NVIDIA is no longer usefully described as a graphics chip company. It is the general contractor of the global AI build-out: it designs the accelerators, CPUs, DPUs, switches, optics and rack-scale systems that constitute an "AI factory," writes the CUDA software layer that makes them programmable, and increasingly underwrites the financing and power infrastructure that lets customers buy them. In FY2026 it generated $215.9 billion of revenue at a 71.1% gross margin and $120.1 billion of net income — profitability without precedent at this scale. Its competitive position rests on three reinforcing assets: a two-decade CUDA software estate with roughly five million developers; annual full-stack platform cadence (Hopper → Blackwell → Vera Rubin → Feynman); and privileged access to constrained TSMC advanced packaging and HBM supply. Its principal vulnerabilities are the mirror image of that strength: extreme customer concentration, a supply chain anchored in Taiwan, exclusion from China, and a demand base increasingly financed by NVIDIA itself.
At FY2026 year end NVIDIA employed approximately 42,000 people across the United States and 37 other countries (38 in total), of whom approximately 31,000 were in research and development. More than 80% of the workforce holds technical roles and more than half hold an advanced degree. FY2026 voluntary turnover was 3.7%, against a semiconductor industry norm in the mid-teens; over 40% of new hires came through employee referrals (Source: FY2026 Form 10-K, Item 1, Human Capital).
---
### The company's own characterisation
NVIDIA describes itself in its FY2026 Form 10-K and in every subsequent press release as "the world leader in AI and accelerated computing." The 10-K frames the business around the proposition that two simultaneous platform transitions are under way — from general-purpose CPU computing to accelerated computing, and from hand-written software to machine-learned software — and that NVIDIA supplies the full computing stack for both. Management's operative metaphor, repeated by Jensen Huang through FY2026 and FY2027 and reproduced in the FY2026 Sustainability Report, is that AI is a "five-layer cake": energy at the base, then chips, then infrastructure/systems, then models, then applications. NVIDIA's strategy is to participate in — and where possible set the standards for — every layer above energy, and to guarantee the layer below it.
### Independent characterisation
NVIDIA is a fabless systems company with a software moat and, since FY2026, a balance-sheet financing arm. Four things are true simultaneously:
1. **It sells systems, not chips.** The transition from Hopper HGX boards to Blackwell full-scale rack solutions (GB200/GB300 NVL72) and now Vera Rubin NVL72 moved the average unit of sale from a ~$30,000 accelerator to a ~$3 million rack. The 10-K explicitly attributes the FY2026 gross-margin decline to this business-model transition. The corollary is that NVIDIA now captures roughly 70% of the bill of materials of a new AI data centre rather than 30–40%.
2. **It monetises almost entirely through hardware.** Despite an enormous software estate (CUDA, CUDA-X, NIM, NeMo, Omniverse, AI Enterprise, Dynamo, Isaac, DRIVE, Cosmos), NVIDIA does not separately disclose software revenue. Management has periodically indicated a software and services run-rate in the low single-digit billions — under 2% of revenue. Software's function is demand creation and lock-in, not P&L. Deferred revenue of $6.4 billion at 26 July 2026 (of which $2.8 billion is customer advances) is small relative to revenue; NVIDIA is a transactional hardware business, not a subscription business.
3. **It is becoming a capital allocator in its own value chain.** By 26 July 2026 NVIDIA held $47.9 billion of non-marketable (privately held) equity securities, up from $3.4 billion eighteen months earlier, plus $42.8 billion of marketable equity securities and $5.0 billion of long-dated locked-up public equity in Other assets. It had $25 billion of remaining equity-investment commitments, $3.5 billion notional of land/power/shell guarantees for AI-cloud partners, and — announced August 2026 — up to $105 billion of residual-value guarantees supporting an OpenAI-affiliated tenant in Ohio. Gains from equity securities contributed $23.7 billion of pre-tax income in H1 FY2027, roughly 17% of pre-tax income.
4. **Its customers are few and enormous.** Five direct customers represented 22%, 14%, 13%, 11% and 10% of accounts receivable at 26 July 2026 — 70% of receivables in five names. Three customers had represented 25%, 18% and 13% at 25 January 2026. NVIDIA also disclosed in the FY2026 10-K that a single "AI research and deployment company" contributed a meaningful amount of revenue indirectly, by purchasing cloud services from NVIDIA's direct customers.
### Revenue model mix
### Value-chain position and customer types
NVIDIA is fabless. It designs silicon; TSMC fabricates it (N4P/N3P today, TSMC A16 1.6nm targeted for Feynman in 2028); TSMC CoWoS and, increasingly, alternative advanced packaging assemble it; SK hynix, Micron and Samsung supply HBM; Foxconn, Wistron, Quanta, Dell, HPE, Supermicro and Lenovo integrate systems. NVIDIA's direct customers are add-in-board makers, distributors, ODMs, OEMs, cloud service providers, AI model makers and system integrators. End markets served: hyperscale cloud, AI-native labs, neoclouds, sovereign/national AI programmes, enterprise IT, telecommunications (AI-RAN), scientific computing/HPC, automotive and robotics, healthcare and life sciences, financial services, media and entertainment, gaming, and professional design/visualisation.
