How NVIDIA Corporation Makes Money

NVIDIA doesn't make chips. That's the first thing to understand. TSMC makes the chips — on 4nm and 3nm process nodes, in fabs that cost $20 billion or more to construct. NVIDIA designs the architecture, writes the software, builds the systems, and captures the margin. It's a fabless model, and it produces economics that shouldn't exist in semiconductors: $215.9 billion in FY2026 revenue, $120.1 billion in net income, a 55.6% net margin, and roughly $6 million in revenue per employee across a workforce of 36,000 people. Those are software-company margins on hardware-company scale. The revenue breakdown tells you where the gravity is. Data Center swallowed the company: approximately $190 billion in FY2026, or 88% of total revenue. This segment includes the GPUs everyone talks about — H100, H200, B100, B200, the GB200 Blackwell systems — but also high-speed networking (InfiniBand, NVLink, Spectrum-X Ethernet), DGX and HGX platforms, and NVIDIA AI Enterprise software. The buyers are a short list with enormous budgets: Microsoft Azure, Google Cloud, AWS, Meta, Oracle Cloud Infrastructure, plus sovereign AI programs from the UAE to Japan to France, plus thousands of startups burning venture capital on foundation-model training runs. Gaming still exists — roughly $15 billion, about 7% of revenue — through GeForce RTX cards, Nintendo Switch processors, and GeForce NOW cloud gaming. It's the business that kept the lights on for two decades while CUDA matured. Professional Visualization (around 2%) serves designers and engineers. Automotive (around 2%) sells DRIVE platforms for autonomous vehicles. Both are rounding errors relative to Data Center, but they keep NVIDIA's tentacles in markets that could matter in five years. Why do the margins hold at 73-75% gross when competitors exist? Because the product isn't really a chip. It's a deployment guarantee. When Microsoft commits $50 billion to AI infrastructure, the procurement team isn't comparison-shopping on FLOPS-per-dollar alone. They're asking: can we get this running in three months without rewriting our entire software stack? CUDA answers that question. Millions of developers, thousands of optimized libraries (cuDNN, TensorRT, NCCL, cuBLAS), every major framework pre-tuned — that's what sustains pricing power. Switching to AMD's ROCm means revalidating code that took years to write. Most organizations won't accept that risk while AI timelines feel existential. The geographic picture adds a wrinkle. The U.S. And China were the two largest markets, but export controls now block NVIDIA's most advanced chips from Chinese buyers. That's both a revenue headwind and a geopolitical accelerant — it pushes Chinese firms toward domestic alternatives faster. Meanwhile, the $5.7 trillion market cap (roughly 26x trailing revenue) prices in a belief that this isn't a cyclical semiconductor peak but a structural shift. If that belief cracks — if AI capex pauses, if custom silicon matures, if four hyperscalers decide they're overpaying — the downside is severe.

NVIDIA's growth story in 2026 comes down to one architectural bet: sell the entire AI factory, not just the GPU inside it. The company has been methodically climbing the stack — from discrete accelerator cards to rack-scale systems to software subscriptions — and the financial results show it working. The Blackwell ramp is the near-term engine. B100, B200, and GB200 systems offer substantial performance gains over Hopper for both training and inference. Q1 FY2027 guidance of $78 billion (up ~44% YoY) is almost entirely a Blackwell story. But the more interesting shift is networking. Data Center Networking revenue surged 263% year-over-year in Q4 FY2026 to $11 billion. That's the Mellanox acquisition ($6.9 billion in 2020) paying off — NVLink and InfiniBand are now essential plumbing for any AI cluster above a certain scale, and NVIDIA controls both ends of the wire. Inference is where the next margin pool lives. Training gets the headlines, but inference workloads are growing faster as models move into production. TensorRT, Triton, and NIM microservices optimize inference performance and — critically — create recurring software revenue that doesn't depend on hardware upgrade cycles. Then there's sovereign AI. Governments from the UAE to India to Singapore are building national AI infrastructure on NVIDIA platforms. This is a genuinely new customer category that didn't exist three years ago. It diversifies revenue away from four U.S. Hyperscalers, which matters because customer concentration is NVIDIA's most obvious vulnerability. The longer-duration bets — robotics (Isaac), autonomous vehicles (DRIVE), digital twins (Omniverse) — are real but small. Automotive was $2.3 billion in FY2026. These won't move the needle until physical AI applications reach the scale that language models hit in 2023. The honest assessment: NVIDIA has one massive bet (AI data center infrastructure keeps growing) and several options on the future. The massive bet is working spectacularly. The options are interesting but unproven at scale.