NVIDIA designs the accelerators (GPUs) that train and run AI models, and sells them increasingly as complete systems: GPUs, CPUs, high-speed networking (NVLink, InfiniBand, Spectrum-X Ethernet) and the software to run them. Taiwan Semiconductor manufactures the chips; NVIDIA designs, integrates and sells.

The customers are the hyperscale clouds (Microsoft, Amazon, Google, Meta, Oracle), AI labs, and increasingly governments and enterprises. Gaming GPUs, once the core, are now a small share.

Revenue in FY2026 (year ending January 2026) was $215.9B, up 65.5% after more than doubling in each of the two prior years.

The silicon is excellent, but what keeps developers is CUDA: twenty years of libraries, compilers and developer habit. Almost every AI framework and model is optimised for NVIDIA first. A rival chip has to match not only performance but the software that lets engineers use it without rewriting their code.

That hold is being tested from the inside: its largest customers design their own chips (Google's TPU, Amazon's Trainium, Microsoft's Maia, Meta's MTIA) precisely to reduce dependence, and AMD offers a credible GPU with an open software stack. Inference, which is larger and more price-sensitive than training, is where alternatives have the best chance.

These are software-like margins on hardware: gross margin 71% and operating margin 60% in FY2026. Free cash flow was $96.7B and stock pay a modest $6.4B. The capital is light because manufacturing is outsourced.

The catch is concentration and cyclicality. A handful of customers account for a large share of revenue, and their spending depends on their own returns from AI. Semiconductor history is a series of shortages followed by gluts.

NVIDIA is moving from selling chips to selling whole AI data centres: GPUs, CPUs, NVLink and networking designed together as racks, with the software to run them. A new generation arrives about every year (Hopper, Blackwell, then Rubin), which keeps rivals chasing a moving target.

The shift under way is from training models to running them (inference): a larger market where the cost of each answer matters more than peak speed, and where custom chips compete hardest. How much of today's 71% gross margin survives that shift is the business's open question.

• The build-out slowing. AI spending by the clouds is lumpy, and chip history is shortages followed by gluts.
• Custom chips and AMD competing for inference, the larger and more price-sensitive workload.
• Export controls limiting sales to China and other markets.
• Concentration. A few customers buy much of the output, and one manufacturer (TSMC, in Taiwan) makes it.

• Gross margin as inference becomes the larger share of demand.
• A major customer moving most of its inference to its own silicon.
• The clouds' data-centre spending plans, year on year.