet for new AI fabric generally buys UK buyers easier interoperability and less single-vendor dependence, which matters when comparing options like Arista 800G AI spine switches against alternatives — see how the major platforms stack up when you compare 800G AI networking vendors. Procurement teams should budget for optics and structured cabling changes early, since 800G's real cost is increasingly driven by the whole link rather than the switch port alone — though, again, the sourced data here doesn't extend to confident sterling guidance on that full-link premium.

Migration Playbook: Moving from 400G to 800G Without Breaking Production

A pragmatic 400G-to-800G migration treats most estates as hybrid for years, not months. That means new AI spine layers and dense GPU pods get built on 800G from day one, while existing leaf-spine and general-purpose fabrics stay on 400G until there's a workload-driven reason to change — the port economics above make that the financially rational default in 2026, not just a caution.

Breakout strategy matters more than switch selection: how an 800G uplink is split down to server-facing ports determines cabling, optics count and future flexibility, and it should be planned against your actual GPU-to-switch ratios rather than a generic template. Operationally, the challenges buyers report most are cable management density, optics testing and validation at higher lane counts, and monitoring tooling that can actually see per-lane performance rather than aggregate port stats. Offloading some of that packet and security processing to a DPU is increasingly part of the answer — see our explainer on DPU and SmartNIC technology — and any higher-speed fabric build should be reviewed against your wider network security posture before go-live, since higher-throughput links raise the stakes of a misconfigured segment.

Optics Supply and the Road to 1.6T

The optics market backs up the switch-side data. Global shipments of 800G-and-above optical transceivers are projected to rise from 24 million units in 2025 to nearly 63 million units in 2026, with 800G-and-above transceivers expected to exceed 60% of total transceiver shipments in 2026, according to 2026 industry commentary linked to TrendForce data. Nokia's own 2026-facing commentary describes 400G as mainstream and 800G as ramping through 2026, while pointing to a market of more than one million 800G units shipping annually by 2030 — a hyperscale-driven curve, not an enterprise one.

Cisco Live's 2025 forecast of a 1600Gbps majority of AI back-end switch ports by 2027 is squarely an AI-cluster and hyperscaler timeline. Spirent's 2026 market report separately estimates more than 65 million ports shipping across 2024–2026 in the high-speed segment it tracks, and more than 8.5 million ports across 2025–2026 alone — a different denominator to the AI-specific figures above, but consistent directional evidence that port volumes are accelerating broadly. For most UK enterprise buyers, the realistic read is that 1.6T decisions are a 2027-plus conversation; what you buy for 800G today — chassis, cabling plant, power provisioning — should be chosen with that next step in mind rather than treated as a dead end.

Sources

Every figure in this article traces to the sources below.

• •IEEE ComSoc Technology Blog — Ethernet overtaking InfiniBand in DC switch/server port counts, January 2026
• •ASP Events / AI Infra Summit 2026 report — Ethernet dominance in new scale-out AI fabric builds
• •Cisco Live 2025 — AI back-end switch-port speed forecasts for 2025 and 2027
• •The Next Platform — Q1 2026 800G switch revenue, sales share and cost-per-Gb/s data
• •Ethernet Alliance — December 2025 panel data on AI bandwidth share by fabric
• •Spirent — 2026 High-Speed Ethernet market impact report on port shipment volumes
• •LinkedIn / TrendForce-linked commentary — 800G-and-above transceiver shipment forecasts
• •Theia Global — Nokia commentary on 400G/800G demand through 2026 and 2030