(FORTUNE Magazine) – After you're garbed head to toe in your bunny suit, the unisex uniform of semiconductor manufacturing, your first thought as you pass through the air lock and walk out on the production floor of Fab 6 is simply "Jeez! It's huge!" Manned round the clock by a work force of some 2,600 technicians and engineers, it is the largest electronic-chip fabrication plant in the world. The clean room, fittingly named the "ballroom," where integrated circuits are meticulously laid down on wafers of ultrapure silicon, stretches over 190,000 square feet, a rectangle bigger than three adjoining U.S. football fields. Opened in the Tainan area of southern Taiwan in spring 2000, Fab 6 will shortly reach a monthly capacity two-thirds greater than that of most chip plants--50,000 wafers containing at least ten million integrated circuits.

Erected at a cost of more than $2 billion, Fab 6 is the newest plant of Taiwan Semiconductor Manufacturing Co., or TSMC. It's about a three-hour drive south of Hsinchu, the company's headquarters and the site of eight of its ten other fabs as well as most of the fabs of its main competitor, United Microelectronics Corp., or UMC. Both companies are merchant mills, better known in the trade as foundries. They make wafers for companies that supply semiconductor designs, and then ship them to other companies that slice them up and encapsulate the individual chips. This is no cottage industry. TSMC's revenues hit $5 billion last year, UMC's were $3 billion, and until last winter the trend was way, way up.

About 85% of the world's integrated circuits are designed and produced by such better-known companies as IBM, Intel, Motorola, and Texas Instruments in their own fabs for sale under their own name. These so-called integrated-device manufacturers also do a little foundry work for others on the side. But TSMC and UMC, with a customer list that includes some of the most innovative U.S. startups, dominate the foundry business with a market share approaching 70%. Foundries are the fastest-growing part of the semiconductor industry. Applied Materials, the leading manufacturer of chipmaking equipment, forecasts that by 2004 about one-fourth of global chip production will be outsourced to them. By 2010, according to another projection, foundries will have half the business. It's a fair bet that TSMC and UMC will have more than half of the half.

Customers of the two companies, most of them in the U.S., might well be concerned about their dependence on a small Asian island, a literally shaky one at that. In September 1999, all of Taiwan's chip plants were jostled by a 7.3-magnitude earthquake. Mainly by cutting off electric power, it knocked out a goodly chunk of one month's production. The siting of Fab 6 in southern Taiwan, a less earthquake-prone area where UMC is also opening a plant and where both companies have plans for more, was intended to alleviate the problem. But there's little the companies can do about another worry: Taiwan's tense relationship with China, a missile shot away, whose testiness was on display during its recent confrontation over the U.S. spy plane forced to land on Hainan.

Geological and political concerns about Taiwan, however, probably won't dim the bright future of TSMC and UMC. They not only are No. 1 and No. 2 in the foundry business, but are also ahead of competitors in gearing up to make chips on 300mm wafers, disks of silicon about the size of a medium pizza, or 12 inches. These promise far greater production efficiencies than the current standard of 200mm, or 8 inches. Five years ago the shift to a 300mm wafer seemed imminent, but it didn't turn out that way. Chilled by a semiconductor slowdown in 1996-98 and again last winter, as well as by fears that the required manufacturing equipment might not work, almost all U.S., European, and Japanese chipmakers have hesitated to pour money into 300mm technology. The new fabs it requires will each cost $3 billion and counting.

TSMC and UMC, however, are forging ahead with the big wafers, undeterred by a steep downturn in orders. Last fall the two were operating at more than 100% of rated capacity; today they are in the 70% range or lower. (Other chipmakers are less forthcoming, but few can be doing much better.) However, while the Taiwanese have taken whacks at their capital-expenditure budgets, they have spared the 300mm effort. Pilot production of the big wafers is already under way at TSMC's Fab 6, and both companies have erected full-scale factories that will begin production in a few months. The first will be a new, $3 billion fab in Japan that UMC has built in a joint venture with Hitachi. More fabs will follow, in Taiwan and elsewhere. Favorable tax treatment by the Taiwan government helps make the huge investment gamble palatable. Whatever the reason, a technological lead is opening up that in the future could put TSMC and UMC in a strong position to win even more market share.

Big chunks of the semiconductor business are effectively off limits to the foundries. They don't make many standard DRAM or memory chips, which are produced in immense quantities by specialists such as Micron in the U.S. and Samsung in South Korea. They don't make Pentium chips either. While proud Intel reaches out to Taiwan for some products, it keeps production of its world-dominating microprocessors in-house. That still leaves plenty of territory for the foundries to conquer. Their typical customer is a designer and marketer of integrated circuits that resells its finished chips to somebody else, typically an equipment maker that installs them in its products.

Foundries will build any kind of semiconductor the customer desires, including analog chips and all digital versions, from microprocessors and other logic circuits to nonstandard and embedded-memory types. The chips wind up in everything from computers to cell phones--last year's big source of growth--as well as other telecommunications equipment. They are also found in consumer electronics devices, automobiles, medical equipment, and on and on.

More than seven out of ten foundry customers are "fabless"; they don't build chips at all. The Fabless Semiconductor Association says there are 625 such companies, including 450 in North America. Among them are many of the innovative firms that fueled investors' now chilled exuberance, such as Broadcom, Conexant, PMC-Sierra, and Xilinx. In 2000, combined revenues of the public U.S. fabless companies reached $20 billion, up 68% from 1999, about double the growth rate of the semiconductor industry as a whole.

Fabless companies can get some of their semiconductors made by integrated-device manufacturers such as IBM and Toshiba, which mainly produce proprietary chips of their own design. But when turning to them as a source, the fabless are at a disadvantage: They are using a potential or actual competitor as a supplier. Foundries like TSMC and UMC, on the other hand, are "pure plays" in the sense that the semiconductor industry uses the phrase. With no circuit designs of their own, they dedicate themselves solely to making what the fabless customer dreams up. (Publicly traded, the Taiwanese chipmakers are also pure plays for investors in the Wall Street sense.) Although UMC sniffs that it got the idea first, TSMC's founder and CEO, Morris Chang, gets credit for it.

A onetime Texas Instruments executive in the U.S., Chang returned to Taiwan and got TSMC started in 1987 as a spinoff from a government research project. Along with $52 million in capital from the Taiwan government, he received technical assistance and capital from Philips, the Netherlands electronics giant, which still owns 22.5% of the common shares. Often seen with pipe in hand, white-haired, 69-year-old Chang now plays the role of elder statesman and resident philosopher. He has penned his autobiography without the usual ghost and is carefully overseeing its translation into English from Mandarin. Throughout the company people are wont to start statements with "As Morris says..."

When Chang started TSMC, there was no reason to expect that chip production would be more successful in Taiwan than elsewhere. The country had an electronic-equipment business that was starting to boom, but even now it accounts for only a small piece of TSMC's and UMC's sales. The principal explanation for their prosperity is simply that there was a supply of the necessities--electricity, water, and an educational system turning out legions of well-trained young engineers with a yen for manufacturing. Just as important, there was a government eager to create better jobs for its citizens by subsidizing an industrial park and enacting tax incentives that are still lavish. TSMC paid no tax on last year's profits, a fat $2 billion, which in fact were boosted by $39 million in tax benefits.

TSMC and UMC have always provided extra, or what Morris Chang calls "bumper," production capacity for the integrated-device manufacturers. But the companies' target customer has been the innovator with more ideas than money. In near-perfect symbiosis, the foundries and the fabless have fed on each other's success. In the past decade 30 fabless U.S. companies have gone from little or nothing to more than $500 million in sales. Among them: Broadcom, based in Irvine, Calif., which designs and markets integrated circuits for TV cable boxes, home and business networking, and other equipment. Last year it doubled its sales and shipped $1.1 billion in chips made for it by TSMC, UMC, and a Singapore foundry.

The dedicated foundry model has the same kinds of efficiencies that have caused electronic-equipment makers to farm out production to contract manufacturers such as Solectron. When one foundry customer is faltering, another is prospering, so the plants stay full, the equipment utilized. There's less overhead--no acres of design engineers, no ranks of marketing and advertising specialists. The focus is on efficient manufacturing.

A couple of years ago TSMC and UMC lagged a generation or more behind Intel, IBM, Texas Instruments, and others in chipmaking technology. But because the Taiwanese companies have grown as dramatically as their best customers, they have generated the profits needed to buy, modify, and invent their way through successive technology generations. They now have caught up with and claim that they have a lead in getting the latest advances into full production. Last year, in a reversal of the usual transpacific technology flow, U.S. chipmaker National Semiconductor licensed technology from, not to, TSMC.

By closing the technology gap, the foundries have given the fabless an extra edge in competing with the integrated-device manufacturers. They need not worry about the cost and headaches of building a fab that will fall behind the technology curve sooner rather than later. But the foundries' relationship with the device manufacturers has also changed. During the 1996-98 downturn, additions to U.S. companies' capacity were delayed just as they have been this year. When the market recovered, the device makers didn't have enough capacity. The Taiwanese did. Now, says Hans Wildenberg, director of the order-fulfillment organization at Motorola, which has close ties to TSMC, "the relationship has intensified a lot." The foundries are no longer just a place to grab a little extra production but have become, as Wildenberg adds, "more and more a source that gets involved in the early stage of a new-product introduction."

TSMC, with 14,700 employees, towers over the business it created. Comparing last year's $5 billion in sales to those of Intel, the world's biggest semiconductor company, with revenues of $33.7 billion, is like comparing apples to apple pie. Unlike Intel, TSMC does not assemble or package the chips it makes; assembly and packaging, plus markups, account for a major chunk of Intel's revenues. So for a true comparison, figures David Wang, executive vice president of Applied Materials, foundry revenues must be multiplied by a factor of 2.5 or three. Do that, and TSMC is second only to Intel as the world's largest semiconductor producer.

While it is now Avis to TSMC's Hertz, UMC is the older firm. It was founded in 1980 as an integrated-device maker but later began to provide foundry services. In the mid-1990s, following TSMC's lead, it sloughed off proprietary products and chip design and put up new fabs, mostly in joint ventures with fabless companies, whose interest in the plants it later bought out. UMC is more inclined than TSMC to get into joint ventures such as its 300mm deal with Hitachi, and to enter technology-sharing agreements with companies like IBM. Its CEO says TSMC is "closed-minded." UMC now has 9,400 employees and eight operating fabs, including one in Japan that it acquired from Nippon Steel. Last year it reported $1.5 billion in net profits.

The Taiwanese dragons are not the only pure-play foundries. There are another six, with a seventh planned in eastern Germany, which will make specialized chips with Intel's help. Four of the companies are tiny and technologically deprived. A fifth, in Korea, is largely a supplier to Texas Instruments. The only significant competitor is Chartered Semiconductor Manufacturing in Singapore, with five fabs in that city-state, including two joint ventures. One is with Agere (just sold by Lucent) and the other with Agilent (an HP spinoff). Chartered made money in 2000 for the first time in four years, $245 million, on revenues of $1.1 billion. It now has brighter prospects, technology deals with Agere and Motorola, and a sixth, 300mm fab rising.

One industry observer views TSMC, UMC, and Chartered as a chip foundry triopoly. But two brash Malaysian upstarts are getting ready to nip a piece out of that dominance. One is 1st Silicon, which is turning out integrated circuits from a site hard by a branch of the Sarawak River on the western side of Borneo. Right. Borneo. An odd spot for a chip plant, arguably the world's most sophisticated, touchy-to-run, and expensive type of manufacturing. Claudio Loddo, the company's expansive, 51-year-old, Italian-born CEO, will tell you 1st Silicon's neighboring town, Kuching, is just an hour or so flying time east from urbane Singapore. But leave the company's parking lot, drive into the nearby rain forest, and you will shortly arrive at a traditional long house occupied by people whose families within living memory considered the ultimate in home decor to be the preserved heads of their enemies.

The other new venture, Silterra, is in a slightly more developed part of Malaysia, an industrial park 45 minutes east of Penang amid low rolling hills covered with palm-oil plantations and groves of rubber trees. For Malaysia, which has a lot of chip-packaging and other electronic businesses, Silterra and 1st Silicon represent an effort to move up the technology chain. Silterra's initial funding includes some $341 million in loans and equity from Malaysia's federal government. 1st Silicon has been staked primarily with $179 million from the Malaysian state of Sarawak, which is well heeled with petroleum revenues.

Both Malaysian fabs are building circuits for foreign partners that have supplied their manufacturing knowhow--Japan's Sharp for 1st Silicon and, for Silterra, America's LSI Logic, which has invested $45 million in the Malaysian company. Each site has room for several more fabs, but for the time being the companies are producing 200mm wafers that are a generation or so behind the latest circuit technology. Catching up will have to wait. Both have equipment loans from foreign banks, and as they ramp up toward their goal of 30,000 wafers a month each, they will need more customers and more capital just to cover their initial budgets of at least $1.5 billion. Right now they are looking for a parlay: a few riskophiles willing to take a piece of the action in return for dibs on output. That's not entirely wishful thinking. Japan's Seiko-Epson, for one, said in April that it will invest $50 million in Silterra in return for a share of production.

When the world's economies improve, the foundry business is going to grow faster than the Taiwan-Singapore triopoly can add production. Moreover, not everybody needs or is willing to pay for the latest 300mm wafers or other technical advances. The Malaysians, if nothing else, will be the newest of the old. Still, filling their fabs with orders won't be easy right now, when even TSMC and UMC are fighting for customers with some not-too-gentlemanly price cutting.

It's a battle between close neighbors. The headquarters of both companies and most of their fabs are about an hour's drive south and a bit east of Taipei, in an industrial park that wouldn't look out of place in San Diego County. It adjoins Hsinchu (pronounced "Sin Ju"), a dreary, traffic-choked industrial city that couldn't be anywhere but Taiwan. In 2000 the companies had little time for sniping. Today the world's once-booming fabless chipmakers are topping each other's layoff announcements. And independent device manufacturers have gotten mealy-mouthed about their commitments to the foundries. Wildenberg of Motorola says that while his company once projected that it would outsource 50% of chip production, it was including chip assembly as well as fabrication in the figure. Besides, he says, 50% was "a statement of intent...the only thing that has changed is whether it is 50% or 40% or 30%...."

TSMC and UMC have done what they can to prepare their Hsinchu fabs for the next earthquake. There's not much they can do about China. The U.S. depends on Taiwan for a lot more than integrated circuits: More than half the laptops, computer monitors, and PC motherboards sold in the world are made on the island. Some Taiwan makers of computer hardware have moved their operations off the island, mostly to China. But after the spy plane incident in April, that may not be especially reassuring to their U.S. and European customers.

The two foundries are building some production capacity outside Taiwan. In addition to the fab it bought in Japan and its joint venture with Hitachi, UMC is building a fab in Singapore with Infineon, a Siemens spinoff. TSMC is also going to Singapore, where this month it is opening a fab built in partnership with its old friend Philips. It has also bought out the other owners of WaferTech, a troubled fab in Camus, Wash., and shipped in a gaggle of engineers to try to improve matters.

Nevertheless, the two companies will keep most of their capacity in Taiwan. Fabs are most economical when they are gathered in clusters like Hsinchu where they can share engineering and services and attract circles of suppliers. TSMC's top management has also been leery of running plants at a distance, especially in North America and Europe, where, they claim, workers have a poor attitude. Drawing on his experience at Texas Instruments, Morris Chang claims that U.S. technicians are not at all like those in Taiwan, who are "very cost-conscious, very company-oriented, very collaborative." It helps, of course, that thanks to a lightly taxed profit-sharing scheme, the Taiwan engineers earn far more in total compensation than their counterparts in the U.S.

Some of the unwillingness to move away from home reflects the anticipation of tough times mastering the shift to 300mm wafers and the associated exotic technology. Better to do this with your engineering and scientific talent as close to headquarters as you can get it. The move to bigger wafers is far more complicated than it is when Denny's puts bigger pancakes on the senior-citizen breakfast menu. First, plant conversion is impractical, if only because the new machinery needs higher ceilings. Next, the technical hurdles are pole-vault high. But conversion will be worth the effort. The potential economic advantages are even greater than those implied by the simple math that reminds us that this 50% increase in the diameter of a wafer results in an area 2.25 times as large. In practice, the number of chips per disk could actually increase by a factor of as much as 2.6.

While the wafer's getting bigger, what goes on it is getting smaller. For years "feature size," the width of the semiconductors and their interconnections on a chip, has been gradually reduced, which lets designers pack more transistors into the same area. For most of the advanced chips in production, feature size is 0.25 micron, or one-fourth of a micron, which is a millionth of a meter. However, the leading edge in chipmaking is 0.18 micron and heading toward 0.15, which is about as thin as any fab now at 0.25 can handle by modifying its machines.

Smaller feature sizes are increasing the viability of the "system on a chip," a single chip that contains several components. Combining them on one sliver of silicon reduces cost and usually improves performance and reliability. A system on a chip might include a microprocessor, another logic circuit, and both conventional and flash memory. The new 300mm fabs will initially make most circuits with a feature size of 0.13 micron and eventually will be able to push down to 0.10 and possibly 0.07. There's no technical barrier to working at this size on smaller wafers, but the economies inherent in the big wafer help offset the costs of working in such small dimensions.

Mastering 300mm production is not going to be a pushover. TSMC is the sort of company that posts a changing series of aphorisms in English and Chinese in easy-to-read locations. A recent one is from U.S. mystery story writer Donald E. Westlake, whose observation should be writ large near Fab 6's 300mm pilot line: "Whenever things sound easy, it turns out there's one part you didn't hear." One thing you have yet to hear is the cost of blank 300mm wafers. A blank 200mm silicon wafer is priced at around $100; the 300mm version is about ten times as expensive because it's tougher to make. With volume, the price will decline, but even three years from now, the bigger substrates will cost three to four times as much.

On the positive side, machinery and equipment prices don't go up in step with wafer size. UMC says the cost won't be much more than 30% higher. TSMC says it's more than that but not double. For several years none of the 300mm machines will be running as fast as those in current plants. But even when 300mm production is just getting under way, the savings will be awesome. John Hsuan, UMC's chairman, notes that annual sales per employee on the company's 200mm lines are $500,000 to $750,000 but will jump to $2.5 million to $3 million with the 300mm wafer.

Nevertheless, there's no boom in 300mm fab construction. A few pilot lines are running at Intel and elsewhere. Some proprietary memory-chip companies like Infineon are putting up 300mm plants. But Strategic Marketing Associates in Santa Cruz, Calif., which tracks fab expansion, counts just eight full-sized, nonmemory 300mm fabs coming online this year and next. One is the UMC-Hitachi plant now ramping up production in Japan. TSMC and UMC will each open one more fab later this year. A second TSMC plant will start in 2002, when Chartered will also open its 300mm fab in Singapore. That leaves three. One is being built in Taiwan by a small, integrated-chip maker. By fall Texas Instruments will be shipping 300mm wafers for the first time from a plant in Dallas. Intel will open its first big-disk fab in Albuquerque in late 2002. That's it. Except for the memory-chip makers, all the other familiar names in semiconductor manufacturing are missing.

By the time IBM and France's STMicroelectronics open their first 300mm fabs in 2003, TSMC will have two running and UMC will have opened its joint venture in Singapore with Infineon, a $3.6 billion project. Come 2005, TSMC and UMC expect that as much as half of their greatly increased capacity will be in 300mm wafers. With each fab costing better than $3 billion, this adds up to a huge risk for the two Taiwanese companies. As Morris Chang concedes, "We have control of market share. We can't do anything about demand." Nevertheless, the "market share" assertion has to sound threatening if you're an integrated company. Says Chang: "It will be tough for the integrated-device manufacturers to keep pace."

It won't be easy, though, for the boards of the integrated-device makers to approve $3 billion-plus investments when their current operating statements are reminding them forcefully that expensive fabs can stand partly empty anytime the market cycles downward. Admitting that Motorola is watching while others are pioneering 300mm, Wildenberg says, "I don't think it's necessary to be first." He concedes that one question is, "Can you stay among the leaders in technology? If you start falling behind, you're out."

Still, skeptical directors will be even tougher to sway if the companies that don't build fabs not only continue to come up with bright ideas for new chips but also have a way to get them made in the least expensive and most technically advanced plants. In the early '90s, a common assertion in the semiconductor industry was "Real men own fabs." Maybe it's time to add, "Smart ones don't."

FEEDBACK: psiekman@fortunemail.com. Stories from FORTUNE's Industrial Management & Technology section can be found at fortune.com/imt. Executives in manufacturing and research and others eligible to receive FORTUNE's Industrial Edition can subscribe by calling 888-394-5472.