Intel wants to be a foundry—again. Will this time be different?10/20/2021
Last month, Intel CEO Pat Gelsinger stepped to a podium on a hazy, wind-whipped day just outside Phoenix. “Isn’t this awesome!” Gelsinger exclaimed, gesturing over his shoulder. Behind him, two large pieces of construction equipment posed theatrically atop the ocher Arizona soil, framing an organized tangle of pipes, steel, and fencing at the company’s Ocotillo campus. “If this doesn’t get you excited, check your pulse,” he said with a chuckle. A handful of executives and government officials applauded at the appropriate points.
Despite the gathering dust storm, Gelsinger genuinely seemed to enjoy himself. He was in Arizona to announce not one but two new fabs that, when finished, will form a $20 billion bet that Intel can return to the leading edge of semiconductor manufacturing, one of the world's most profitable, challenging, and cutthroat businesses.
“Semiconductors are a hot topic these days,” Gelsinger continued. “What aspect of your life is not being increasingly driven by digital transformation? If there was any question on that, COVID eliminated it.”
Over the last year and a half, as the pandemic has everyone turned to their screens, demand has surged for devices (phones and laptops) and cloud services (Netflix and Zoom), all powered by a range of advanced semiconductors. Manufacturers have raced to squeeze more chips out of their fabs, but many were running near their limits before the pandemic. Still, Intel and its competitors didn’t rush to build new fabs—fabs are startlingly expensive, and without continued demand, semiconductor firms are loath to build more.
But now, as the global pandemic continues to disrupt supply chains, chipmakers have decided that the current spike in demand isn’t going away. Intel’s $20 billion investment is only one example. Samsung announced in May that it would spend $151 billion over the next decade to boost its semiconductor capacity. TSMC made a similar announcement in April, pledging to invest $100 billion in the next three years alone.
The investments required to stay at the leading edge—where the most advanced chips are made—has whittled down the number of semiconductor competitors from more than 20 in 2001 to just two today. “There’s really only so much room at the leading edge, just because of the huge capital costs involved,” said Will Hunt, a research analyst at Georgetown University’s Center for Security and Emerging Technology.
That cost is driven by the price of the equipment that’s required to etch ever-smaller features onto chips. A few years ago, the industry began to use extreme ultraviolet lithography (EUV) to shrink transistor sizes. EUV machines are marvels of physics and engineering, and one tool costs upwards of $120 million. To stay relevant, companies will need to buy a dozen or more annually for the next several years.
For those sorts of investments to make sense, semiconductor manufacturers must produce and sell an enormous volume of chips. “When you have volume orders, then you can do yield experiments, you can improve your yield, and yield is everything because that’s how you cover your costs,” said Willy Shih, a professor of management at Harvard Business School.
Which is why Intel, under Gelsinger, is doing something now that it historically has shunned. “We are now a foundry,” Gelsinger said at the Arizona groundbreaking. In the coming years, he said, Intel will “open the doors of our fab wide for the community at large to serve the foundry needs of our customers—many of them US companies that are dependent on solely having foreign supply sources today.”
But becoming a leading-edge foundry isn’t just about building fabs and telling customers you’ve got space to make their chips. Gelsinger will have to change Intel’s culture and, to some extent, its technology, both of which are deep-seated. “He has to turn a huge ship around,” said Robert Maire, president of Semiconductor Advisors.
In the coming years, Intel has several challenges to master at once. As the company rolls out a new business model, it also needs to redouble its R&D efforts while still being careful with cash flow. (Intel has fallen so far behind that it now plans to outsource production of its most advanced chips—and a portion of the profits that accompany them—to TSMC.) The transition will demand intense focus. “The foundry business could be a distraction,” Shih said.
At the same time, he added, Apple, Google, Amazon, and other companies are moving away from Intel’s standardized chips toward their own customized designs. If Intel doesn’t change with the times, it risks being left behind. “There will be many challenges, and there will be tests that will face them,” Shih said. “It’s going to be hard.”
Trial and failureIt was a brisk fall morning in Burlington, Vermont, when Governor Peter Shumlin addressed a crowded room in a two-story brick building just blocks from Lake Champlain. Clouds were rolling in from the south, reflecting the dark mood that had settled over the Green Mountain State after news had leaked that IBM was handing its chip business over to GlobalFoundries. IBM’s fab in Essex Junction, six miles from where Shumlin stood, employed around 4,000 people at the time, making Big Blue the largest private manufacturer in the state.
The night before, Shumlin had talked with executives from IBM and GlobalFoundries, and they assured him they had no plans to cut staff. “This is good news for Vermont,” Shumlin said, “because GlobalFoundries is in the chip manufacturing and development business. That’s what they do.”
Still, Vermonters had every right to be worried. “We all know that in recent years, there has been a lot of uncertainty about the future of IBM in this particular business,” Shumlin acknowledged. At one time, the plant had employed more than 8,000 people, but by 2014, when the sale was announced, IBM’s Microelectronics division had shrunk significantly after rounds of layoffs. The unit also had been a drag on IBM’s profits, losing $700 million in 2013 and another $400 million in the first half of 2014.
It was an ignominious end for one of the industry’s most storied names. IBM was among the first companies to license the patent for the transistor from Bell Labs, and for decades it made the chips that distinguished its mainframes and servers from the competition. The company's chip-making technologies were—and still are—lauded for their ingenuity.
For much of their history, IBM and other US semiconductor firms designed the chips they built, making them so-called integrated device manufacturers or IDMs. Over the years, they realized they could sell their fabs’ excess capacity to keep the pricey machinery busy, but it remained a relatively small part of their business.
This frustrated Carver Mead, then a professor at Caltech. In the 1970s, Mead was teaching his students how to design and make chips, and he was growing weary of managing the logistics required to turn a design into etched silicon. He was able to sneak his students’ work into test runs at Intel because he knew founders Bob Noyce and Gordon Moore and had taught many of the engineers who worked in the fabs. But he also knew that option wasn’t available to everyone.
Mead wanted anyone to be able to design chips and let someone else handle the rest, and he thought there might be a viable business in the idea. “I started with Intel,” Mead said, “and then I would go to CEOs of the various companies and explain to them why that could be good business.” All of them turned him down, saying they wanted to own the designs. “I basically ran out of prospects," he added.
His fortunes took a turn, though, when he was invited to Taiwan to speak to the government’s Industrial Technology Research Institute. “They got very excited about it,” Mead recalled. This was in the early 1980s, and Taiwan was aiming to modernize its economy. ITRI’s mandate was to bolster the manufacturing sector. Taiwan didn’t have much expertise in designing, marketing, or selling advanced chips, but it was getting good at making them. The Taiwanese government latched onto Mead’s vision and brought in Morris Chang, a former Texas Instruments executive, to start a foundry dedicated to making other companies’ chips.
TSMC was founded in 1987 with funding from the Taiwanese government and from Philips, the Dutch electronics firm. Over the years, it attracted larger customers; even Intel began making some of its lesser-known chips with TSMC.
Other companies took notice. IBM started selling its excess capacity in earnest in the early 1990s, and by 1998, it was making Intel-compatible x86 chips for AMD and Cyrix. Part of its pitch to potential customers, many of which were in California, was that working with IBM could diversify their supply chains. “What is in the customer’s interests?” an IBM executive said at the time. “Any easy ability to second source parts. Some other fabs want to be the only source.”
In 2002, IBM had another reason to sell its excess capacity—its $3 billion cutting-edge fab in East Fishkill, New York. No other project in company history had cost as much, and Big Blue, needing to recoup some of its investment, redoubled its efforts to win foundry customers. By the next year, the company was the third-largest foundry by revenue—still well behind TSMC but with a roster of blue-chip clients that included Sony, Qualcomm, Nvidia, and AMD. Its fab technologies were the envy of the industry. “IBM is the absolute best,” an industry analyst said at the time. “You pay through the nose for it, but it’s great stuff.”
Not all potential customers were convinced that access to the best technology was worth the risk, though. Some worried that IBM would push its foundry customers to the end of the line once demand for its high-margin servers picked up. Others were concerned that IBM’s foundry kept its best technologies for itself, only sharing them after their edge had dulled. The ones it did share didn’t cater to a wide swath of customers, and the company showed little interest in accommodating them. IBM’s chip technologies were “defined by the requests from their server group,” said Harry Levinson, principal lithographer at HJL Lithography, who spent decades at AMD and GlobalFoundries. “They basically crafted everything towards one customer, and that was their own internal customer.”
One famous example was Apple, which struggled with IBM’s hand-me-down technology. In the early 2000s, IBM designed and supplied Apple with the PowerPC G5, a derivative of the POWER4 server processor. It worked well in Apple’s spacious Power Mac towers, but Apple struggled to put the hot, inefficient chip in a laptop. At a time when more customers were buying laptops, Apple’s PowerBooks began slipping further behind Windows PCs.
Ultimately, Apple switched to Intel’s energy-efficient Core processors, and IBM didn’t appear heartbroken about the decision. Apple was a famously demanding customer, and PC chips had lower margins than their server cousins. “IBM knew who the customer was—it was their server business,” Levinson said. “It just didn’t admit that was incompatible with their foundry ambitions.”
IBM’s chip division was tossed a lifeline when Sony, Nintendo, and Microsoft all adopted the PowerPC architecture for their game consoles in the mid-2000s, but it wasn’t enough. “IBM did not focus on foundry as a profit center,” Maire said. As a result, the company never invested the money required to make the business succeed, and soon its semiconductor business was on the chopping block. IBM ultimately paid GlobalFoundries $1.5 billion to take its fabs off its hands.
Small stepsMeanwhile, TSMC was rising. Throughout the 2000s, revenues flowed, enabling them to invest in bigger, more advanced fabs. The company was proving that the foundry business wasn’t just viable but wildly profitable.
Soon TSMC was courting Apple, a move that would eventually catapult the Taiwanese company into the lead. Apple had started designing its own chips for the iPhone, and while early versions were made by Samsung, a patent dispute led Apple to search for alternatives.
Apple tested TSMC’s chips for several years before pitting them against Samsung’s in the iPhone 6s in 2015. Samsung nominally had the edge—its transistors were smaller and should have been more efficient than TSMC's. But for the most part, that wasn’t true. TSMC’s chips held their own, and when the iPhone 7 rolled around, Samsung was out. Analysts lauded the phone's “incredibly thin” and powerful processor, and TSMC has been in the iPhone ever since.
TSMC had mastered the art of low-power processors in part because it had refined its techniques across hundreds of diverse customers. “When you have a large number of foundry customers, these customers don’t all have their product cycles synchronized,” said Chenming Hu, a professor at the University of California, Berkeley, and former chief technology officer at TSMC. “Almost any time you have a new technology, you will have some customer that’s willing to pay for it.”
Foundries that can juggle multiple clients and technologies can swiftly advance on their competitors. For example, TSMC’s scale allowed it to master extreme ultraviolet lithography faster than anyone else, which reduced the number of steps it took to make advanced chips and boosted the throughput of its fabs.
Large foundries have dozens or hundreds of customers, Hu said, which encourages them to take smaller steps because there’s always an interested customer. “When you take bigger steps, there’s more chance of slipping than taking a larger number of smaller steps," he said. It’s also easier for companies to recover from stumbles.
Intel has historically taken big leaps that attempt to mirror Moore’s law, which describes a doubling of transistor densities every 18–24 months. For most of the company’s history, Intel succeeded, rolling out impressive updates that kept the company one or more steps ahead of its competitors. But then in 2015, it slipped. Intel announced that chips made on its 10 nm node would be delayed. In 2017, it announced another delay. Soon the industry titan wasn’t just even with its competition, it was behind.
At the same time, other companies began to follow Apple’s lead, designing their own chips rather than buying off-the-shelf parts like the ones Intel sold. As TSMC pulled ahead, more companies sent their designs to it, which gave the Taiwanese company more opportunities to refine its processes. Today, around 90 percent of leading-edge chips are manufactured by TSMC, and the rest are made by Samsung.
“Developing a new-generation technology is tremendously difficult,” Hu said. “Intel falling behind TSMC and Samsung in the very leading edge technology can be traced to the fact that Intel did not participate in the foundry.”
Trusting the foundry“The real question is, does Intel moving into the foundry business help them get back to cutting edge?” said Paul Triolo, practice head of geotechnology at the Eurasia Group.
Success in the semiconductor industry is heavily reliant on scale. More sales mean more opportunities to perfect your process, and perfecting your process helps get you to the next node. “That’s what happened with TSMC. At some point, they reached a critical mass of capability and customer relationships. And that was this virtuous cycle—that’s what you need,” Triolo said.
“Can the US maintain its technology lead without getting into foundry? That really is the question that Gelsinger grappled with,” Hu said. “I certainly agree with this conclusion, that no, Intel cannot maintain a leading technology without getting into the foundry business.”
For Intel and the US, embracing the foundry model represents a significant shift. Historically, leading US firms have either functioned as IDMs that design and make chips or as fabless designers that outsource the production to another company, usually in Asia. Part of that is because much of the profit in computer chips comes from designing and selling them, not making them. TSMC’s success as a pure-play foundry “is almost historically anomalous,” said Jennifer Kuan, deputy director of innovation and research at California State University—Monterey Bay. Not only that, but “TSMC has shown that it’s a profitable business,” she added. Many people hadn’t thought it could be.
Now, Intel appears to be going after the foundry business with gusto. So far, Intel has announced Qualcomm and Amazon Web Services as customers, and Klaus Schuegraf, vice president of foundry strategy and planning, said that 100 more companies have expressed interest. “That’s a long order book, and they come from all walks of life.” For now, Intel appears focused on the high-performance market. “We see the growth in the business over the next five, ten years is predominantly from the leading edge,” Schuegraf said.
The true challenge of the foundry market, though, isn’t wooing customers or even developing better technologies—it’s making sure each customer’s needs are served. One thing that sets TSMC apart is its “ability to meet a wide, wide range of customer needs,” particularly those of chip design companies, said Hunt, the Georgetown researcher. “That’s something that Intel does not have that depth of experience with.”
And as a pure-play foundry, TSMC customers don’t have to worry that unique elements of their designs will somehow leak over to the manufacturer’s own chips. “When you talk to TSMC, one of the basic tenets is customer trust,” said H.-S. Philip Wong, a professor at Stanford who led TSMC’s research and development from 2018–2020. “This is as important to them as manufacturing capability.”
Intel appears to have learned from its failed previous attempt at launching a foundry, which sputtered out after five years in 2018. Unlike the last time, the new Intel Foundry Services is a standalone business unit that reports directly to Gelsinger. Schuegraf said the company has been developing key relationships with suppliers over the years so that fabless companies won’t have to shoehorn their designs into Intel’s way of doing things. And their process is set up so customers don’t have to worry about their secret sauce popping up in other chips.
“We’re in a position to segregate customer confidential information, compartmentalize that information, and service those customers while we’re leveraging the core of Intel,” Schuegraf said. Any customization of Intel’s process for a particular customer “is privy only to those who need to know.” The company is also committed to being “fair and transparent in how we allocate capacity,” he added. “It’ll look a lot like dealing with other foundries.”
An acquisition might help Intel quickly gain some of those skills. Earlier this year, Intel was rumored to be in talks with GlobalFoundries, but nothing has materialized so far. “The main failing of Intel’s previous foundry attempt was not being customer centric or focused on building this plethora of parts for a huge number of diverse customers,” Maire said. “GlobalFoundries has done that.”
Schuegraf acknowledges that accommodating a wide range of customer requirements could test the company, which historically relied on a tight relationship between its in-house designers and its manufacturing arm. But he thinks it’s a challenge Intel can overcome. “It will stretch us a bit. But fundamentally, we’re geared to do that.”
America at a crossroadsIt will be a while before anyone can judge Intel’s foundry ambitions as a success or failure. Observers think it will be at least three years, and more likely five, before that can happen. The fabs will take a couple of years to build, and new chip designs will take months or years to test and produce. “These things always take some amount of time,” said Shih, the Harvard professor.
In the long term, Intel’s success is also dependent on US industrial policy. TSMC enjoys lower costs than Intel in part because it receives significant support from the Taiwanese government. The Semiconductor Industry Association of America estimates that building and operating a fab in the US for a decade costs 30 percent more than in Asia, and about half that cost difference comes down to government subsidies.
Congress is mulling an injection of around $50 billion into the semiconductor industry that would incentivize research and development and the construction of domestic fabs. That would go some way to leveling the playing field, but it would also pose new challenges. If the bill passes, tens of thousands of new jobs would be created in the semiconductor industry every year, said Tsu-Jae King Liu, dean of the College of Engineering at the University of California, Berkeley. (Liu is also a board member at Intel, though she spoke with Ars only in her capacity as dean.) “This means that we need between 5,000–10,000 new graduates per year. No single university—or even a university system like the University of California—can meet that workforce development need,” she said.
In the short term, high-skilled immigration will help fill the gap, Hunt said. But longer-term, the US needs to find ways “to increase the amount of talent that we have coming up,” he said. “It’s really everything, from people with a vocational degree up through people with PhDs.” Along those lines, Liu and her colleagues have been developing a program that would span industry, universities, and community colleges to fill the need. “We want to take best practices in semiconductor education and propagate them across the country,” she said.
The US and Intel have a steep hill to climb to regain the leading edge in semiconductor manufacturing, but no one Ars spoke with was ready to discount their chances. “I think there’s an argument to be made that having a ‘whole of company’ focus on manufacturing has improved TSMC’s odds of out-competing Intel. But I don’t think there’s anything inevitable about what’s happened,” Hunt said. “The United States has talent from around the world, we have great IP protection, and we have an ecosystem of design firms that are eager to work with a foundry in the United States.”
Hu, the Berkeley professor, agreed. “We have a leader, namely Intel, and we have the technology base, which is basically on par with the best in the world," he said. "And we still have the best universities in this area.” What’s been missing, perhaps, is a sense of urgency. Whether or not you can achieve success, Hu added, “really depends on whether you think you have to do it.”
For now, plenty of people think the US—and Intel specifically—have little choice. “We’ve seen that a relatively small shortage in semiconductors got exacerbated by COVID and has thrown the global economy into a tailspin,” Maire said. “The success of the United States, to a very large extent, is highly dependent upon this.”
https://arstechnica.com/tech-policy/2021/10/intel-slipped-and-its-future-now-depends-on-making-everyone-elses-chips/ ....
This is an interesting piece published on the future of the global semiconductor industry.
This news would appear to be positive at first glance. With semiconductor production ramping up, shortages in the automotive industry should be filled. Greater scale production of ASICs and GPUs could also reduce their market price, further increasing the viability of crypto mining.
It may not all be good news. There could be negatives present. Intel, samsung and TSMC investing $20 billion to $100 billion into their foundries could prohibitively raise the barrier to market entry to levels unapproachable for smaller players. Governments are the only demographic who could afford to pay $20 to $100 billion into a start up.
Anyways this should be good news for the future of crypto mining.
How does everyone see these massive investments into future semiconductor production panning out?
