# NVDA How the company works

> How NVIDIA (NVDA) works, in its own words from its annual report for fiscal 2026: how it makes money, what it sells, unit economics, competitive position, cost inputs, and the anatomy of the extremes: what has to break for the result to land far from the forecast, either way.

Canonical: https://pricedin.markets/nvda/company/

## How the company works

What it sells, how it makes money, and the anatomy of the extremes: what has to break for the result to land far from the forecast, either way. In NVIDIA's own words, from its annual report for fiscal 2026.

NVIDIA designs the chips, networking, systems and software that AI data centers are built from, has them made by foundries in Asia, and sells them to the cloud providers, AI model makers and enterprises building those data centers.

#### How it makes money

Fabless: NVIDIA designs, TSMC and Samsung make the wafers, contract manufacturers assemble. It sells "data center-scale" systems, chips plus NVLink and InfiniBand networking plus the CUDA software stack, so a customer buys the whole rack, not a part. Revenue was $215.9B in fiscal 2026, up 65%, with a gross margin of 71.1% and net income of $120.1B. Data Center is $193.7B of it; Compute & Networking, the segment that holds it, earned $130.1B of operating income.

#### What it sells

| Market, fiscal 2026 | Revenue | A year earlier |
| --- | --- | --- |
| Data Center compute Blackwell and Blackwell Ultra GPUs, Grace CPUs, rack-scale systems; Rubin ships in the second half of fiscal 2027 | $162.4B | $102.2B |
| Data Center networking NVLink, InfiniBand and Ethernet switches, adapters, DPUs; up 142% | $31.4B | $13.0B |
| Gaming GeForce RTX GPUs for PCs, GeForce NOW, console chips | $16.0B | $11.4B |
| Professional Visualization RTX PRO workstation GPUs | $3.2B | $1.9B |
| Automotive DRIVE computers and software for driver assistance and self-driving | $2.3B | $1.7B |
| OEM and other | $619M | $389M |

Revenue by end market as NVIDIA reports it. Two reportable segments: Compute & Networking $193.5B and Graphics $22.5B.

#### Unit economics

Design once, sell into several markets on "a unified underlying programmable architecture". Research is $18.5B a year and 31,000 of 42,000 employees work in it; more than half the engineers write software. The buyer is concentrated: one direct customer was 22% of revenue and another 14%, and most sales flow to a handful of cloud providers. Supply is bought ahead: non-cancellable orders, prepayments and deposits secure wafers, memory and CoWoS packaging, and a demand miss is paid for, as the $4.5B charge on H20 chips for China showed in the first quarter of fiscal 2026.

#### Competitive position

The platform the AI industry builds on: CUDA runs on every NVIDIA GPU, over 7.5 million developers use it, and the company says it powers over 78% of the TOP500 supercomputers. The named rivals are AMD, Intel and Huawei on chips; Amazon, Microsoft, Alphabet, Alibaba and Baidu designing their own accelerators in-house; Broadcom, Marvell, Arista and Cisco on networking. The company's own words on the risk: its "effective foreclosure from the China market helped our competitors build larger developer and customer ecosystems to challenge us worldwide".

#### Biggest cost inputs

Wafers from TSMC and Samsung, memory from SK Hynix, Micron and Samsung, CoWoS advanced packaging, and assembly at Hon Hai, Wistron and Fabrinet, a supply chain "mainly concentrated in Asia". Then research and compensation, $18.5B of research in fiscal 2026. Export controls are a cost of their own: licenses, inventory charges and a 25% tariff on H200 chips routed through US inspection for China.

#### Anatomy of a large **upside**

- **The data center is the new unit of computing, and NVIDIA sells the whole data center.** Every AI build-out buys GPUs, CPUs, NVLink, switches and software from one supplier. Data Center revenue grew 68% in fiscal 2026; networking alone grew 142%; Rubin arrives in the second half of fiscal 2027 with "up to a 10x reduction in cost per token" on Blackwell, keeping customers on the upgrade cycle.
- **AI moves from training to agents, inference and physical AI, and the platform follows.** Blackwell Ultra, Dynamo, Nemotron and Cosmos aim at agentic and physical AI, robotics and self-driving, so the installed base keeps buying as the workload changes, and software licenses (NVIDIA AI Enterprise) add a recurring line.
- **China reopens.** A product that satisfies both governments returns the company to "China's data center computing/compute market", from which it says it is now effectively foreclosed.

#### Anatomy of a large **downside**

- **The AI build-out digests.** Revenue rests on a few cloud providers' capital spending and on "the availability of sufficient data center capacity or energy for customers to procure"; a pause leaves prepaid supply, non-cancellable orders and inventory to be written down, as the $4.5B H20 charge showed.
- **Customers build their own chips.** Amazon, Microsoft, Alphabet, Alibaba and Baidu design accelerators in-house and AMD, Intel and Huawei sell alternatives; pricing pressure and a lower share of each build-out cut a 71.1% gross margin.
- **Export controls spread.** Controls now foreclose China's data center market and a replacement for the rescinded AI Diffusion rule "may impose new restrictions on our products or operations" worldwide, while they help rivals "build larger developer and customer ecosystems to challenge us worldwide".

Written from NVIDIA's annual report for fiscal 2026, in its words and numbers. The extremes are the long view, what has to break for the result to land far from the forecast; the quarter's own cases, each model's, are on the quarter forecast.