Data Centers

Data Center Compute & Network

AI Server Vendors: Quarterly Revenue

The three US-listed server OEMs. Supermicro shows the sharpest AI inflection because it is closest to a pure play; Dell and HPE carry large legacy businesses that mute the same signal. High revenue, thin integration margins.

Who Builds the AI Servers

The market structure behind the revenue lines: ODMs ship most hyperscale systems, branded OEMs fight for enterprises and neoclouds, and NVIDIA's rack-scale reference designs decide where the integration value sits.

Vendor	Position	Notes
Dell (DELL)	Largest branded AI server vendor	Roughly $10B of AI-optimized servers shipped in fiscal 2025 with a multi-billion backlog; anchor customers include xAI and CoreWeave. AI servers carry materially lower gross margin than its legacy mix.
Supermicro (SMCI)	Speed-to-market specialist	First to ship each NVIDIA platform generation at rack scale with direct liquid cooling; revenue roughly tripled across the 2023 to 2025 ramp. Governance turbulence (auditor resignation, late filings, both since resolved) remains the caveat.
HPE (HPE)	Enterprise and supercomputing	AI systems built on the Cray line plus GreenLake consumption models; the 2025 Juniper acquisition added the data center networking leg. More enterprise-weighted than Dell or Supermicro.
Foxconn, Quanta, Wiwynn (ODMs)	The hyperscale majority	Most hyperscaler AI systems ship direct from Taiwanese ODMs, bypassing branded OEMs entirely. Foxconn's AI server business has grown to rival its smartphone assembly revenue. Not US-listed; tracked as context.
NVIDIA rack-scale designs	The integration standard	NVL72-class rack systems define power, cooling, and interconnect for everyone above. Each generation shifts more integration value into NVIDIA's reference design and its blessed component list. NVIDIA is rostered in AI Chips.

Storage Vendors: Quarterly Revenue

The nearline hard drive duopoly rode a brutal downcycle into an AI-driven recovery: training data and inference archives fill exabytes. Western Digital's step-down reflects the 2025 Sandisk flash spin-off, not lost business.

Networking and Optics: Trailing Revenue, Indexed

Relative momentum across the cluster fabric cohort since the end of 2023. The AI interconnect specialists compound fastest off small bases; the optics ramp lifts Coherent and Fabrinet; Arista grows steadily from a base roughly 15 times Credo's.

The Cluster Fabric Stack

Five connectivity layers, five different battles. Ethernet took more than two-thirds of AI back-end switch sales in 2025, co-packaged optics loom over the pluggable transceiver market, and copper keeps winning short reach on cost and power.

Layer	Champions	Status
Scale-up (inside the rack)	NVIDIA NVLink; UALink consortium forming the open alternative	Proprietary NVLink domains (72+ GPUs) define the current generation; the open UALink challenge is years from shipping at scale
Scale-out back-end (training fabric)	InfiniBand (NVIDIA) vs Ethernet (Arista, Cisco, Broadcom silicon)	InfiniBand owned the early builds. Ethernet took more than two-thirds of AI back-end switch sales in 2025 and still about two-thirds in Q1 2026, even as InfiniBand rebounded on 800G Blackwell Ultra upgrades.
Front-end and data center interconnect	Arista, Cisco; Ciena for long-haul DCI	Conventional Ethernet switching, growing with cluster count; multi-site training makes campus-to-campus links a new bottleneck
Optical transceivers	Coherent, Innolight, Eoptolink; Fabrinet manufactures, Lumentum supplies lasers	800G ramping to 1.6T with each GPU generation; co-packaged optics (NVIDIA's photonics switches) start displacing pluggables at the switch face from 2026
Copper interconnect and retimers	Credo (active electrical cables), Astera Labs (PCIe retimers, Scorpio fabric)	Copper stays dominant for short reach inside racks because it is cheaper and lower-power; AECs displace passive cables as speeds rise

Ethernet vs InfiniBand in AI Back-End Networks

Dell'Oro's published snapshots, not an invented share chart. InfiniBand was nearly 80 percent of AI back-end switch sales in 2023; Ethernet was more than two-thirds and more than double InfiniBand for full-year 2025, and still about two-thirds in Q1 2026.

Period	Dell'Oro finding	Notes
2023	InfiniBand nearly 80% of AI back-end switch sales	Dell'Oro's 2026 predictions note: two years before 2025, InfiniBand accounted for nearly 80 percent of data-center switch sales in AI back-end networks.
2025 (full year and Q4)	Ethernet more than two-thirds; more than double InfiniBand	Ethernet switch sales in AI back-end networks more than tripled and accounted for more than two-thirds of AI-cluster switch sales. Celestica and NVIDIA together held about 50 percent of Ethernet AI back-end switch share; Arista ranked third.
2026-Q1	Ethernet still about two-thirds	Ethernet more than doubled and held about two-thirds of AI-cluster switch sales even as InfiniBand sales more than tripled, helped by 800G switches shipping with NVIDIA's Blackwell Ultra platform, including brownfield upgrades.
Speed mix	800G is the majority; 1.6T 2H 2026	800 Gbps switches accounted for the vast majority of Ethernet AI back-end shipments and revenue in late 2025 and Q1 2026. 1600 Gbps expected to begin volume shipments in the second half of 2026.

Seagate hard-drive capacity shipped

Nearline, the data-center mass-capacity product, rose from 497 exabytes in FY2025 to 695 in FY2026. Non-nearline was roughly flat. Western Digital does not disclose a comparable absolute split in its 10-K, so it is not plotted.

Arista Networks Quarterly Revenue

The dollar run-rate of the largest public Ethernet franchise in AI clusters, from $0.88B in 2022-Q1 to $2.71B in 2026-Q1. Distinct from the indexed six-name networking cohort, which shows relative momentum off different bases.

Prefer Sterling in Google

Add sterlingcharts.com as a preferred source so our charts can appear with a preferred badge in Google Top Stories, AI Mode, and AI Overviews.

Add as preferred source