Snapshot

InfiniBand is a high-speed interconnect protocol — a way for GPUs in an AI training cluster to talk to each other at sub-2-microsecond latency, far faster than standard Ethernet. The physical products are switch ASICs, network adapters (called HCAs — Host Channel Adapters), cables, and management software. NVIDIA acquired the dominant InfiniBand vendor Mellanox in 2020 for $6.9B and now controls an estimated 82% of InfiniBand port shipments. Marvell sells complementary interconnect PHY chips (the physical-layer silicon that converts digital signals to electrical/optical signals on the wire) and PAM4 DSPs (digital signal processors for 800G links) used alongside InfiniBand and high-speed Ethernet fabrics. The total InfiniBand market — switches, adapters, cables, and software — was $25.7B in 2025 and is projected to reach $35.1B in 2026.

NVIDIA's networking business generated over $31 billion in its fiscal year ending January 2026 — more than 4x Mellanox's entire $1.3B annual revenue at acquisition. This includes InfiniBand, NVLink (GPU-to-GPU scale-up fabric), and Spectrum-X Ethernet, but InfiniBand remains the core product for large AI training clusters. Ethernet is now growing faster: Dell'Oro Group found that Ethernet accounted for more than two-thirds of AI back-end network switch spending in 2025, up from under 20% when tracking began in late 2023.

The product & how money is made

An InfiniBand fabric consists of three physical product types that generate revenue:

• Switch ASICs and switch systems (45.3% of InfiniBand revenue, ~$11.7B in 2025 est.): the core routing chips and assembled switch boxes. NVIDIA's current-gen Quantum-2 (NDR, 400 Gb/s per port) and new Quantum-X800 (XDR, 800 Gb/s per port, 144 ports per switch) are the only InfiniBand switch ASICs in production. Each AI cluster needs hundreds to thousands of switches in a leaf-and-spine topology — a 100,000-GPU cluster uses roughly 10,000+ switches. Revenue per switch system ranges from $10K-50K+ depending on port count and speed.
• Host Channel Adapters (HCAs) — network cards that plug into each GPU server. NVIDIA's ConnectX-7 (400G NDR) and ConnectX-8 (800G XDR) are the adapters. Every GPU server needs at least one, and multi-GPU servers need several. The adapter-to-GPU ratio in a large cluster is roughly 1:1 to 2.4:1.
• Cables, transceivers, and software (31.6% of revenue est.): active optical cables, passive copper direct-attach cables, and NVIDIA's UFM (Unified Fabric Manager) management software. A reference architecture for ~1,000 GPUs requires ~2,400 optical modules.

Money flows in a simple chain: hyperscalers and AI labs (Meta, Microsoft, xAI, Oracle, government labs) buy GPU clusters, and each cluster requires a matching interconnect fabric. Interconnect spending runs roughly 15-20% of the total cluster cost est.. NVIDIA sells the complete stack — switches, adapters, cables, software — as an integrated solution bundled with GPU purchases. Marvell's role is different: it sells PAM4 DSP chips and electro-optics components that go into both InfiniBand and Ethernet transceivers and cables, plus PCIe/CXL retimers and data center switching ASICs. Marvell does not sell InfiniBand products directly — its exposure is to the broader high-speed interconnect layer regardless of protocol.

Source: Mordor Intelligence (InfiniBand Market, 2025); NVIDIA Quantum-X800 product page; NVIDIA DGX SuperPOD reference architecture documentation.

Demand

Contracted & deployed

• xAI Colossus: 100,000 H100 GPUs interconnected with Quantum-2 InfiniBand switches — currently the largest known single InfiniBand deployment. contracted
• Oracle: building a 131,000 GB200 GPU cluster; committed $40B to NVIDIA GB200 systems including InfiniBand fabric. contracted
• Stargate AI Data Center (SoftBank/OpenAI): 64,000 GB200 systems with 800 Gb/s InfiniBand. contracted
• Japan METI: JPY 72.5B (~$470M) for five AI supercomputers using InfiniBand fabric. contracted
• U.S. Navy Nautilus: 8.2 PF/s supercomputer on HDR InfiniBand, $35M contract. contracted
• InfiniBand connects 270+ of the world's top supercomputers (NVIDIA stated figure, as of 2026).

Forecast demand drivers

• GPU cluster scaling: as training clusters grow from 100K to 1M+ GPUs, interconnect port count and switch layers grow super-linearly. A fat-tree topology for N GPUs needs roughly O(N) switch ports, but the spine layer scales as O(N/radix), meaning larger clusters need disproportionately more spine switches. est.
• Speed generation transitions: the shift from NDR 400G to XDR 800G roughly doubles port bandwidth, driving a forklift upgrade cycle. XDR 800G is the fastest-growing speed tier at 40.7% CAGR through 2031 est.. Next: 1.6 Tb/s ports, expected in initial shipments H2 2026.
• Hyperscaler AI capex: the four largest US tech buyers (Meta, Alphabet, Amazon, Microsoft) roughly doubled combined capex from ~$230B (2024) to the high-$300Bs (2025), with 2026 guidance higher still. est.
• Inference compute growth: reasoning and agentic AI consume far more compute per query than simple generation. Each new reasoning step multiplies GPU-seconds, expanding how many interconnected GPUs are needed for serving, not just training.
• Total AI back-end networking: 650 Group projected the combined InfiniBand + Ethernet AI back-end networking market at nearly $20B in 2025. est.

Source: Mordor Intelligence; Dell'Oro Group (Q4 2025); 650 Group (2025 forecast); NVIDIA press releases; Oracle, xAI, SoftBank public announcements.

Supply

Capacity & production

• NVIDIA is the sole InfiniBand supplier. There is no second source for InfiniBand switch ASICs. ConnectX adapters come only from NVIDIA. Switch manufacturing is outsourced (TSMC for ASICs, contract manufacturers for system assembly), but the design and IP are exclusively NVIDIA.
• InfiniBand switches and cables have been in allocation alongside GPUs since 2023. Customers do not choose InfiniBand independently — it ships as part of the GPU platform.
• Quantum-X800 (XDR, 800G) is the current shipping generation. ConnectX-8 adapters provide 800G per port (up to 1.6 Tb/s per GPU with dual ports). Co-packaged optics (CPO) versions — where the optical transceivers are built directly onto the switch silicon — are in development and claim 3.5x power efficiency improvement.
• Marvell's supply: Marvell does not manufacture InfiniBand products. Its interconnect-relevant products — PAM4 DSPs, electro-optics engines (from the 2021 Inphi acquisition), PCIe/CXL retimers, and active electrical cables (AECs) — ship into both InfiniBand and Ethernet ecosystems. Marvell's data center switching revenue exceeded $300M in FY2026 (ending Jan 2026), with guidance to surpass $600M in FY2027. AEC/retimer revenue was ~$200M in FY2026 and expected to double.

The competitive threat: Ethernet

• Ethernet has overtaken InfiniBand in AI back-end network spending. Dell'Oro Group reported that Ethernet represented more than two-thirds (>66%) of all data center switch sales for AI clusters in 2025. When they began tracking in late 2023, InfiniBand held over 80%.
• The Ultra Ethernet Consortium (UEC), founded August 2023 with AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, and Microsoft, released UEC 1.0 in June 2025. It targets InfiniBand-equivalent latency (<2μs) with open multi-vendor sourcing at roughly one-third the hardware cost of InfiniBand est..
• NVIDIA itself sells Ethernet: Spectrum-X Ethernet surpassed a $10B annualized run rate in Q2 FY2026 (Aug 2025).
• Broadcom's Tomahawk 6 (102.4 Tb/s aggregate) shipped in 2025, approximately one year ahead of NVIDIA's Spectrum-X1600 at the same bandwidth.
• Key Ethernet adopters: Amazon, Microsoft, Meta, Oracle, and xAI are all deploying Ethernet for AI (Dell'Oro, 2025). Dell'Oro projects nearly $80B in Ethernet data center switch sales for AI over the next five years est..

Source: Dell'Oro Group (Feb 2026, Ethernet vs InfiniBand report); Ultra Ethernet Consortium; NVIDIA Q2 FY2026 CFO commentary; TrendForce (2025); SDXCentral citing Dell'Oro.

The gap

Pricing direction: InfiniBand hardware costs approximately 3x comparable Ethernet switches (TrendForce) est.. This price premium persisted because InfiniBand was the only viable option for large-scale AI training. UEC 1.0 Ethernet claims performance parity at one-third the cost. NVIDIA's response: Spectrum-X Ethernet ($10B+ run rate) — effectively cannibalizing InfiniBand to keep the networking dollar inside NVIDIA's ecosystem. Marvell benefits regardless of protocol winner because its PAM4 DSPs and electro-optics go into both InfiniBand and Ethernet transceivers.

Source: TrendForce (IB vs Ethernet cost comparison); Dell'Oro Group; 650 Group; Mordor Intelligence.

The players

Other relevant players: Broadcom (AVGO) — primary Ethernet switch ASIC competitor with Tomahawk 6 (102.4 Tb/s); Arista Networks (ANET) — #3 in AI Ethernet switch deployments; Cisco (CSCO) — accelerating Ethernet AI sales; Cornelis Networks — only non-NVIDIA InfiniBand competitor, shipping Omni-Path CN5000 with an 800 Gb/s roadmap targeting 2026, claims 35% lower switch-to-switch latency vs NDR.

Source: NVIDIA FY2026 earnings (nvidianews.nvidia.com); Marvell FY2026 10-K and Q4 earnings call; Calcalist (NVIDIA networking); Dell'Oro (market share); Mordor Intelligence (Cornelis).

The price of exposure

Dilution math for interconnect exposure: NVIDIA's networking revenue (~$31B) is roughly 14% of its $215.9B total. At the $5.4T market cap, pro-rata revenue allocation implies ~$756B for the networking business (rough approximation — the GPU compute business likely commands a higher multiple, so the networking portion may be valued at less). Marvell's interconnect-specific revenue is not separately reported; the entire $6.1B data center segment includes custom ASICs ($1.5B) alongside interconnect PHY/optics. Assuming interconnect is roughly $3-4B of Marvell's data center revenue, pro-rata allocation implies $125-170B of market cap for that exposure est..

FCF yield: NVIDIA ~3.0% (~$162B operating cash flow on $5.4T cap). Marvell ~0.65%.

Source: stockanalysis.com (NVDA, MRVL statistics pages); Calcalist (NVIDIA networking $31B); Marvell Q4 FY2026 earnings call.

What to deep-dive next

• NVIDIA 10-K FY2026 (filed with SEC): contains the actual compute vs networking breakdown within the data center segment. The $31B networking figure cited above comes from third-party reporting of CFO commentary, not the press release.
• The Ethernet migration rate: Dell'Oro's paid quarterly report ("Data Center Switch — AI Back-end Networks") contains the exact dollar and unit-share split between InfiniBand and Ethernet.
• Marvell's interconnect-specific revenue: Marvell does not break out PAM4/electro-optics vs custom silicon within data center. The FY2027 guidance implies interconnect grows >50% YoY — the earnings call transcript (Mar 5, 2026) has the most granular color.
• NVLink vs InfiniBand within NVIDIA's $31B: NVLink is scale-up (within a rack), InfiniBand is scale-out (between racks). They are bundled in "networking revenue." As NVLink's share grows with rack-scale Blackwell/GB200 systems, InfiniBand's relative share of NVIDIA's own networking revenue may be shrinking even as absolute InfiniBand dollars grow.
• Cornelis Networks Omni-Path CN5000: the only potential second-source InfiniBand-class product. Their 800 Gb/s roadmap (targeting 2026) and DOE/government contracts are relevant to the supply picture.

Sources & confidence