“The only operators are the ones in the ceiling,” says Chris Belfi, wrapped up in a Tyvek bunny suit, tinted yellow under the photo-safe lights. The robots rush by on overhead tracks, blinking and whirring. Every few seconds, one pauses above a giant machine. Out of its laundry-basket-size belly, a plastic box drops on thin wires, like Tom Cruise in a catsuit. It holds precious cargo: up to 25 shiny silicon wafers, each the size of a 12-in. pizza. The process of transforming them into tiny computer brains—call them microchips, semiconductors or just chips—takes nearly three months. “I use an analogy like baking a cake,” says Belfi, an automation engineer at chipmaker GlobalFoundries. “The only difference is our cake is about 66 layers.”
This $15 billion complex tucked away behind trees north of Albany, N.Y., is one of only a handful of advanced semiconductor factories, or “fabs,” in the U.S. Its receiving docks pull in 256 specialty chemicals, like argon and sulfuric acid. Its shipping docks send out finished wafers, ready to be cut up, encased in metal and ceramic shells, and assembled into everything from airbags to blenders, headphones to fighter jets.
Since it opened in 2011, Fab 8, as it’s known, has kept a low profile. But as with toilet paper and chicken wings, the pandemic shocked the global semiconductor supply chain, leading to shortages in surprising places—and pulling the U.S. semiconductor industry to center stage. The car industry has been hardest hit of all. When the initial lockdowns caused car sales to collapse, automakers cut their orders for parts, including semiconductors. (A typical new car can contain more than a thousand chips.) Chip manufacturers saw the slack and shifted their output to serve the surging demand for consumer electronics, like webcams and laptops.
But when car sales snapped back last fall, a dramatic misstep became apparent: the automakers couldn’t get enough chips. They still can’t. Missing chips are now expected to lower global output by 3.9 million vehicles in 2021, or 4.6%. Ford alone expects to produce 1.1 million fewer vehicles, leading to a $2.5 billion earnings hit. (Even before it gets to its silicon-heavy electric F-150.)
As automakers and chipmakers scrambled for equilibrium, the White House stepped in to help, urging industry leaders to untangle the supply chain and increase production. The problem wasn’t only that there weren’t enough chips being made in America. The problem was that no one was paying attention to where chips were being made at all—and, more important, how long it took to make them. In June, the Senate passed a bipartisan bill with $52 billion in funding aimed at increasing chip production and cutting-edge research—competing directly with China’s ambitions of becoming the global semiconductor champion. But new chip fabs take years. Analysts now worry that the auto chip shortage will slosh back to consumer electronics, affecting manufacturing all the way to Christmas. “Never seen anything like it,” tweeted Tesla CEO Elon Musk.
Microchips, long revered as the brains of modern society, have become its biggest headache. The stakes extend beyond pandemic-era shortages. Because chips are a crucial component of so many strategic technologies—from renewable energy and artificial intelligence to robots and cybersecurity—their manufacturing has become a geopolitical thorn. In the 20th century, oil was the supreme global resource. But this year’s shortages have prompted a 21st century catchphrase among policymakers and diplomats: Chips are the new oil. As the U.S. resets to post-pandemic life, a steady supply of semiconductors has become a high-priority benchmark of preparedness and resilience. Except there’s a bigger problem: semiconductors were invented in the U.S., but fabs like GlobalFoundries have become a dying breed. In 1990, 37% of chips were made in American factories, but by 2020 that number had declined to just 12%. All the new pieces of the growing pie had gone to Asia: Taiwan, South Korea and China. Chip fabs aren’t just factories, but linchpins of American self-reliance.
Semiconductors are astonishing, with billions of transistors packed into a space the size of a dime, and they are astonishingly hard to make. If Henry Ford imagined an assembly line, a silicon wafer’s path through a factory is more like a labyrinth. At GlobalFoundries, the journey from raw material to finished chip—what engineers like Belfi call the “process flow”—is typically 85 days and encompasses more than a thousand steps. The whole time, the chips travel in sealed pods called FOUPs, entirely untouched by human hands. The robots do the driving, careening on their suspended tracks above machines the size of small RVs. One polishes wafers with a slurry that acts like liquid sandpaper. Another uses lasers to imprint circuits just 12 nanometers wide—about the length your fingernail grows in 12 seconds. Electron microscopes inspect the wafers for imperfections, and a robotic arm immerses them 25 at a time into a chemical bath like a carnival dunk tank. “We basically are bouncing wafers to and from each section of the fab, all day every day,” Belfi says. “It’s a lot of putting things on, taking it off, printing, putting more on, taking more off.” Humans intrude only when something goes wrong. The showstopper is a problem with one of the lithography machines, which set the pace of the whole operation. Each costs more than $100 million. “When one of those go down, it is all hands on deck,” says Belfi.Yet when COVID-19 hit, the fab never stopped. “We never shut down a single factory—not for an hour,” recalls Tom Caulfield, CEO of GlobalFoundries, in his office two levels above the fab floor. Long accustomed to wearing masks and full PPE, the engineers kept their usual watch on the robots. The business shocks proved harder to handle. As in so many industries, Caulfield’s initial financial models left him bracing for the worst. “We told our team, ‘We entered this pandemic together; we’re going to exit together. The world needs us to continue to make semiconductors.’” It was not an understatement. Chips powered the pandemic response—webcams, laptops, COVID-19 testing machines. In New York City alone, the department of education purchased 350,000 iPads.
The only thing it seemed no one needed was a new car, at least at first. Sales were off by a third in April, May and June 2020. Auto-component makers—not the brand-name car companies but their suppliers, and their suppliers’ suppliers—canceled orders.
But a semiconductor fab can’t turn on a dime. Foundry is the chip-industry term for a contract manufacturer, like a $15 billion Kinko’s. GlobalFoundries alone prints chips for more than 250 customers, which in turn supply components to device manufacturers—big, familiar names like Apple or Samsung, as well as industrial brands like Continental or Bosch. The supply chains are long. It takes three months to bake a chip, but it won’t end up in a car engine or smart speaker for months beyond that—and in consumers’ hands for more months still. On any given day at GlobalFoundries, there might be only 10 different kinds of chips in some phase of production. Each unique new chip design arrives on a 6-sq.-in. piece of quartz glass called a reticle. Like an old photographic slide, it contains a map of the chip, ready to be projected onto the silicon wafer with lasers. A reticle is its own trade secret, a protected piece of intellectual property belonging to the company that designed it, and adjusted to the unique specifications of GlobalFoundries’ proprietary process. Switching fabs isn’t easy, and definitely isn’t quick.
Around Thanksgiving, eight months into the pandemic, Caulfield’s phone started ringing. Auto executives who had never heard of GlobalFoundries were realizing they couldn’t make cars without them. “There is no way if you’re a supply-chain owner of an auto-manufacturing company, and you didn’t ship a car because you didn’t have a $5 or $10 chip, that you’re ever going to let this happen again,” Caulfield says. By New Year’s, the implications were alarming. In 2019, the auto industry spent $43 billion on chips—but they made up just 10% of the total chip market. The world’s largest foundry, Taiwan Semiconductor Manufacturing Company (TSMC), supplies more chips than anyone else to the automotive industry—but the automotive industry makes up just 3% of its revenue. (Apple makes up more than 20%.) At GlobalFoundries, chips destined for cars accounted for less than 10% of its business—enough to matter, but not enough to set the clocks.
That changed this year, when the political stakes for chipmakers rose dramatically. Caulfield called on his engineers to “remix their output,” sidelining some orders and prioritizing car chips. “We made very difficult decisions,” Caulfield says. “Wherever we could create more capacity, we gave it to automotive to make sure we were no longer the gate to manufacture.” The implications were obvious to him. “I didn’t need a letter from the White House to do the right thing.”
But the White House was paying attention. Before this year, lawmakers saw chipmaking as a local economic issue. Senate majority leader Chuck Schumer championed the construction of GlobalFoundries’ fab in his home state of New York, and lobbied for more. But the pandemic revealed how the decline in chipmaking wasn’t only about jobs and regional economic impacts, but also had strategic implications at a global scale. In February, President Biden signed an Executive Order launching a review of critical supply chains. If a pandemic could cause chip-induced shocks across industries, what else might happen? “We explicitly saw geopolitics as one of the risks to our supply chains,” says a senior Administration official.
In the earliest days of semiconductor manufacturing, in the 1960s, Intel co-founder Gordon Moore observed that he and his colleagues were able to double the number of transistors they could squeeze onto a chip every year. A decade later, that pace of innovation slowed (a little) to a doubling merely every two years. But then, it held. Chips got faster, cheaper and more efficient, eventually achieving a kind of social liftoff—powering computers that fit in a pocket.
But Moore’s law, as it became known, isn’t a fact of nature but the hard-earned result of enormous expenditures on research and development to conceive of new chip designs, and to find ways to manufacture them. Each new generation of semiconductor requires, in effect, a new factory to make it. Each new “process,” as it’s known, is named for the size of the chip’s smallest feature, like a jeweler whose fingertips keep getting finer. In the 1970s, chips were measured in micrometers, or one millionth of a meter. Since the 1980s, “leading edge” fabs have measured their handiwork in nanometers, or one billionth of a meter. Today, the benchmark is set by TSMC, which runs the 5-nanometer fab that makes Apple’s new M1 chips, for its latest computers and tablets. Each step down in process size—and increase in performance—requires the fab to be retooled with the latest generation of lithography machines to “print” the chips, along with the fleet of equipment that rings them. The newest fabs cost at least $12 billion.
Today, the semiconductor industry has predominantly split between the “fabless” companies that design chips and the foundries that make them. “There’s really two giant pieces of fixed costs,” says Chad Bown, an economist at the Peterson Institute for International Economics. “One set of fixed cost is all of the R&D—you need to come up with the chip’s ideas. The other fixed cost is all of the capital equipment—you need to build one of these fabs.” For decades, the leading companies—like Intel—did both. “Real men have fabs,” Jerry Sanders, former CEO of Advanced Micro Devices, famously insisted in the 1990s, in a comment as sexist as it is now outdated. AMD went fabless in 2009—selling its chip factories to GlobalFoundries.
But it was TSMC in Taiwan that pioneered the division between fab and fabless, and that dominates the industry today. Its founder, Morris Chang, was born in China and educated in the U.S. When he worked at Texas Instruments in the 1960s and 1970s, his engineering creativity and management acumen were legendary. He helped improve a chip fab’s “yield”—the number of chips good enough to be sold—and drive down prices. But after managing TI’s entire semiconductor division, Chang came to the conclusion that he would never be an American CEO. “I felt that essentially I had been put out in the pasture,” Chang recalled in a published oral history. “My hope of further advancement was gone.”
While at Texas Instruments, Chang had seen the rising cost of chip factories, and the way in which it can hold back innovation. His colleagues were eager to strike out on their own with new innovations, but never could. The barriers to entry were too high, if you had to run your own fab. He seized on the idea of a “pure-play foundry,” as he called it—a merchant semiconductor company focused on making chips for others—and, with support from Taiwan’s government, founded TSMC in 1987.
It was an auspicious moment in global trade policy. The Reagan Administration had enacted policies to counter rising Japanese production of semiconductors—but the first Bush Administration took a more hands-off approach. “Potato chips, computer chips, what’s the difference?” Michael Boskin, an economic adviser to George H.W. Bush, famously said. Along with TSMC in Taiwan, South Korean and Chinese companies began ramping up chip manufacturing, constructing high-tech fabs, often with the help of government subsidies. By the 1990s, even Intel began shifting some production overseas. “And while that’s all happening, we are adhering to our market-economics principles strongly—obsessively,” says John Neuffer, CEO of the Semiconductor Industry Association, a leading trade group. The prevailing attitude was to leave companies alone—to keep government out of business.
When in 2010 Apple announced its first custom-designed chip, the A4, it was self-evident in the semiconductor industry that it would be made by a foundry in Asia. This particularly stung Intel, which until then had both designed and manufactured the chips Apple used in its devices. But that division in the industry had become the norm, and it continues today: the most sophisticated chips are likely to still be designed in America but manufactured overseas. Silicon Valley still deserves its name—but only thanks to the dozens of fabless chip-design startups, not the foundries that once replaced its fruit orchards. Without the government incentives offered by Asian nations, that’s unlikely to change. “The U.S. policy was ‘We don’t create winners; we let capitalism work,’” says GlobalFoundries’ Caulfield. “That’s a great philosophy if everybody around the world plays that way.”
Now U.S. lawmakers are changing course. June’s Senate bill, officially the U.S. Innovation and Competition Act, is squarely aimed at competing with China, in part by subsidizing semiconductor manufacturing. “There’s been a growing bipartisan consensus over several years now that the U.S. needs to make more domestic investments to keep our competitive edge, particularly in an era of sort of strategic competition with China,” says a senior Administration official. “It was an easy political oversight,” says Neuffer, of the Semiconductor Industry Association. “It’s not an oversight anymore.”
The aggressive trade policy of the Trump Administration opened the door for Republicans to support greater economic intervention. Now they are clamoring for more. “There really has been a mistake here,” says Oren Cass, executive director of American Compass, a conservative think tank, who sees semiconductors as the “ultimate case study” for this necessary shift in American economic policy. “They are so obviously high-tech, they were an American area of dominance for so long, they so obviously have national-security implications, and you can so nicely quantify who is ahead or behind,” Cass says. “It crystallizes the issue in a way few other things could.”
For President Biden, semiconductors are an opportunity to stimulate high-tech American industry. “This is infrastructure,” the President said in April, holding up a glinting silicon wafer at the White House.
But bringing leading-edge semiconductor manufacturing to the U.S. will take years. In March, Intel announced plans for a $20 billion expansion of its factory in Arizona designed specifically to manufacture chips for others, as part of a newly launched division, Intel Foundry Services. But unlike TSMC—itself building a new factory estimated to cost $10 billion to $12 billion in Arizona, and possibly more–it will not be able to manufacture the cutting-edge chips. Smaller means faster, because you can squeeze more transistors into less space. But Intel has struggled to bring even its 7-nanometer node online, while TSMC is moving beyond its 5-nanometer node and preparing a 3-nanometer node for production. (Apple will reportedly again be its major customer.) It’s about “capacity and capability,” says Intel’s Al Thompson, vice president of U.S. government relations. “We’re going to spend a great deal of money in the U.S., creating a lot of jobs, to put us on a path to ensure that we’re doing our part to protect our nation’s economic and national security.” But there is a long way to catch up.
In April, a Silicon Valley startup called Cerebras announced a new computer called the CS-2. It’s meant not for Zoom calls or Netflix parties, but for the most sophisticated research in artificial intelligence—like discovering cancer drugs or simulating fusion reactions. At its heart is a custom-designed chip with a remarkable new design.
Rather than chop up a 12-in. silicon wafer into hundreds of tiny chips—punching each one out like a gingerbread cookie—Cerebras has found a way to make a single giant chip, like a cookie cake. Today’s smartphone chips contain billions of tiny transistors, etched into silicon like a miniature city. Cerebras’ custom chip contains trillions of transistors. To make it work, Cerebras engineers found a way to work around a basic flaw of silicon wafers. Typically, they are sliced from an ingot, like deli salami. But even the most sophisticated of these crystalline disks have imperfections, which ruin a chip. Semiconductor designers and manufacturers get around this by keeping each chip small, and throwing out the bad ones. (The yield is what’s left.)
The engineers at Cerebras created a design with 850,000 identical blocks, like wallpaper, and a system to turn off any flawed sections without ruining the entire chip. Most supercomputers chain thousands of individual chips together. But moving information between those chips slows things down. Cerebras keeps it all on a single giant chip. “We found a way to use the fact that silicon moves information at nearly the speed of light, and at tiny fractions of the power taken to move bits elsewhere,” says Cerebras CEO Andrew Feldman. For customers like the drugmaker GlaxoSmithKline and the Argonne National Laboratory, it provides the horsepower needed for breakthroughs in artificial intelligence—a key ambition of U.S. technology policy.
It’s the kind of bold idea that defined the early days of Silicon Valley innovation, and the ongoing creativity of American chip designers. But if in the 1960s and 1970s, chip fabs were all over the valley, when it was time for Cerebras to find a fab for its chip, there were no local options. “There are only two choices if you want to build chips at the cutting edge,” Feldman says portentously. “You can swim to China from one, and you can throw a stone to the DMZ at the other”—meaning Taiwan’s TSMC and South Korea’s Samsung. Like Apple—whose headquarters are just a 10-minute drive away—Cerebras uses TSMC for its manufacturing, using its 7-nanometer fab.
If the U.S. Innovation and Competition Act survives its journey through the House and becomes a law, billions of federal dollars will flow into the semiconductor industry—already one of the most profitable. But it will take years to turn that investment into new chip factories, new chip designs and a new pipeline of engineering talent. The challenge for the industry isn’t merely to catch up to where Taiwan is today, and China plans to be by 2025, but to meet them where they’ll be in the future—or go further.
Except chips improve nonlinearly. A 3-nanometer node is more expensive than a 5-nanometer. “The amount of money it takes to stay ahead of that curve keeps going up,” says Alisa Scherer, an independent chip-manufacturing expert. And the time span does too. It is almost impossible to skip a generation. The environmental permitting alone can take years. “These aren’t Taco Bells,” says Feldman. “You don’t knock them out. They don’t arrive in a box.” For his giant AI chips, TSMC and Samsung are the only two choices, “as far as the eye can see.”
Until then, chip manufacturing—and all of the geopolitical and economic reverberations it causes—will continue to depend on a global web, stretched delicately across the oceans.
—With reporting by Barbara Maddux and Simmone Shah
Blum is the author of Tubes: A Journey to the Center of the Internet and The Weather Machine: A Journey Inside the Forecast