An AI chip is a processor built to do machine-learning arithmetic quickly, and making one takes several separate companies: a designer, a foundry that manufactures the silicon, memory suppliers, and firms that package and assemble the result. Annual reports filed with the U.S. Securities and Exchange Commission describe this chain, and none of the companies named in them performs every step.
Key takeaways
- Georgetown's Center for Security and Emerging Technology groups AI chips into GPUs, FPGAs, and ASICs built for AI.
- NVIDIA's annual report for the fiscal year ended January 25, 2026 describes a fabless model, with TSMC and Samsung named as wafer foundries.
- That report names SK Hynix, Micron Technology, and Samsung as memory suppliers and describes its supply chain as mainly concentrated in Asia.
- TSMC's 2025 annual filing says its ten largest customers made up approximately 78 percent of net revenue.
- Quarterly reports issued after those filings (NVIDIA in August 2026, TSMC in October 2026) show both companies still growing fast, but the annual filings remain the latest full descriptions of the supply chain, and neither ranks companies.
What makes a chip an AI chip
The Center for Security and Emerging Technology (CSET) at Georgetown University defined the term in an April 2020 report by Saif M. Khan and Alexander Mann. It describes AI chips as chips that reach high efficiency and speed on AI-specific calculations, at the cost of slower performance on other tasks. The report sorts them into three types. Graphics processing units (GPUs) mostly train algorithms, field-programmable gate arrays (FPGAs) mostly run trained algorithms on new data, and application-specific integrated circuits (ASICs) built for AI can do either job.
According to the report, three design choices explain the advantage over general-purpose processors: running many calculations in parallel, using lower-precision arithmetic that needs fewer transistors, and speeding up memory access. CSET states that AI chips can be tens or even thousands of times faster and more efficient than central processing units for these workloads. The figure dates from 2020, and the report analyzes the technology without comparing products.
Who does which job
The table maps the roles that appear in two primary sources: NVIDIA's annual report on Form 10-K for the fiscal year ended January 25, 2026 (the latest annual report as of October 2026), and TSMC's annual report on Form 20-F for 2025, filed April 16, 2026. It simplifies. A company named in one row is not the only one that could fill it, and the filings do not say which foundry makes which product.
| Step | What happens | Companies named in the filings |
|---|---|---|
| Design | The company specifies the processor and its software | NVIDIA describes itself as using a fabless and contracting manufacturing strategy |
| Wafer fabrication | A foundry manufactures the silicon | TSMC and Samsung (named by NVIDIA) |
| Memory | High-speed memory is supplied to sit next to the processor | SK Hynix, Micron Technology, and Samsung |
| Packaging | Chip and memory are joined into one module | NVIDIA says it uses CoWoS packaging technology, a TSMC trademark |
| Assembly and test | Modules become finished boards and systems | Hon Hai Precision Industry, Wistron, and Fabrinet |
What the NVIDIA filing shows about scale and concentration
NVIDIA's Form 10-K states that revenue for fiscal year 2026 was USD 215.9 billion, up 65 percent from a year earlier, and that Data Center revenue grew 68 percent. It also reports that sales to one direct customer represented 22 percent of total revenue and sales to another represented 14 percent, and that 31 percent of revenue came from outside the United States. On supply, the filing says the chain is mainly concentrated in Asia while the company is expanding into the United States and Latin America.
These numbers describe one company in one fiscal year. They do not measure the whole market, and they are not a forecast. Because a filing is a legal disclosure with a risk-factor section, its attention to concentration signals that revenue and supply rest on a small number of counterparties.
What the foundry filing adds
TSMC's Form 20-F says a vast majority of its revenue comes from customers that use its products in high performance computing (including AI applications), smartphones, the Internet of Things, automotive, and digital consumer electronics. In its risk factors, it reports that its ten largest customers accounted for approximately 70 percent of net revenue in 2023, 76 percent in 2024, and 78 percent in 2025. The largest customer made up 19 percent of 2025 revenue and the second largest 17 percent. The filing also notes that some system companies now design their own chips and work directly with foundries.
That last point is why a list of "AI chip companies" can mislead. A cloud operator that designs an accelerator for its own data centers is a chip designer in one sense and a customer in another, while the manufacturing still passes through a foundry. The roles in the table can sit inside different corporate boundaries from one product to the next. Stanford HAI's 2026 AI Index makes the same point about geography: it reports that TSMC fabricates nearly every leading AI chip, which concentrates the supply chain in Taiwan.
Quarterly results since the annual filings
Both companies have reported again since the annual documents above. NVIDIA's press release of August 26, 2026, for the quarter ended July 26, 2026, gave revenue of USD 96.2 billion (up 106 percent from a year earlier) and Data Center revenue of USD 89.0 billion (up 117 percent), with an outlook of USD 108.0 billion, plus or minus two percent, for the next quarter. TSMC's report for the third quarter of 2026, released October 15, 2026, gave consolidated revenue of USD 46.72 billion and said technologies of 7nm and more advanced made up 79 percent of wafer revenue. A press release is not an audited annual report and does not replace the supply-chain descriptions in the 10-K and 20-F. The next annual filings, for NVIDIA's fiscal year 2027 and TSMC's 2026, are expected in 2027.
Questions to ask about any AI chip claim
- Is the claim about design, manufacturing, memory, packaging, or the finished system?
- Does the number come from a filing or regulator, or from a market analyst's estimate?
- Which fiscal year and unit (revenue, units shipped, or capacity) does it use?
- Is a performance figure measured on a public benchmark with disclosed rules?
- Does the company depend on a single foundry or a few customers, as its risk factors describe?
FAQ
Which company is the largest AI chip maker?
The filings cited here do not rank companies, and no neutral source for a ranking was verified. "Largest" depends on whether the measure is revenue, units, or computing capacity.
What does fabless mean?
A fabless company designs chips but does not own the factory that manufactures them. NVIDIA's annual report uses the phrase "fabless and contracting manufacturing strategy" and names outside foundries for wafer fabrication.
Are GPUs the only kind of AI chip?
No. CSET lists GPUs, FPGAs, and AI-specific ASICs, and the TSMC filing notes that some system companies design their own chips.
Does this article say which chip to buy?
No. It explains the structure of the industry from public filings and gives no purchasing, investment, or performance guidance.
Related reading
- AI and robotics
- Leading AI researchers
- How AI helps companies cut costs
- AI companies and their products
More in AI Tools and Industry.
Last verified: October 2026
Sources
- AI Chips: What They Are and Why They Matter, Saif M. Khan and Alexander Mann, Center for Security and Emerging Technology, April 2020 (source page)
- Form 10-K for the fiscal year ended January 25, 2026, NVIDIA Corporation, U.S. Securities and Exchange Commission
- Form 20-F for the year ended December 31, 2025, Taiwan Semiconductor Manufacturing Company Limited, U.S. Securities and Exchange Commission, filed April 16, 2026
- NVIDIA Announces Financial Results for Second Quarter Fiscal 2027, NVIDIA Corporation, August 26, 2026
- Form 6-K, Third Quarter 2026 results, Taiwan Semiconductor Manufacturing Company Limited, U.S. Securities and Exchange Commission, October 15, 2026
- The 2026 AI Index Report, Stanford Institute for Human-Centered AI, 2026