Core Business

NVIDIA designs and sells AI computing infrastructure. Its core product is the GPU (graphics processing unit), a processor purpose-built for parallel computation that is now the foundational building block of AI training and inference workloads. NVIDIA's GPUs power the training of large AI models (like GPT and Gemini), the deployment of those models at scale (inference), and a growing array of reasoning and agentic AI applications. Beyond the GPU itself, NVIDIA sells full-stack AI infrastructure: complete rack-scale systems that combine GPUs, CPUs, networking interconnects, and software into an integrated AI computing platform.

NVIDIA's customers include:

• Hyperscalers and cloud service providers (CSPs): Amazon, Microsoft Azure, Google Cloud, Oracle, and others, which collectively represent about half of data center revenue. These customers build large AI factories to serve enterprise and consumer AI applications.
• AI model builders: OpenAI, Anthropic, xAI, Meta, Mistral, and others building and running frontier AI models.
• Enterprises and startups: Companies deploying AI for specific workflows, from fraud detection to drug discovery to autonomous vehicles.
• Sovereign governments: Nations building national AI infrastructure, with NVIDIA tracking over $20B in sovereign AI revenue for FY2026.
• Gamers and creators: Consumers who buy NVIDIA's GeForce GPUs for gaming PCs.

NVIDIA sells to large data center customers directly and through OEMs, ODMs, and system integrators who build and deploy NVIDIA-based infrastructure.

Segments

NVIDIA reports two segments:

Compute & Networking (~90%+ of revenue) includes:

• Data Center: The dominant revenue driver. NVIDIA sells GPU compute systems (currently the Blackwell architecture, transitioning to Rubin in late FY2027), CPUs (Grace), and data processing units (DPUs/BlueField). Systems are delivered as rack-scale units — the flagship GB200 NVLink 72 system connects 72 Blackwell GPUs and 36 Grace CPUs into a single rack-scale computer.
• Networking: NVIDIA sells InfiniBand (Quantum) and Ethernet (Spectrum-X) networking hardware for scale-out connectivity between compute nodes, and NVLink for scale-up connectivity within a rack. Networking generated ~$8.2B in Q3 FY2026, up over 160% YoY. NVLink, InfiniBand, and Spectrum-X Ethernet are all growing.
• Automotive: NVIDIA's DRIVE platform provides AI computing hardware and software for autonomous vehicles. Revenue was ~$570M+ per quarter in FY2026, growing rapidly and expected to reach ~$5B for FY2026 overall.
• Software: NVIDIA AI Enterprise, CUDA-X libraries, NIM inference microservices, and Dynamo inference framework are increasingly part of NVIDIA's full-stack offering.

Graphics includes:

• Gaming (GeForce): NVIDIA's GeForce GPUs for gaming PCs. Revenue was ~$4.3B in Q3 FY2026, up ~30% YoY, driven by the Blackwell-generation RTX 50 Series launch.
• Professional Visualization (Quadro/RTX Pro): GPUs for enterprise workstations used in design, simulation, and AI. Revenue was ~$760M in Q3 FY2026, up 56% YoY, driven partly by DGX Spark (an AI supercomputer in desktop form factor).

Business Model

NVIDIA makes money primarily by selling AI computing hardware at a premium driven by performance leadership. The key dynamics:

Performance-per-watt as the core value proposition: Data centers are constrained by power. Since AI factories directly monetize token generation (the output of AI inference), a data center's revenue is a direct function of tokens produced per watt of electricity consumed. NVIDIA argues that each generation of its architecture delivers substantially higher performance per watt than its predecessors — and that this advantage translates directly into higher revenue for customers operating power-limited facilities. This is the primary reason customers pay NVIDIA's prices rather than seeking alternatives.

Full-stack pricing: NVIDIA does not just sell chips. A single GB200 NVLink 72 rack contains 1.2 million components, weighs ~two tons, and integrates GPUs, CPUs, networking switches, and NVIDIA's software stack. The rack-scale system architecture means NVIDIA captures value across compute, networking, and software within a single customer deployment. NVIDIA estimates it captures roughly $30B+ per gigawatt of AI data center capacity in the current Blackwell generation.

Annual product cadence: NVIDIA ships a new GPU architecture roughly annually (Hopper → Blackwell → Blackwell Ultra → Rubin). Each generation delivers significantly higher performance per watt, which improves customer economics and drives upgrade cycles. Customers plan multi-year CapEx cycles around NVIDIA's roadmap. The pace of innovation is itself a competitive moat: the rapid cadence makes it very difficult for competitors to close the performance gap before the next generation arrives.

Software ecosystem as a retention mechanism: NVIDIA's CUDA programming platform, first introduced in 2006, has over 7.5 million developers and supports more than 6,000 applications. Software written for CUDA runs on all NVIDIA GPUs, and NVIDIA continually improves software performance for older hardware. Management notes that A100 GPUs shipped six years ago are still running at full utilization today due to software improvements — a TCO argument that competitors without a comparable ecosystem cannot match.

Gross margins: NVIDIA targets non-GAAP gross margins in the mid-70% range as Blackwell ramps. Margins are temporarily pressured during new architecture ramps due to manufacturing complexity and cost of expediting. NVIDIA's fabless model (relying on TSMC for wafer production) means its cost structure is largely variable.

Capital allocation: NVIDIA maintains large inventory and purchase commitments to secure supply chain capacity, using its balance sheet to guarantee offtake to suppliers. It also makes strategic equity investments in key AI model companies (OpenAI, Anthropic, xAI, Mistral) to deepen technical partnerships and expand the CUDA ecosystem.

Competition and Market

The AI accelerator market is currently a de facto oligopoly with NVIDIA holding a dominant position.

Key competitors:

• AMD (MI300X and successors) is the most credible GPU competitor but trails NVIDIA significantly in software ecosystem and benchmark performance, particularly for inference on complex models.
• Custom ASICs from hyperscalers: Google (TPU), Amazon (Trainium/Inferentia), Microsoft (Maia), and Meta are building internal chips optimized for specific workloads. Broadcom and Marvell supply ASIC design capabilities to these hyperscalers.
• Huawei (Ascend) is the primary alternative for Chinese customers given U.S. export controls, and is gaining scale domestically.
• Intel (Gaudi) has largely failed to gain traction in AI accelerators.

Why customers choose NVIDIA:

• Software ecosystem (CUDA): CUDA is the de facto standard programming model for AI. Every major AI framework (PyTorch, TensorFlow, JAX) is optimized for CUDA, and every new model architecture is typically developed and validated on NVIDIA GPUs first. Switching away from CUDA imposes significant retraining and porting costs.
• Breadth of workload coverage: NVIDIA GPUs handle the full AI pipeline — data processing, pretraining, post-training (reinforcement learning), and inference — on a single architecture. Custom ASICs are optimized for specific tasks and become inefficient or obsolete as model architectures evolve rapidly.
• Availability everywhere: NVIDIA is deployed across all major clouds, on-premises, at the edge, and in robotics systems. This ubiquity makes NVIDIA the default platform for developers and startups, reinforcing the ecosystem flywheel.
• Performance on inference for reasoning models: As AI has shifted toward long-thinking, reasoning-based inference (where a single query requires orders of magnitude more compute than one-shot responses), NVLink 72's rack-scale memory bandwidth advantage has become particularly important. NVIDIA claims 10-15x inference performance advantage over H200 for mixture-of-experts reasoning models.

Barriers to entry are high but not insurmountable at the chip level. The harder challenge for competitors is matching the CUDA software ecosystem, which represents 20+ years of developer investment. Jensen Huang's argument — that NVIDIA runs every AI model, is available in every cloud, and can handle every phase of AI workload — is the clearest articulation of the platform's structural advantage. That said, hyperscalers' custom ASIC programs are a genuine long-term competitive threat, particularly for inference workloads where model architectures are more stable.

China market: U.S. export controls have effectively closed NVIDIA's data center compute business in China, which NVIDIA estimates as a ~$50B annual market. Huawei's Ascend chips are the primary beneficiary. NVIDIA estimates the lost China opportunity is a material and ongoing competitive harm, as Chinese customers building on Huawei's platform contribute to an alternative ecosystem that could challenge NVIDIA globally.

Growth Strategy

NVIDIA articulates its growth opportunity around three simultaneous platform transitions:

1. CPU to GPU accelerated computing: Much of the world's existing cloud computing still runs on CPUs. As Moore's Law slows, NVIDIA argues that transitioning these CPU workloads (data analytics, simulation, classical ML) to GPU accelerated computing is both inevitable and ongoing — and these workloads happen to run on NVIDIA's existing infrastructure.

2. Classical ML to generative AI: Hyperscalers' core revenue-generating workloads — ad recommendation systems, search ranking, content moderation — are migrating from classical ML to generative AI. This transition is still in progress and requires substantially more compute per workload.

3. Generative AI to agentic and physical AI: Agentic AI (AI systems that reason, plan, and take multi-step actions) requires orders of magnitude more inference compute than one-shot generative AI. Physical AI (robotics, autonomous vehicles, industrial automation) requires training in simulated environments (Omniverse/Cosmos) and dedicated in-device computing (Jetson/DRIVE platforms). NVIDIA sees these as multi-trillion-dollar long-term demand drivers.

Management estimates total AI infrastructure investment will reach $3–4T by the end of the decade, up from ~$600B in annualized CapEx among just the top four hyperscalers today.

Sovereign AI is an emerging growth vector — countries building national AI infrastructure — with NVIDIA tracking over $20B in sovereign AI revenue in FY2026, more than double the prior year.

Product roadmap: NVIDIA is on an annual cadence — Blackwell (FY2025–2026), Blackwell Ultra (FY2026 H2), Rubin (FY2027). Each generation targets a roughly 10x reduction in cost per token relative to the prior generation, which continuously expands the economically viable market for AI inference.