Nvidia designs advanced computer chips called GPUs, has them manufactured by contract foundries like TSMC, bundles them with its own CUDA software, and sells the package mostly to companies building AI data centers — plus, in a much smaller way, PC gamers. It owns no factories of its own.
Nvidia designs the chips and TSMC builds them
Nvidia is a "fabless" chipmaker — it designs chips but pays other companies to build them. TSMC manufactures the silicon; SK Hynix and Samsung supply the high-bandwidth memory that sits next to it. Nvidia's own job is chip design plus CUDA, the software layer that makes its GPUs usable for AI workloads, and it sells the combined package at a gross margin of roughly 71–75%.
SK Hynix, Samsung (HBM)
71–75% gross margin
90% of revenue
Simplified. Segment mix is FY2026. Source: NVIDIA Q4 FY2026 earnings release, Feb 25, 2026.
Data Center is $193.7B of $215.9B in sales
Revenue by product line — FY2026
Source: NVIDIA Q4 FY2026 earnings release, Feb 25, 2026.
Source: 10-K FY2025, Item 7 (MD&A).
A narrow base of cloud buyers
Nvidia sells almost entirely to other businesses — cloud operators directly, or system integrators and distributors as a middle step. That customer base is narrow: in FY2025, three direct customers accounted for 12%, 11%, and 11% of revenue — 34% combined, all from the Data Center segment.
- AMD — the one company making GPUs that compete head-on with Nvidia's. Cheaper, but still well behind on the CUDA software ecosystem that locks developers in.
- Google, Amazon, and other hyperscalers' in-house chips (TPU, Trainium) — Nvidia's biggest customers are simultaneously building silicon to reduce how much they need from Nvidia. Its customers are becoming its competitors.
- Huawei — a domestic Chinese alternative gaining ground in the gap left by U.S. export restrictions on Nvidia's China sales.
Source: 10-K FY2025, Item 1A (Risk Factors — Competition), p.13.
Gross margin and one segment carry the story
For any chipmaker, gross margin is the clearest read on pricing power. And because Nvidia is now, in practice, a data-center company, the growth rate of that one segment is the whole investment story.
| Fiscal year | Revenue | YoY growth |
|---|---|---|
| FY2023 | $15.0B | — |
| FY2024 | $47.5B | +217% |
| FY2025 | $115.2B | +142% |
| FY2026 | $193.7B | +68% |
| Fiscal year | Gross margin |
|---|---|
| FY2023 | 56.9% |
| FY2024 | 72.7% |
| FY2025 | 75.0% |
| FY2026 | 71.1% |
The FY2026 dip has a specific cause: when Washington required export licenses for China-bound H20 chips in April 2025, Nvidia took a $4.5 billion inventory writedown in a single quarter — a real, dollar-denominated instance of export-control risk hitting the income statement.
Source: 10-K FY2025, p.41 & p.51 (FY23–25); 10-Q Q1 FY2027, "H20 excess inventory" disclosure (FY26).
A founder still running the company after 33 years
Jensen Huang founded Nvidia in 1993 and has been President and CEO for all 33 years since — one of the longest tenures of any founder still running a company this size. He previously worked as an engineer at LSI Logic and AMD. Huang and family trusts hold 3.58% of shares outstanding; all officers and directors combined hold 3.94%. The largest institutional holders are BlackRock (7.43%) and Vanguard (7.31%), as of March 2026.
Source: DEF 14A, filed May 12, 2026 — Security Ownership and executive biography sections.
A symbolic dividend, with buybacks doing the work
The dividend is symbolic rather than material: $0.01 per share quarterly, about $974 million total in FY2026 — just 2.4% of the $41.1 billion Nvidia returned to shareholders that year. The other 97.6% came through buybacks: roughly $40.1 billion in FY2026, up from about $33.7 billion in FY2025, against 24.31 billion diluted shares outstanding as of March 2026.
Source: NVIDIA Q4 FY2026 earnings release (Feb 25, 2026); DEF 14A (May 12, 2026).
What breaks Nvidia: hyperscalers build their own chips
- Customer concentration meets customer-as-competitor risk — the top three customers are 34% of revenue, and those same companies are the ones with the strongest motive and resources to build a substitute.
- U.S. export controls keep tightening — the 2025 H20 restriction already cost $4.5 billion. Further rules modeled on the "AI Diffusion" framework could restrict sales beyond China too.
- Supply chain concentrated in Taiwan and Asia-Pacific — TSMC and Nvidia's assembly-and-test partners are clustered in a single geopolitically sensitive region.
Source: 10-K FY2025, Item 1A (Risk Factors), p.13–24; 10-Q Q1 FY2027 (H20 disclosure).
Revenue went from $27B to $216B in three years
| FY21 | FY22 | FY23 | FY24 | FY25 | FY26 | |
|---|---|---|---|---|---|---|
| Revenue | 16,675 | 26,914 | 26,974 | 60,922 | 130,497 | 215,938 |
| YoY growth | — | +61% | +0.2% | +126% | +114% | +66% |
| Operating income | 4,532 | 10,041 | 4,224 | 32,972 | 81,453 | 130,387 |
| Operating margin | 27.2% | 37.3% | 15.7% | 54.1% | 62.4% | 60.4% |
| Operating cash flow | 5,822 | 9,108 | 5,641 | 28,090 | 64,089 | ~102,700 |
| Capital expenditure | 1,128 | 976 | 1,833 | 1,069 | 3,236 | ~6,000 |
| Free cash flow | 4,694 | 8,132 | 3,808 | 27,021 | 60,853 | ~96,700 |
| Total debt | n/a | n/a | n/a | 9,709 | 8,463 | ~8,500 |
Against $8.5 billion of debt, Nvidia held $62.6 billion in cash and securities at the end of FY2026 — a net cash position of $54.1 billion. Balance-sheet risk is close to non-existent.
Sources: 10-K FY2022 (FY21–22), 10-K FY2025 (FY23–25), Q4 FY2026 earnings release (FY26, press-release rounding). FCF = operating cash flow − capex, calculated directly.
Open questions: FY2026 figures still awaiting the 10-K
- The full FY2026 10-K wasn't yet available at analysis time — some FY2026 figures above (OCF, capex, debt) are press-release estimates pending confirmation against the filed annual report.
- Exact China revenue isn't disclosed — filings break out revenue by billing location only, not true end-use geography.
- Whether the next-generation Rubin architecture ships on schedule to reaccelerate growth can't be judged from this filing alone — the next earnings call is the place to check.
What we'd watch next
These three checks would show whether this read is holding up. They are prompts for your own follow-up, not predictions.
- Data Center share of revenueIt was $193.7B of $215.9B in FY2026. A falling share would mean AI spending is broadening beyond the giants, or cooling.
- Customer concentrationA few cloud buyers take most of the GPUs. Check each 10-K for any single buyer growing past its current share.
- Gross marginIt carries the pricing-power story. A sustained drop would be the first sign that hyperscalers' in-house chips are biting.
Frequently asked questions
What does Nvidia do?
Nvidia designs advanced computer chips called GPUs, has them manufactured by contract foundries like TSMC, and bundles them with its own CUDA software before selling the package — mostly to companies building AI data centers, plus PC gamers in a smaller way.
Does Nvidia manufacture its own chips?
No. Nvidia is a "fabless" chipmaker — it designs the chips but pays contract foundries, primarily TSMC, to actually manufacture them. It owns no factories of its own.
What is Nvidia's market cap?
As of this article's data, Nvidia's market cap was about $5.45 trillion, with FY2026 revenue of $215.9B and a P/E ratio of roughly 34.5x.