Back to News
Market Impact: 0.55

90% of Nvidia's Customers Now Buy This -- and It's Not GPUs

NVDAMETAMSFTORCLANETCSCONFLXNDAQ
Artificial IntelligenceTechnology & InnovationProduct LaunchesCorporate EarningsCompany FundamentalsAntitrust & CompetitionAnalyst Insights
90% of Nvidia's Customers Now Buy This -- and It's Not GPUs

Nvidia reported explosive networking growth, with $8.2 billion in networking revenue in Q3 fiscal 2026—up 162% year-over-year—and a networking attach rate near 90%, indicating most buyers of full AI systems also purchase Nvidia networking products. The company is pushing rack-scale Rubin systems (including the Vera Rubin NVL72) and Spectrum-6 Ethernet switches (up to 102.4 Tb/s), giving it an 11.6% share of the data-center Ethernet switch market behind Arista and Cisco. Market forecasts underpin the opportunity (MarketsandMarkets: AI networking $14.9B in 2025 to $46.8B by 2029 at a 33.8% CAGR; McKinsey cites large data-center capex needs through 2030), although the piece cautions on demand uncertainty and potential overbuild risk for AI infrastructure.

Analysis

Market structure is shifting from GPU-only capture to system+network capture: NVDA has a ~90% networking attach rate, reported $8.2bn networking revenue (up 162% YoY) and an 11.6% data‑center switch share — winners are Nvidia (NVDA), Ethernet optics/switch suppliers, and hyperscalers (META, MSFT, ORCL) buying turnkey racks; losers are incumbents exposed to lower‑end enterprise switching (CSCO) and any GPU‑only vendor that can’t sell rack systems. Competitive dynamics: Nvidia’s rack push (Rubin NVL72, Spectrum‑6 102.4 Tb/s) increases its pricing power at the high end and forces peers to choose between competing on silicon, software, or taking share in lower‑margin segments; expect pricing tension in 800GbE ports and accelerated product cadence. Risks: tail scenarios include hyperscaler overbuild leading to 30–60% capex retrenchment by 2027, antitrust/vertical‑integration probes within 12–36 months, and supply‑chain constraints for optics/silicon that could delay deployments by quarters. Hidden dependencies include hyperscaler purchasing cadence (multi‑quarter funnels), customers’ internal silicon roadmaps that can substitute GPUs but still rely on third‑party networking, and inventory cycles that can amplify swings; catalysts to watch in next 90–180 days are hyperscaler capex guidance and NVDA networking margins reported in quarterly results. Trade implications: bias long NVDA exposure to capture high‑margin networking and rack sales but size for volatility — establish 2–3% core longs and use options to cap cost; own ANET (Arista) selectively for durable data‑center switching demand but avoid legacy CSCO exposure which faces margin pressure. Options/relative trades: use 12–24 month NVDA LEAPS (buy calls or call spreads) for directional exposure, and consider a pair trade long NVDA / short CSCO (notional 1–2% each) to express structural share shift; target exits on +30% move or on proof‑point misses over 12 months. Contrarian angles: consensus underestimates cyclicality — networking attach could revert toward 50–60% if hyperscalers internalize networking or pause builds, making current growth partly front‑loaded; historical parallels include server/storage booms that reversed after inventory build (2018–2020). Unintended consequences: Nvidia’s verticalization may provoke faster in‑house switch SOCs from hyperscalers or accelerated Broadcom/Marvell product pushes; monitor customer procurement timing and vendor inventory days as leading indicators over the next 3–6 months.