Chips · Supply Chain

Three Companies Stand Behind Every Advanced AI Chip

Last updated: 29 June 2026

As of June 2026, the most advanced AI chips depend on a chain that runs through three companies on three continents. ASML in the Netherlands builds the only machines that can print the smallest circuits. TSMC in Taiwan uses those machines to manufacture the chips. Nvidia in the United States designs the GPUs that train and run AI models, yet owns no factory. Break any single link and the leading edge of AI hardware stops.

A note on labels: figures marked claim are an interested party's stated number and not independently audited; derived figures are calculated here; projection figures are forecasts that have not yet happened.

One chain, three companies

A modern AI chip is the work of a division of labor so deep that no single company can complete it. Nvidia designs the GPU: the architecture, the logic, and the CUDA software that developers build on. It does not own a factory. It hands its designs to TSMC, the world's largest contract chipmaker, which prints them on silicon. TSMC, in turn, cannot make its most advanced chips without lithography machines that only ASML builds. Each company depends on the one before it, and the most advanced AI accelerators in the world all pass through the same three sets of hands.

What the machine actually does

A chip is a grid of transistors, the tiny switches that hold the ones and zeros of computation. The smaller the transistor, the more you can fit on a slice of silicon, and the faster and more efficient the chip. Printing features this small is a problem of light. You can only etch a pattern as fine as the wavelength of the light you use, and ordinary light is far too coarse.

ASML's answer is extreme ultraviolet light, or EUV, with a wavelength of 13.5 nanometers, thousands of times narrower than a human hair. Making it is brutal. The machine fires a laser at droplets of molten tin tens of thousands of times per second, vaporizing them into a plasma that emits the light. Because EUV is absorbed by air and by glass, the whole process runs in a vacuum and uses mirrors instead of lenses. Those mirrors, made by Germany's Zeiss, are among the smoothest surfaces ever manufactured. A single machine contains more than 100,000 parts and draws on a supply chain of thousands of specialist firms.

This is also where the price confusion starts, and it is worth being precise. A standard EUV machine (ASML's Low-NA, or NXE, line) sells for about $183 million. ASML's next-generation High-NA system (the EXE line) runs about $380 million. The two are not interchangeable, and the difference matters for the rest of this story.

Where the leading edge sits now

The frontier moved in late 2025. TSMC began volume production of its 2nm process in the fourth quarter of 2025, with Apple, Qualcomm, AMD, and Nvidia among the early customers. The 3nm node is now mainstream. Nvidia's current Blackwell GPUs, the chips at the center of the AI boom, are built on a custom TSMC 4nm-class process and pack 208 billion transistors. Nvidia's next architecture, Rubin, moves to a 3nm process.

Here is the part that trips up most coverage: all of today's best chips, including those 2nm and 3nm parts, are made on the standard $183 million EUV machine. The $380 million High-NA tool is not yet doing volume work anywhere. Intel installed the first one for research in late 2024, and TSMC has signaled it will not deploy High-NA in volume until around 2030. High-NA is the next frontier, not the current one. For now, the machine that makes the world's best chips is the standard EUV system, and that is the machine that matters for who is ahead.

Why no one else builds these machines

ASML's monopoly is not an accident, and it is not for lack of rivals trying. Japan's Nikon and Canon dominated lithography in the previous era and both looked hard at EUV. Both walked away, judging the technical risk and the capital required too high for the payoff. ASML spent roughly three decades and well over $10 billion bringing EUV to production while the outcome was still in doubt. That head start is now an accumulated lead measured in generations.

The deeper barrier is the supply chain. ASML does not make most of the machine itself. The optics come from Zeiss, the product of a development effort that ran more than 15 years. The high-power lasers come from Germany's Trumpf. ASML coordinates a network of thousands of suppliers, many of them effectively exclusive to it. A new entrant would have to rebuild that entire ecosystem from scratch. The economics do not invite anyone to try: the total customer base is only a handful of chipmakers, and ASML ships just a few dozen EUV systems a year. You cannot justify a decade and tens of billions of dollars to fight for a handful of buyers a rival already holds. And even a successful copy would be chasing a moving target, because ASML is already shipping High-NA and researching the generation beyond it.

How three chipmakers funded the monopoly

ASML's origins are humbler than its position suggests. It was founded on 1 April 1984 as ASM Lithography, a 50/50 joint venture between Philips, the Dutch electronics conglomerate, and ASM International (ASMI), a chip-equipment maker. Its entire mandate was to commercialize a wafer stepper, the PAS 2000, that Philips had developed in-house but could not sell. The new company started in a leaky shed next to a Philips office in Eindhoven.

Two people sit at the center of that founding. Wim Troost was the Philips manager who took on the stepper project when it moved out of the research labs in the late 1970s. He fought to keep it alive inside a conglomerate that had little appetite for its own lithography machines, at one point using a personal discretionary budget to fund it, and he spent years hunting for an outside partner after deals with Cobilt, Perkin-Elmer, and others fell through. He would later become ASML's second chief executive. Arthur del Prado was the entrepreneur on the other side of the table. He had founded ASM International and is widely regarded as the father of the European semiconductor-equipment industry. He had floated ASM on the Nasdaq in 1981, the first Dutch company to do so, which meant that when Troost came looking for a partner, del Prado had both the appetite and the cash to say yes.

Notably, there was no venture capital. ASML emerged from a conglomerate as a joint venture between two larger firms. Each parent contributed about $2.1 million, though the contributions were lopsided: ASMI put in cash, while Philips counted physical assets, including 17 PAS 2000 machines, toward its share. The 47 Philips engineers who transferred negotiated the right to return to Philips within four years if it failed, a clear sign of how little anyone expected. It nearly did fail. ASMI sold its half to Philips in 1988, Philips considered shutting the venture, and only a later machine, the PAS 5500, brought in the customers that made ASML viable. It went public on the Amsterdam and Nasdaq exchanges in 1995.

The most revealing chapter came in 2012. To finance the multibillion-dollar EUV gamble, ASML turned to the companies that needed EUV most. Under a customer co-investment program, Intel, TSMC, and Samsung together bought a 23 percent stake in ASML for 3.85 billion euros and committed a further 1.38 billion euros to its lithography research. Intel alone took a 15 percent stake, part of a set of agreements that totaled about $4.1 billion. The chipmakers that depend on the machines helped pay to invent them. The industry, in a precise sense, funded the monopoly it now cannot live without.

Why China cannot simply copy it

The United States and its allies have built export controls that go well beyond banning finished chips. They cover the EUV machines themselves, the chip-design (EDA) software, and certain specialized tools and materials. Crucially, the ban covers both EUV tiers. China cannot buy the $183 million standard machine that makes today's best chips, and it cannot buy the $380 million High-NA machine either. Acquiring its way to the frontier is not an option on the table.

China has made real progress within those limits. Its leading foundry, SMIC, has produced 7nm chips since 2022 and is ramping a 5nm-class process, both using older deep-ultraviolet (DUV) machines run through multiple exposure passes to reach features that would normally require EUV. The workaround is real but costly: yields are reported at roughly 20 to 40 percent claim, far below the levels TSMC achieves, and the economics lean on state subsidies. Huawei's Ascend accelerators, built on these lines, now serve a domestic AI market that Nvidia is largely barred from supplying.

China is also trying to build the machine itself. Reuters reported in December 2025, citing two people with knowledge of the project, that a team of former ASML engineers in Shenzhen had built a domestic EUV prototype by reverse-engineering older machines claim. The prototype can reportedly generate EUV light but has not yet produced working chips, and ASML's chief executive said in April 2025 that China would need many years to master the technology. A laboratory prototype is a long way from a production tool, and most observers see high-volume manufacturing this decade as unlikely. As for the gap, the popular framing of China being a decade or more behind overstates it in one direction and understates it in another. On process node, most analysts put China roughly 5 to 7 years behind the leading edge. On the equipment that defines the frontier, especially EUV, the gap may be longer, because reproducing the Zeiss optics and the tin-plasma light source is exactly the wall that stopped everyone else.

When the Bubble Pops

Frontier labs are spending enormous sums to train and run models, and many are not yet profitable. OpenAI projects a net loss of about $14 billion for 2026 projection, while Anthropic projects its first operating profit in the second quarter of 2026 projection (covered here). That gap between spending and profit is what drives the bubble question, and the answer depends on which layer you mean.

The infrastructure layer, the machines and fabs and chips, has demand anchored to models that already exist and have to run. The application layer, the new models and startups chasing the next product, is far more speculative: many will fail or be absorbed, and a few will not. A correction in the second does not automatically empty the first, which is why the supply chain in this article is a useful place to watch.

Who is most exposed if spending slows

The three companies in the chain are not equally exposed, and their financials show why. Nvidia is the most concentrated bet on AI: its data center segment generated $51.2 billion of $57.0 billion in total revenue in the third quarter of its 2026 fiscal year, roughly 90 percent of the company derived. TSMC is more cushioned, because it also makes chips for phones, cars, and laptops. ASML may be the most defensive of the three: its installed-base service and upgrade revenue reached nearly 2.5 billion euros in the first quarter of 2026, recurring income from machines already sitting in customers' fabs regardless of the next order.

From there the analysis runs to the buyers. The companies best positioned to keep funding AI through a downturn are the diversified platforms, Microsoft, Google, and Amazon, that can pay for it out of established businesses like cloud, advertising, and software, rather than needing AI to turn a profit on its own. None of this is a forecast of who wins. It is a map of who is insulated, drawn from where each company's money actually comes from today.

The bottom line

The AI story is usually told through models and demos. The leverage sits one or two layers down, in the tools that are hardest to replicate and the businesses diversified enough to ride out a cycle. ASML's single-supplier hold on EUV, TSMC's manufacturing scale, and the recurring revenue under both of them are more durable than any one model release. Understanding the supply chain does not make the technology less impressive. It just shows where the real choke points, and the real resilience, actually are.

Last updated: 29 June 2026

Sources

1. ASML, EUV lithography systems and TWINSCAN EXE:5000 product pages, asml.com (accessed June 2026).
2. ASML, "ASML's founding story," asml.com, 2024.
3. ASML Q1 2026 results: installed-base management revenue, reported April 2026.
4. EUV machine prices are ASML-stated figures (the company publishes no list price), reported consistently by Bloomberg, Reuters, and Tom's Hardware: High-NA about $380 million (350 million euros) versus about $183 million (170 million euros) for Low-NA NXE systems, 2024.
5. TSMC, statement on 2nm (N2) volume production in Q4 2025, via AFP and TechXplore, 31 December 2025.
6. Nvidia, Blackwell architecture (custom TSMC 4NP process, 208 billion transistors), nvidia.com.
7. Nvidia, Form 8-K, Q3 fiscal 2026 results (revenue $57.0B; data center $51.2B), U.S. SEC, 19 November 2025.
8. CSIS, "In Chip Race, China Gives Huawei the Steering Wheel," on export controls and SMIC 7nm, 2026.
9. American Enterprise Institute, "Lithography Loophole" report on SMIC and Huawei DUV multi-patterning, April 2026.
10. Reuters, exclusive on China's domestic EUV prototype (citing two people with knowledge of the project), 17 December 2025; corroborated by South China Morning Post, Engadget, and TrendForce relaying the same report.
11. ASML and Intel, 2012 Customer Co-Investment Program press releases and SEC filings (Intel 15 percent stake; Intel, TSMC, and Samsung an aggregate 23 percent for 3.85 billion euros).
12. ITIF and SemiAnalysis, analyst estimates of China's process-node gap (roughly 5 to 7 years).
13. AI Race Facts, "Anthropic Projects First Profit in Q2 2026, OpenAI Faces $14B Loss," 2026.

Pricing, model availability, and export rules change frequently. Verify against the primary source before relying on any figure here.