EUV

On the morning of April 11, 2001, in a high-bay clean room at Sandia’s Livermore campus, a few dozen engineers and a handful of suited officials stood in front of a contraption the size of a delivery truck and waited for someone to make a speech. The machine had no chassis in the usual sense. It was an exposed rack of stainless-steel vacuum chambers, optical benches, mirror cells, and laser plumbing, threaded together by cabling and cryogenic lines and watched over, screen by screen, from a control room next door. It was called the EUV Engineering Test Stand. Its lead designers, Daniel Tichenor and Glenn Kubiak of Sandia, had spent the better part of three years assembling it. That morning it had projected the first full-field images ever produced on American soil at a wavelength of 13.4 nanometers. The patterns it printed were a hundred nanometers across, a quarter the size of what the best deep-ultraviolet steppers in the industry could then resolve.

Chuck Gwyn, the program manager of an unusual entity called Extreme Ultraviolet Limited Liability Company, stepped to the microphone. He was a softly spoken physicist who had spent four decades inside Sandia and was now, in his last career, running a private consortium that had taken custody of one of the most ambitious public-private research programs the United States had ever attempted. The completion of the prototype, he told the audience, marked a major milestone for the program, since it had proven that EUV lithography worked. Behind him, Tom Hunter of Sandia and senior officials from the Department of Energy’s National Nuclear Security Administration nodded approvingly. Engineers from Intel, Motorola, AMD, Micron, Infineon, and IBM, the six private firms then bankrolling the consortium, watched a tool that the previous decade of work had insisted would never function actually function. Most of them believed they were within five years of putting a production version of this machine into a high-volume fab.

They were eighteen years away.

The miscalculation was not one of competence. It was a miscalculation of how hard a wavelength can be. The light that the machine in Livermore had just produced sat in a part of the electromagnetic spectrum that, for most practical purposes, no one had ever tried to use industrially. Visible light was easy. Ultraviolet was harder. Deep ultraviolet, at 248 and 193 nanometers, ran every advanced fab on Earth in 2001 and would still be running most of them a decade later. EUV at 13.5 nanometers was a different country. Air absorbed it. Glass absorbed it. Photoresist absorbed it greedily and unevenly. Every existing lithographic intuition about lenses, transmission, refraction, and atmosphere had to be discarded. To build a stepper at 13.5 nanometers meant accepting that nothing in the optical path could be made of anything you could see through, and that the entire machine, source, illumination, mask, projection, wafer, would have to live inside a vacuum, looking at its own reflections. Industry physicists had known this for a decade and a half. The thing they had not yet known was how many adjacent physics problems each of those constraints would generate.

The story had started, as so many semiconductor stories did, with a single skeptical audience in Japan. Hiroo Kinoshita, an engineer at Nippon Telegraph and Telephone’s R&D laboratories outside Tokyo, had been working through the early 1980s on the problem of pushing chip features below the diffraction limit of optical lithography. The conventional path was X-ray, but X-rays at the wavelengths that lithographers cared about were difficult to focus. In 1984, Kinoshita came across a paper by James Underwood and Troy Barbee describing a class of mirrors made of alternating ultra-thin layers of two materials, in their case molybdenum and silicon, that exploited Bragg diffraction to reflect specific narrow bands of soft X-ray and EUV light. A single Mo/Si pair reflected almost nothing. Stack fifty pairs of them, each only a few nanometers thick, and the reflectivity at 13.5 nanometers approached seventy percent, not a great mirror by visible-light standards but vastly better than the zero that any monolithic surface offered at that wavelength.

Kinoshita saw a path. A reduction stepper at 13.5 nanometers, built entirely out of multilayer mirrors instead of glass lenses, could in principle resolve features an order of magnitude smaller than the best lenses in the industry. He tested the idea on a benchtop and, by 1985, had projected an image with a Mo/Si reflective system. In 1986 he stood up at the annual meeting of the Japan Society of Applied Physics and presented the result. The reception, as he would tell colleagues for years afterward, was hostile. Senior figures in the room did not believe the pictures. NTT itself, more interested in the deep-ultraviolet roadmap that its chip-equipment partners were pursuing, did not push the technology forward, though Kinoshita’s manager allowed him to continue the work as a back-burner project. The first publication of EUV lithography, in other words, was greeted by its own audience as an embarrassment.

Across the Pacific, the same physics had attracted a smaller and even more obscure constituency. At Lawrence Livermore National Laboratory, a soft-X-ray imaging group led by figures including Natale Ceglio and Andrew Hawryluk had been developing multilayer optics for fusion diagnostics and for X-ray microscopy. At Bell Labs, a researcher named Tania Jewell, working with William Silfvast and Don White on plasma sources, had begun thinking about whether the same multilayer mirrors could be turned into a projection lithography system. In 1989, at a conference in Monterey, the American and Japanese strands met. Jewell, hearing Kinoshita’s results for the first time, pressed him with questions through an exhausting afternoon. Within a year, Bell Labs had printed fifty-nanometer features using a soft-X-ray projection rig. The hostile reception of three years earlier started, in the way these things did, to look like an overlooked breakthrough.

Through the early 1990s the problem migrated into the standard apparatus of American semiconductor R&D. Sandia, Lawrence Livermore, and Lawrence Berkeley pooled their soft-X-ray and multilayer expertise into a Virtual National Laboratory and began a serious program in what was now beginning to be called extreme ultraviolet lithography, the new name distinguishing the projection approach from the older proximity-mask soft-X-ray technique. The Defense Advanced Research Projects Agency, which had funded a parallel program of advanced lithography options through its Advanced Lithography Program since 1991, paid attention. So did Intel. By 1992, according to subsequent accounts assembled by industry historians and by the Center for Security and Emerging Technology, Andy Grove had approved roughly two hundred million dollars of internal Intel research into next-generation lithography, much of it directed toward EUV. Through the mid-1990s, Sandia and Livermore between them had spent on the order of thirty million dollars of Department of Energy money on the problem. The American program was cohering.

In 1996, Congress nearly killed it. As part of a broader squeeze on DOE applied-research budgets, the agency’s funding for EUV work was scheduled to be terminated. The decision would have left a substantial body of expertise in three national labs without a sponsor and stranded several years of multilayer-mirror, mask, and source research. Inside Intel, the prospect was unacceptable. The company by then was forecasting, on its own roadmap, that 193-nanometer optical lithography would run out of resolution somewhere around the early 2000s and that the industry would need a successor. Without EUV, the alternatives, electron-beam direct write, ion projection, X-ray proximity, were each, on Intel’s reading, either too slow for volume manufacturing or too immature to bet on. EUV was the option Intel believed would actually scale.

Grove, who had been Intel’s chief executive since 1987 and had become chairman in May 1997, took the decision personally. Intel committed to underwrite the national-lab work itself. On September 11, 1997, in a press release that Intel issued jointly with Motorola, Advanced Micro Devices, and the Department of Energy, the four parties announced that they had formed a private consortium to fund EUV development at Sandia, Livermore, and Berkeley. The legal instrument was a Cooperative Research and Development Agreement between the new consortium and the DOE labs. The name of the consortium, picked for the legal flexibility of the entity rather than for any romantic value, was EUV Limited Liability Company, soon shortened in the trade press to EUV LLC. The starting commitment from the three private members was about two hundred and fifty million dollars over six years, with Intel picking up the dominant share. By the time Micron, Infineon, and IBM joined in 2000 and 2001, the cumulative private commitment had grown well past three hundred million. The Energy Secretary at the time would describe it, in the press release that accompanied the deal, as the largest investment private industry had ever made into a Department of Energy program. Charles Gwyn, drawn out of Sandia management to run the consortium, became its program manager.

Gwyn inherited a political problem alongside a technical one. Through the 1980s and 1990s, the United States had ceded the lithography-tools market almost entirely to Japan and a small Dutch firm called ASML. Of the world’s three remaining commercial stepper makers, none were American. If EUV LLC’s research succeeded, the resulting tools would have to be built by someone who actually shipped lithography equipment, and Intel’s leadership knew that someone was not going to be a domestic supplier. The natural candidates were Nikon, the world leader, and Canon, its closest competitor. Intel pushed hard, through 1997 and 1998, for both Japanese firms to be admitted to the consortium. The DOE and the Clinton administration’s Commerce Department, looking at the long arc of Japan’s chip-trade fights with the United States and at the political optics of handing publicly funded research to firms that had crushed the American stepper industry a decade earlier, refused. Nikon withdrew. Canon was blocked. The third candidate was ASML. ASML was Dutch, not Japanese, and the Netherlands was a NATO ally with no recent history of trade conflict with the United States. After a year of internal argument, Washington agreed in 1999 to admit ASML to the EUV LLC arrangement, on the understanding that the company would source a substantial fraction of its EUV components from US suppliers and would establish a US factory to build the eventual production tools. ASML accepted the terms in writing. The factory and the component-share commitment were, in subsequent years, never seriously enforced.

Inside the American lithography community, the decision was not received well. The chief executive of Silicon Valley Group, the surviving US stepper maker, complained at the time that Intel had done everything in its power to give the technology to ASML on a silver platter. Two years later, ASML moved to acquire SVG outright in a deal valued at roughly $1.6 billion. The Bush administration, with Andy Grove and Craig Barrett of Intel publicly arguing for approval, cleared the merger in May 2001 over objections from members of Congress and from US lithography veterans who saw the last of the American optical-stepper industry being absorbed by a Dutch competitor. By the time Gwyn unveiled the Engineering Test Stand at Livermore that April, the long-term commercial home of EUV was already, in effect, decided. The technology had been incubated by Intel and three American national laboratories. The machines that would carry it to production would be built in Veldhoven.

The EUV LLC consortium ran through to 2003, met all of its formal technical milestones, filed more than a hundred and fifty patents, and produced the working Engineering Test Stand on schedule. When the consortium dissolved, ASML inherited the IP, the engineering know-how, and the responsibility for converting a national-lab prototype into a tool that could survive the floor of an Intel or TSMC fab. ASML’s chief technology officer, Martin van den Brink, a quiet engineer who had been at the company since its 1984 founding, had been arguing for years inside the firm that EUV was the only cost-effective long-term path beyond 193-nanometer immersion. In 2006 he gave a keynote at the EUV symposium in Lake Tahoe in which he laid out the case publicly. Inside ASML, the path was set. Van den Brink and a longtime ally, the physicist Frits van Hout, would spend the next decade and a half driving the development.

The first ASML prototypes, called Alpha Demo Tools, shipped in 2006 to two research consortia, IMEC in Leuven and the College of Nanoscale Science and Engineering at SUNY Albany. The tools functioned. They printed thirty-five-nanometer features. They ran at a throughput, by ASML’s later admission, of about one wafer in twenty-three hours, against a commercial target of more than a hundred and fifty wafers per hour. The gap between proof-of-concept and production was, in effect, four orders of magnitude in source power. Closing that gap turned into the longest, most expensive, and most technically improbable engineering campaign in the history of semiconductor manufacturing.

The mirrors came first. Carl Zeiss SMT, the spun-off semiconductor optics arm of the historic German optical firm in Oberkochen, had been making lenses for ASML since the 1980s. EUV did not need lenses. It needed mirrors so flat that, scaled up to the area of a continent, the largest hill on the surface would be half a millimeter high. Zeiss engineers had to learn how to deposit, on top of meters-long elliptical and aspheric substrates, fifty alternating layers of molybdenum and silicon, each layer between two and four nanometers thick, with no defect anywhere across the figure that exceeded a few atoms in height. Direct-current magnetron sputtering, the workhorse of Mo/Si deposition, was barely up to the task. The figure of merit was the surface deviation in picometers. By the late 2010s, Zeiss claimed an RMS surface roughness of about fifty picometers across a 450-millimeter mirror. There were perhaps six or seven mirrors per EUV system, each costing single-digit millions of euros, each requiring months of polishing and weeks of acceptance testing. Asianometry’s Jon Y, drawing on Zeiss technical disclosures, would later describe the result as a house of mirrors, the most precise reflective optical system anyone had ever built.

The light source was harder. The original EUV LLC research had explored two architectures for generating photons at 13.5 nanometers: a discharge-produced plasma and a laser-produced plasma. By the late 2000s, after years of competition between source vendors, the laser-produced plasma approach had won, in part because it could be scaled and in part because it kept debris from the plasma away from the collector mirror. The architecture, in outline, was simple to describe and almost unimaginable to build. A small generator at the back of the source vessel released spheres of molten tin, each about thirty micrometers in diameter, into a vacuum chamber at a rate of fifty thousand droplets per second, traveling across the chamber at about thirty meters per second. A high-power CO2 laser, generating ten-micrometer infrared light at twenty kilowatts and rising, fired pre-pulses and main pulses at every droplet as it passed through the focus, in a master-oscillator-power-amplifier configuration, hitting each droplet twice. The first pulse flattened the droplet into a disk; the second pulse vaporized the disk into a plasma at half a million kelvin, which radiated, among other wavelengths, the desired narrow band at 13.5 nanometers. A multilayer collector mirror gathered the EUV out of the plasma at solid angle and steered it into the rest of the optical column. The geometry had to repeat fifty thousand times a second, every second, for months on end, without the collector mirror accumulating enough tin debris to lose reflectivity.

Cymer, a San Diego firm that had spent the previous two decades supplying the high-power excimer lasers used in deep-ultraviolet steppers, was the company that had managed, through the 2000s, to push the laser-produced plasma source from a benchtop curiosity into something resembling an industrial subsystem. Its engineers, working out of a development site near the Pacific, demonstrated the first MOPA pre-pulse architecture, scaled the source from forty watts to one hundred and twenty-five watts, and pursued a roadmap that promised, eventually, two hundred and fifty watts. The progression was the bottleneck of the entire EUV program. Without source power, throughput was unviable; without throughput, the cost per wafer was unsupportable; without cost-per-wafer parity with deep-ultraviolet immersion, no fab would buy the tool. By 2010, ASML’s leadership had concluded that owning the source vendor outright was the only way to maintain the schedule. In October 2012, ASML announced an agreement to acquire Cymer for approximately 1.95 billion euros, or about 2.5 billion dollars; the deal closed in May 2013 after antitrust clearances in Washington, Brussels, and Seoul. The CO2 driver laser itself, the brute that produced the infrared pulses for Cymer’s source, was supplied by a fourth firm, the German laser builder Trumpf, which by the late 2010s was delivering forty-kilowatt pulsed CO2 systems containing, by Trumpf’s own count, more than four hundred and fifty thousand individual parts per laser.

By the time the Cymer acquisition closed, ASML’s three biggest customers had concluded that the company itself was at financial risk. The total cost of the EUV program, originally projected in the late 1990s as a roughly billion-dollar undertaking, had, by industry estimates, blown past five billion and was still climbing. Development was eating ASML’s earnings. In 2012, in a structure ASML labeled the Customer Co-Investment Program, Intel, TSMC, and Samsung agreed jointly to put 1.38 billion euros of cash directly into ASML’s research and development over five years, on top of buying minority equity stakes that altogether transferred 23 percent of ASML’s shares to the three customers for a further 3.85 billion euros in cash. Intel’s stake was 15 percent, TSMC’s was 5 percent, and Samsung’s was 3 percent. Intel’s chief executive Paul Otellini and ASML’s chief executive Eric Meurice presented the arrangement publicly as a roadmap accelerator. Privately, the ASML-side teams, as the Dutch journalist Marc Hijink would later document in his 2024 book on the company, described it as the rescue that kept the EUV program alive. Without the customer co-investment, the development program would have been forced to slow.

Even with the rescue, the schedule kept slipping. The first pre-production tools, the NXE:3300 series, shipped in 2013, but with light sources that produced only about thirty watts at the wafer, not enough for an economically viable wafer pass. The next-generation NXE:3350 lifted that to forty watts. Customers began running R&D wafers but not yet commercial silicon. Through 2014 and 2015 ASML and its supplier base pushed the source toward eighty watts, then a hundred and ten, then a hundred and twenty-five. Each increment required a new generation of pre-pulse timing, droplet-generator stability, collector-mirror coatings, and CO2 laser power-handling. Each one looked, from the outside, like another year of disappointment, and inside the program, like a step that would not have been possible without the previous five.

In 2016, ASML began taking orders for the NXE:3400B, the first scanner intended explicitly for high-volume manufacturing. Its source produced two hundred and fifty watts at the wafer, sustained at a duty cycle that allowed wafer-per-hour throughput in the range commercial fabs needed. ASML shipped the first units in 2017. The second-half-of-2018 readiness milestone, which the company had been promising for nearly a decade, was met. Through 2018, ASML built and shipped eighteen NXE:3400 systems. By the start of 2020, the company would mark the delivery of its hundredth EUV system to a customer. None of those systems was made anywhere except in Veldhoven.

The customer that pulled EUV across the line into commercial production was not Intel. It was TSMC. Through the early 2010s, the Taiwanese foundry had matched Intel and Samsung as one of the three customers actively co-funding the development. By the second half of the decade, TSMC had committed in a way that the others had not. The company’s R&D leadership, with Mark Liu and Y.J. Mii driving the manufacturing roadmap, had decided that TSMC’s path to seven-nanometer-class processes would adopt EUV first as a limited insertion, on the most pattern-dense layers of the so-called N7+ node, and then as a deep insertion at the five-nanometer node that would follow. In April 2019, TSMC’s first wafers from N7+, an enhancement of its existing N7 process that used EUV on four of the most challenging metal layers, came out of the fab in Hsinchu. By the second quarter, the node was in volume manufacturing. In its press release of October 7, 2019, TSMC was able to say, with no fear of contradiction, that N7+ was the first EUV process delivering customer products to market in high volume. Within months, Samsung’s Exynos 9825, fabricated on its 7LPP process, became the first mass-market mobile chip built end-to-end using EUV. Apple’s A14, manufactured by TSMC on N5 and shipped in late 2020, would be the first high-volume processor in any consumer product to depend pervasively on the technology.

Intel was not in that list. The company had, through the 2010s, taken a calculated bet that 193-nanometer immersion lithography combined with aggressive multi-patterning, the practice of exposing the same layer through two or three or even four different masks to get effective feature pitches finer than what a single exposure could produce, would carry it through its 10-nanometer node and into 7-nanometer without committing to EUV at the same nodes as TSMC and Samsung. The bet was internally consistent, was endorsed by some of the company’s most respected manufacturing engineers, and was supported by Intel’s then-current cost models. It was wrong. The full account of how wrong it was would only become clear in the decade that followed. What mattered for the EUV story was that, by the time the technology actually entered high-volume manufacturing, the only customer that had committed to a wholehearted insertion at the leading edge was the foundry the United States no longer had a domestic peer for. The asymmetry of that decision, between an American chipmaker that had funded the technology from its DARPA-era beginnings and a Taiwanese foundry that had spent the decade learning to deploy it, would shape the next round of the geopolitics.

The arithmetic of EUV, taken in aggregate, was hard to read except as a verdict on what the chip industry, by 2020, had become. The wavelength itself had been demonstrated in a Japanese lab in 1985 by an engineer whose audience did not believe him. The American national-lab program had spent more than a decade and roughly a billion dollars in mixed public and private money before any tool capable of printing real wafers existed. Intel’s $250 million commitment in 1997 had grown, by the time of the 2012 customer co-investment, into a financial exposure across Intel, TSMC, and Samsung approaching ten billion dollars. ASML’s own internal R&D spend on EUV between 2000 and 2020 ran past sixteen billion euros. Counting Cymer, Trumpf, Zeiss SMT, the SUNY Albany consortium, IMEC, the SEMATECH mask-blank programs led by Stefan Wurm and others, and the years of national-lab funding that came before, the cumulative bet on a single number, 13.5 nanometers, almost certainly exceeded forty billion dollars across two and a half decades. No other technology effort in modern industry came close on either dimension. The Manhattan Project ran shorter and on a comparable real-dollar budget; Apollo had been bigger but had run a quarter the duration. EUV was the largest single technological wager any private industry had ever attempted to win on its own terms.

The wager paid off, but it paid off into an unrecognizable industry. The country that had funded the science had no firm at the finish line. The country that had supplied two of the three largest customers, the United States, had no domestic stepper supplier. The country that built the optics, Germany, did so as a single subsidiary of a single firm, in a single building in Oberkochen. The country that built the laser source, also Germany, did so out of a single Trumpf facility. The company that integrated all of it sat in a small Dutch town on the southern edge of the European Union, employed about thirty-three thousand people worldwide, and shipped fewer than fifty machines per year, each weighing about one hundred and fifty metric tons, comprising on the order of a hundred thousand individual parts, requiring six months and roughly two hundred and fifty engineers to install on a customer’s floor, and selling, by 2020, for about one hundred and seventy million dollars apiece. Whether any other entity on Earth could have replicated that supply chain, and whether anyone other than its three biggest customers could now safely rely on it, were the questions that the next decade of geopolitics would be organized around.

In Livermore in April 2001, Charles Gwyn had told the press that the work had proved EUV lithography worked. He had been right. He had also, without quite realizing it, marked the moment at which the United States stopped being the natural home of the technology its own laboratories had brought into existence. The thing that mattered, in retrospect, was not whether the bet would be won. It was who would be holding the chips when it was.