Shortages and Supply Chains

At a quarter past two in the morning, Tokyo time, on Friday, March 19, 2021, an automated alarm in the N3 building of Renesas Electronics’ Naka factory, an hour and a half northeast of Tokyo in the city of Hitachinaka, signaled an overcurrent in a piece of plating equipment on the first-floor cleanroom. The plating tool, used to deposit thin metal layers onto twelve-inch silicon wafers, sat in the middle of an automotive microcontroller line that produced what the industry blandly called MCUs. Each MCU was a small piece of logic that ran a single function inside a modern car: the antilock brakes, the engine management, the airbag deployment, the windshield-wiper rain sensor, the digital instrument cluster. A car built in 2021 contained somewhere between fourteen hundred and three thousand of them. The Naka line shipped, in any given month, somewhere between five and ten percent of the world’s supply.

The overcurrent ignited the casing. By the time the on-site fire team and the Hitachinaka municipal brigade had the flames out, five and a half hours later, six hundred square meters of cleanroom on the first floor of N3 had burned. Five percent of one floor. The cleanroom remained largely intact, but the air inside, after a fire involving plating chemistry, carried chlorine residues and soot at levels no process tool could survive without deep decontamination. When Renesas’s chief executive, Hidetoshi Shibata, walked the floor with his manufacturing leadership a few hours later, he found that the eleven machines they had initially flagged as damaged were not the whole story. By the time the deep inspections concluded the following week, twenty-three pieces of process equipment, each costing five to twenty million dollars and each with delivery lead times measured in months, would have to be replaced or rebuilt. The fab would not return to its pre-fire shipping volume until summer.

In Wolfsburg, Detroit, Stuttgart, and Toyota City, the planning departments of the world’s automakers spent the next forty-eight hours trying to figure out how much of their second-quarter production they had just lost. The answer was a great deal of it, on top of losses already taken from the Texas winter storm five weeks earlier and from a pandemic-era supply gap that had been worsening since the previous summer. The redundancy of the system, it turned out, had been quietly thinning for a decade.

The fire was the moment the shortage became, briefly, a story the public could see. It was not the moment the shortage began.

The beginning sat in a video meeting Volkswagen’s procurement organization held in late March 2020, when most of the world had been under lockdown for a week or two and the global automobile market had collapsed by close to forty percent. New-vehicle sales in China had cratered in February. Sales in the United States and Europe had cratered in March and were heading into April with no visible floor. Across the industry, the procurement teams whose job was to manage long-tail components made decisions that, in normal times, would have been textbook prudent. They cancelled orders. They notified their tier-one suppliers, the Boschs and Continentals and Aptivs, that January production schedules no longer applied. The tier ones cancelled their orders with the chip foundries that made the silicon those components depended on. The largest of those foundries was Taiwan Semiconductor Manufacturing Company. The second was United Microelectronics Corporation. The third was Samsung’s foundry division. Each had been running, in early 2020, with their twenty-eight-, forty-, sixty-five-, and ninety-nanometer lines fully loaded.

Mature-node capacity, in foundry economics, was not a luxury. It was the bulk of the business. The leading-edge nodes the financial press wrote about, seven and five and three nanometers, ran the application processors for iPhones and the GPUs Nvidia sold to data centers, and earned the highest gross margins. But in wafer volume, the older nodes ran most of the chips that ran most of the world’s products. The power-management ICs lived there. So did the display drivers, the touch controllers, the wireless transceivers, the USB interfaces, and the bulk of automotive microcontrollers. They had been built out in fabs that had long since paid back their original capital and ran on fully depreciated economics. New mature-node capacity had grown at perhaps two percent annually. Leading-edge capacity had grown at twenty-five percent or more, because that was where the money was.

When the auto industry cancelled its orders in April and May, TSMC and UMC and Samsung did not leave the wafer slots empty. They could not. The rent on the buildings, the depreciation on the equipment, and the contracts with suppliers continued whether the wafers shipped or not. They reallocated the slots in real time to the customers who were ordering. Those customers, that summer, were the consumer-electronics brands. The world had moved abruptly into its homes. Children needed laptops for school. Parents needed laptops for video calls. Households that had not bought a new television in a decade replaced theirs. The Nintendo Switch sold out at every retailer. Webcams, microphones, gaming GPUs, and Wi-Fi routers were on backorder by midsummer. Acer, Asus, Lenovo, HP, and Dell each shipped record laptop volumes through the back half of the year. Global PC shipments in 2020 ran north of three hundred million units, the highest figure since 2014. The chips that went into those products came off the same twenty-eight, forty, and sixty-five-nanometer lines the auto industry had abandoned.

By the time the auto industry came back, in the fourth quarter of 2020, the slots were taken. Volkswagen said so first. On December 18, Bloomberg reported the company was bracing for a six-figure production gap in the first quarter of 2021 because semiconductors were no longer available at the volumes the procurement organization had assumed. Continental and Bosch, asked the same week, said the bottlenecks could last six to nine months. The rest of the industry caught up over the following six weeks. Ford idled shifts at its Chicago and Louisville Assembly plants in the last week of January. GM halved production at one of its Korean plants and began extending downtime at three North American factories. Stellantis, Renault, and Honda made similar announcements. North American auto production in Q1 fell by roughly a hundred thousand vehicles below plan; Volkswagen alone said it would build a hundred thousand fewer cars in the period than it had intended.

The cause was easy to describe and structurally hard to fix. A modern car required, on the most ordinary of mature nodes, a long parts list of microcontrollers and analog chips that could not be substituted for one another. An engine-control unit qualified for a Volkswagen MQB platform was not interchangeable with one qualified for a Toyota TNGA platform. Each qualification had taken years and millions of dollars of validation work. A new chip from a different foundry, even a chip with the same nominal specifications, could not be drop-replaced into a vehicle whose entire electrical architecture, safety certification, and crash-test history had been built around the original part. The auto industry could not, when it ran short, simply buy more silicon at the spot price. It had to wait for existing suppliers to rebuild the volume on lines now backed up by eighteen months. The wait would be longer if any one of those lines failed.

Five weeks before the Renesas fire, one of the lines had failed.

On the night of February 14, 2021, a polar vortex broke through the jet stream and pushed across the central United States. By the morning of the fifteenth, temperatures in Austin had dropped below ten degrees Fahrenheit, a level the Texas grid had not been engineered to withstand. Gas-fired generators whose intake equipment had not been winterized began tripping offline. By afternoon, ERCOT had instructed its utilities to shed roughly half of the demand on the system. Austin Energy began rolling its largest customers off. Among them were three semiconductor fabs in the city’s northeast quadrant: Samsung Austin Semiconductor, NXP’s Oak Hill and East Austin facilities, and Infineon’s Austin operation. Together they accounted for roughly ten percent of U.S. wafer production.

Samsung began an orderly shutdown at one o’clock on the afternoon of February 16. NXP and Infineon followed the same day. A modern fab does not absorb a sudden power loss the way an ordinary factory does. The thousands of wafers in process at the moment of the cut were ruined. The deposition reactors and ion implanters cooled out of their operating temperatures and had to be requalified before the next wafers could enter them. The pure-water plant, the chemical-distribution lines, and the abatement systems all required staged restart procedures that took longer to execute than the storm itself. Samsung’s Austin facility would not return to full output until late March. NXP’s Austin lines, which made automotive chips for the same customers Volkswagen and Ford and General Motors had been begging for capacity, were down for similar stretches. The aggregate loss was on the order of seventy thousand wafer-starts.

The Renesas fire, when it came, hit a system that had no slack left. Within hours of the news reaching Tokyo, the Japanese government reached out to Taipei to ask whether TSMC and UMC could shift any capacity, however small, to backfill the lost Naka volume. The answer, polite and immediate, was that they could not. TSMC was already running its automotive microcontroller lines at what its chairman Mark Liu later called a “super-hot run,” compressing the normal cycle time by up to fifty percent so that more product could move through the same equipment. Liu told reporters in May that TSMC had heard about the shortage in December, had begun reallocating in January, and had increased microcontroller output for the first half of 2021 by thirty percent over the prior year, with a goal of sixty percent for the full year. The work was constrained by physics. A process step took the time it took. A new fab, even one already announced, took three to five years to come online. The Naka damage would have to be absorbed largely by Renesas itself, on a recovery schedule the company estimated at a hundred to a hundred and twenty days.

Four days after the fire, the container ship Ever Given, four hundred meters long and outbound from Yantian for Rotterdam, was caught broadside by a sandstorm in the southern reach of the Suez Canal and grounded across both banks. It would take six days to free her. Three hundred and sixty-seven other vessels accumulated at the mouths of the canal. Several were carrying chips from Asian fabs to European auto-assembly lines. The Suez delay was not, by itself, decisive for the shortage; chips travel by air more often than by sea. But the image of the Ever Given, one ship blocking one canal, became the period’s defining illustration of how a global system that had spent thirty years optimizing for cost rather than redundancy now reacted to small disturbances. Goldman Sachs, in a spring 2021 analysis its economists circulated to clients, identified a hundred and sixty-nine downstream industries with measurable exposure to the chip shortage. The list, derived from input-output tables of GDP rather than from headlines, included expected names like automotive and consumer electronics and unexpected ones like ready-mixed concrete, soap manufacturing, and breweries. Almost everything that ran a microcontroller, which by 2021 was almost everything, was on it.

The most consequential downstream industry, by dollar value and political weight, was automobiles. AlixPartners kept revising its forecast of lost revenue upward through 2021. In May, the firm projected a hundred and ten billion dollars in foregone industry sales. By September, it had doubled the figure to two hundred and ten billion and projected the industry would build seven and seven-tenths million fewer vehicles than planned. The forecast did not capture the second-order effects. Used-car prices in the United States rose more than thirty-nine percent from March 2020 through late 2021, according to the Bureau of Labor Statistics, as buyers locked out of empty new-car lots bid up existing inventory. By the autumn of 2021, used-vehicle prices were a meaningful contributor to the headline inflation number that would come to define the U.S. economy in the Biden administration’s first year.

Toyota, alone among the major automakers, kept building cars at near-plan volumes through the first half of 2021, because Toyota had spent the decade since the 2011 Tohoku earthquake quietly stockpiling up to four months of certain critical chips, against the orthodoxy of just-in-time manufacturing the company itself had invented. The strategy held until August, when the Delta variant spread through the Southeast Asian factories that supplied Toyota’s deeper tiers, and the chip stockpile finally drained. On August 19, the company announced a forty percent cut to its September global production plan, against an original schedule of nine hundred thousand vehicles. Bloomberg’s coverage that morning treated the news as a marker. If the chip shortage could ravage Toyota, the company that had treated supply-chain risk as a strategic priority for ten years, no automaker on earth was insulated.

Volkswagen’s Herbert Diess said the company was in “crisis mode.” The procurement team in Wolfsburg had been put on a war footing. Volkswagen began bypassing its tier ones and approaching foundries directly: NXP, Infineon, TSMC. The auto industry had not done that for forty years. At Ford, Jim Farley, who had inherited the chip shortage as one of his first operational problems after taking the chief executive’s office in October 2020, declared the future of the company would be a build-to-order model rather than the dealer-stocked inventory system Detroit had run since the 1950s. By May, his Kentucky Truck Plant in Louisville was producing F-150 and Super Duty pickups without their final electronic modules and parking the unfinished trucks at the Kentucky Speedway in Sparta, an hour up the interstate, where they covered the racetrack’s thirty-acre infield in rows visible from low Earth orbit. By autumn, between sixty and seventy thousand incomplete trucks were sitting on the asphalt, waiting for a single chip per vehicle.

Mary Barra at General Motors decided early on that her finite chip allocation would go disproportionately to the company’s highest-margin products, full-size pickups and large SUVs. Smaller cars and lower-margin nameplates would absorb the production cuts. Spring Hill in Tennessee, Lansing Delta Township in Michigan, and Fairfax in Kansas each took two-week or longer idle periods through April and May. GM’s CFO told investors the shortage would cost the company between one and a half and two billion dollars in 2021 operating profit. Ford, with less aggressive prioritization, lost more. The global production gap by year-end exceeded eleven million vehicles by some accountings.

The political response moved with the speed political responses tend to move. On April 12, 2021, three weeks after the Renesas fire, President Joseph Biden held a virtual summit in the Roosevelt Room with about twenty chief executives from the affected industries: Barra and Farley and Stellantis’s Carlos Tavares, Pat Gelsinger from Intel, Sundar Pichai from Alphabet, and senior leaders from TSMC, Samsung, and the major American chipmakers. Biden held up a four-inch silicon wafer for the cameras. “These chips, these wafers, batteries, broadband,” he said. “It’s all infrastructure.” The argument underneath was simple. A country whose automotive, defense, consumer electronics, and household-appliance industries all depended on chips made in Taiwan, on tools made in California and the Netherlands, with no domestic redundancy worth the name, was a country that could be paralyzed by an overcurrent in a Japanese cleanroom.

It would take another sixteen months for subsidy legislation to clear Congress. Pat Gelsinger, who had announced in January 2022 that Intel would build a twenty-billion-dollar fab complex in Licking County, Ohio, withheld the June groundbreaking ceremony as a public protest at the bill’s drift, telling reporters he would change his investment plans if the Hill did not act. The CHIPS and Science Act passed the Senate on July 27, the House the next day, and was signed on August 9. It appropriated fifty-two and seven-tenths billion dollars for domestic semiconductor manufacturing and research, with thirty-nine billion earmarked for direct fab subsidies and a separate twenty-five percent investment tax credit for equipment. The execution of the program belongs to a later chapter.

The sharper diagnostic lived inside the Department of Commerce’s own data. In a survey published in January 2022, the department reported that the median manufacturer’s chip inventory had fallen from forty days in 2019 to fewer than five days by 2021. Five days. The system built over the previous quarter century to run on a global flow of just-in-time silicon now had less buffer than the next weather event was likely to require. Tom Caulfield, the CEO of GlobalFoundries, told CNBC in October 2021 that “for the better part of the next five to ten years, we’re going to be chasing supply, not demand.” The industry had spent twenty years investing in the leading edge because that was where the margins were and where the financial press paid attention. The mature nodes, where the actual shortage lived, had been allowed to atrophy. There were cars sitting in parking lots, Caulfield said, missing chips made on forty-five or sixty-five-nanometer processes the industry had perfected fifteen years earlier and stopped expanding capacity for.

Renesas’s Naka factory returned to its pre-fire shipping volume by late June 2021, slightly behind Shibata’s hundred-day target but inside the company’s revised hundred-and-twenty-day projection. The shortage continued through 2022 and tapered through 2023. By 2024, automotive lead times had returned to something like the long-running normal. By 2026, with AI demand pulling on a different but parallel set of mature-node capacities, a new shortage was already taking shape on the same vulnerable infrastructure.

The deeper lesson of the 2020-22 episode was not that any one of the events had been catastrophic. The Renesas fire had affected five percent of one floor of one building. The Texas storm had taken three fabs offline for a few weeks. The Ever Given had blocked a canal for less than a week. None of those events, on its own, would have been visible in the global trade statistics. The lesson was that they did not have to be. The system had been engineered for efficiency at the expense of margin, in a sequence of rational corporate decisions no individual chief executive would have made differently in the moment, and the result was a network in which one node hiccuping anywhere in the world could move trillions of dollars of downstream activity within weeks. The visible failure mode was a parking lot in Sparta, Kentucky, full of unfinished pickup trucks waiting for a single absent chip. The deeper failure mode was the absence of any other way the system could have responded.