GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
← Back to feed

Arista Revenues Surge 38% as AI Data Center Ethernet Demand Outpaces GPU Supply

Arista Networks reported a 38% revenue surge, citing AI data center demand as the primary driver. The company has extended its list of major AI clients to Meta, Microsoft, Oracle Cloud Infrastructure, Anthropic, and Google Cloud — all five of the largest current AI infrastructure spenders. The guidance is strong enough that analysts are calling the current environment an AI Ethernet supercycle.

What Arista Builds

Arista sits at the network layer inside AI compute clusters. Its core products — Etherlink, the 7700/7800-series switches, and the EOS/CloudVision software suite — handle three distinct network roles inside a GPU data center:

  • Scale-up networks: High-bandwidth interconnects within individual AI accelerator racks (GPU-to-GPU)
  • Scale-out networks: Backend training networks spanning multiple racks or pods
  • Frontend networks: Storage and service access for model inputs, outputs, and checkpointing

As GPU deployments move from clusters of 8,000 to 100,000+ accelerators, the number of switch ports, optical interconnects, and bandwidth management layers scales with them. Every additional megawatt of GPU capacity requires additional network infrastructure that Arista supplies.

The Supercycle Framing

Arista management used the phrase “high-speed connectivity revolution” in its guidance. The investment thesis is structurally similar to memory: as AI clusters scale, network demand multiplies faster than compute demand because topology complexity grows non-linearly. A 16,000-GPU cluster doesn’t need 2x the switching of an 8,000-GPU cluster — it needs 4-6x, depending on the topology.

Revenue Concentration

The customer list is narrow. Meta, Microsoft, Oracle, Anthropic, and Google collectively represent the majority of hyperscale AI cluster spend. Arista’s 38% growth rate is therefore a proxy for the pace of AI data center expansion at those five organizations — not a broad-market signal. If any of the five delays build-outs, Arista’s growth rate slows.

The flip side: all five have publicly committed to accelerated AI capex through at least 2027-2028. Google raised its 2026 capex estimate to $205B, Microsoft’s AI capital budget now runs at $190B annually, and Meta’s Louisiana campus is targeting 5GW. Each of those expansion projects runs through Arista’s switch portfolio.

Alternative Technology Risk

The company’s core risk is InfiniBand displacement. NVIDIA’s proprietary NVLink/InfiniBand stack handles scale-up networking inside Blackwell and Vera Rubin DGX SuperPOD configurations, bypassing Ethernet for the highest-bandwidth GPU-to-GPU paths. Arista competes on the scale-out and frontend layers — the portions of the network that remain Ethernet regardless of NVLink adoption.

For the current generation of frontier AI clusters, that split is durable. Large-scale training runs span multiple NVLink domains, requiring Ethernet for cross-domain traffic. The supercycle thesis holds as long as cluster sizes keep growing beyond NVLink domain boundaries — which, at current trajectories, they will.