The labyrinthine vastness of Intel’s nearly completed D1D semiconductor factory in Hillsboro, Ore., is every bit as breathtaking as the microscopic intricacy of the microprocessors it will soon start making. Step into the sprawling subbasement of the $2.5 billion facility, and it feels like taking a fantastic voyage into the heart of a computer chip: Thousands of workers and technicians purposefully dart about the dense concrete catacombs like electrons coursing through a chip’s aluminum and copper pathways–testing mazes of high-voltage circuits, connecting miles of meandering pipes and ducts, and installing exotic support gear for the actual fabrication area two floors above. There, in a gleaming clean room the size of three football fields, dozens of equipment specialists scurry about in bunny suits and hardhats. Some painstakingly maneuver delicate chipmaking machines the size of Zambonis into position on dollies that glide on a cushion of compressed air. Next spring, when the place starts cranking out silicon wafers the size of a medium pizza–each embedded with hundreds of state-of-the-art microprocessors, containing nearly a quarter-billion transistors apiece–D1D will officially become the world’s most prolific chip “fab.”
But D1D is more than a Brobdingnagian metaphor for the Pentium. It’s also the latest steel-and-concrete embodiment of Moore’s Law, the guiding principle for the entire semiconductor industry that Intel co-founder Gordon Moore articulated 37 years ago when he first suggested that every year or so, chipmakers should be able to double the number of transistors and electronic components they can etch on a chip. Likewise, D1D is the physical manifestation of Intel itself, a company whose mission statement could well be “It’s the factory, stupid.” That’s because D1D and three fabs just like it that Intel plans to build by 2004 represent a controversial $10 billion bet that the company can erect what might be called Moore’s Wall. Intel is gambling that by pushing the state of the art in chipmaking faster than rivals are able to, it will reach a point where it can use sheer manufacturing prowess and capacity to undercut any competitor in price, performance, and variety. That means not just fending off would-be archrival Advanced Micro Devices and continuing to dominate the business of making chips for PCs, but also challenging Texas Instruments, IBM, Motorola, and a spate of smaller competitors in chips found in everything from cellphones to cars.
“Capacity is strategy,” says Andy Grove, Intel’s chairman and former CEO. “Henry Ford used it to revolutionize the automobile industry; the Japanese used it to push us out of the memory-chip business 25 years ago; we used it a decade ago to ignite the explosion of the PC industry. Now we’re using it again so we can broaden our business beyond the PC.”
Intel is making its move amid the deepest slump in the history of the IT industry. The company’s revenues–projected at about $26 billion for 2002–are running 25% lower than two years ago, dampened by the contraction of the PC business in the U.S. Neither Intel’s sales nor its largest customers’ fortunes show signs of an imminent uptick. At the same time another key consumer of semiconductors–the telecom industry–is barely breathing. And while Intel remains profitable, with an estimated $3.3 billion in net income this year, its earnings are down 8% from last year and 71% from 2000. Intel stock is down 80% from its all-time high of $74.87 in 2000.
Times are even tougher for rival chipmakers, many of which are bleeding red ink, laying off employees by the thousands, and scuttling plans for new plants. AMD is facing cash pressures due to unexpectedly wide losses. Other chipmakers, like Motorola, have announced intentions to eventually quit manufacturing chips altogether. Instead they will rely on so-called foundries in Taiwan that make chips for hire. That could mean having to take a number and stand in line to wait for capacity to open up, and settling for less-than-leading-edge process technology.
Which is one reason Intel is redoubling its efforts now. By investing heavily during a tech recession, Intel thinks it can leap a generation ahead in chip know-how and manufacturing ability. “This is the beginning of the consolidation and bifurcation of the semiconductor industry into a handful of leaders and lots of followers,” says CEO Craig Barrett, who headed manufacturing before moving to the corner cubicle at Intel headquarters in Santa Clara, Calif. “There will be an equivalent bifurcation in products and profit margins. The products that should command the highest margins are those at the leading edge of design and performance, which result from only one thing–having the best manufacturing technology.”
Sounds like typical saber rattling from a notoriously aggressive company. But there’s a second motivation in Intel’s bet on Moore’s Law–a kind of nerdy idealism that includes a pure, simple faith in technological progress and a genuine desire to help revive IT and telecom. Intel thinks its manufacturing capabilities will speed the introduction of incredibly powerful chips that take the Internet to the next level, enabling hundreds of millions of computers, phones, and other devices to be always tied to wireless networks. “We’re talking about a half-billion transistors on a chip, and perhaps even a billion,” says Paul Otellini, Intel’s president, COO, and likely the next CEO. “Suddenly there will be very little limit to what you can design into a single integrated circuit. If you want to talk about a golden age for semiconductors, that’s when it will be, and the IT and telecom and consumer electronics industries will be the biggest beneficiaries.”
Still, there’s no guarantee that the golden age is imminent, and without it, Intel’s profits and stock price could look pretty dull. Even if Intel widens its dominant 81% market share for PC microprocessors, it won’t generate enough incremental sales to use all that new capacity, nor will it get back to growing at its historical double-digit rates. Telecom might benefit from Intel’s creations, but it’s unclear whether that industry–even after it finds its feet again–will use enough chips to keep Intel growing. Skeptics abound, especially on Wall Street. “Intel has a phobia about capacity,” says Salomon Smith Barney analyst Jonathan Joseph. “They’re very concerned that they’ll miss the next upturn. But they clearly have too much capacity coming online; they aimed forward to hit the duck, and the duck isn’t there. Sooner or later they’ll have to adjust to maintain profitability, and that will mean closing some existing plants.”
There is a certain relentlessness to Moore’s Law: It predicts that technology will keep improving exponentially, no matter what the state of any one company, the tech industry, or the economy. For a chipmaker, falling off the pace of innovation is potentially fatal. Not that Gordon Moore ever intended his idea to have such an effect when he came up with it in a 1965 article for Electronics magazine. “I’m willing to take credit for whatever is associated with Moore’s Law,” quips Moore, now 73 years old.
Moore’s original edict was simply an extrapolation of trends based on how quickly the nascent chip industry had been able to incorporate more components into a single integrated circuit. (Carver Mead, the renowned Caltech physicist, was the one who later would dub the prediction Moore’s Law, which he has described as a “self-fulfilling prophecy more than a law.”) When Moore wrote the article, the number of components was doubling about every year, but then again, that meant growing from, say, 16 transistors to 32. In 1975, with chips becoming increasingly complex, he adjusted the figure to every two years. (For more, see “The Surprise Behind Moore’s Law.”) Lately Intel has been able to reduce the doubling period to about 21 months.
Doubling the number of transistors in a given space requires more than ingenious design. Chipmakers have to work with semiconductor-equipment manufacturers to devise manufacturing methods and gear that can etch transistors and other microscopic features that are one-third smaller than the previous norm. That entails developing better methods of lithography, finding more subtle ways to precisely deposit or remove minute quantities of metals and other materials on silicon wafers, and designing ever cleaner clean rooms. The transistors on the chips pounded out at D1D will be smaller than 90 nanometers across–so small that ten of them would fit in the diameter of a human hair–vs. 130 nanometers at the current state of the art.
Each generation usually involves building whole new factories or completely refitting old ones. Not surprisingly, the cost of fabs over the years has risen on a Moore’s Law-like curve, as has Intel’s capital spending (see charts). But the payoff can be as stupendous as the cost. Denser chips translate into higher capacity and output; open a new, state-of-the-art chip plant, and the cost of producing a chip usually falls by about a third literally overnight. That’s not all: The smaller parts perform better. Says Moore: “It’s a funny technology. By making everything smaller, everything gets better. But it also means if you lag behind your competition by a generation, you don’t just fall behind in chip performance, you get undercut in cost.”
In that statement lies the tyranny of Moore’s Law. “Every time we don’t live up to Moore’s Law, somebody else does,” says Andy Bryant, Intel’s chief financial officer. “If you stumble, something bad happens.”
And that’s why, as Barrett puts it, “we don’t adhere to Moore’s Law for the hell of it. It’s a fundamental expectation that everybody at Intel buys into. We dangle Moore’s Law in front of the new young minds that come here to work and say, ‘Hey, your predecessors were smart enough to figure this out for the past 20 or 30 years–why the hell aren’t you?’ We simply don’t accept the growing complexity of the challenge as an excuse not to keep it going.” For Barrett, Moore’s Law is a very big yardstick that he wields mercilessly.
The man responsible for the actual enforcement of Moore’s Law is Sunlin Chou, a courtly, 56-year-old Intel lifer who is senior vice president and general manager of the company’s technology and manufacturing group. Barrett gives him plenty of resources to work with, including nearly half of Intel’s R&D budget, which this year will top $4 billion, to develop the manufacturing methods that wind up in fabs like D1D. (That R&D figure is yet another number at Intel that has grown nearly exponentially over the years.)
Chou’s endeavor is an endless loop. D1D’s 90-nanometer fab isn’t even up and running yet, and already his researchers are devising the manufacturing methods and working with equipment suppliers for the next generation of Moore’s Law, which will reduce feature size to 60 nanometers, and which Intel hopes to begin rolling into fabs in 2004. Moreover, starting with the 90-nanometer fabs, Chou is introducing processes that will allow the factories to be more than one-trick ponies dedicated to a particular kind of chip. Processors, flash memory, and communications chips all have subtly different fabrication processes, but the new plants should handle all of them without slowing overall output.
Chou also chose the 90-nanometer generation to coincide with the once-a-decade segue to a larger wafer size at Intel’s new fabs, a shift many chipmakers around the world have put on hold to save money. From now on, new Intel fabs will handle wafers 300 millimeters (about one foot) in diameter, vs. the 200-millimeter wafers (about eight inches) at ten existing fabs. Explains Chou: “You increase the wafer size to offset the increasing cost of process complexity as you shrink feature sizes. It’s purely an economic enabler. It gives you more than twice as many chips per wafer yet again, and that further reduces the cost of an individual chip.”
One of Chou’s most important duties is to make sure the manufacturing process and fab design of D1D will be replicated precisely in the other three 90-nanometer facilities that will come online in the next 18 months (one in New Mexico, one in Ireland, and another in Oregon that will be retrofitted). This discipline–called the “copy exactly” technique–is something CEO Barrett pioneered during previous big fab build-outs to ensure that each new factory is immediately productive when fired up.
The secret to the copy-exactly strategy is counterintuitive–tying the hands of the design engineers. “Once we come up with a manufacturing process, we don’t let the chip design team tinker with it,” says Chou. “Believe me, it takes enormous discipline for the designer not to try to tweak things, because it goes against an engineer’s very nature.”
The net result of Chou’s efforts will be a double gain in productivity for Intel as D1D and its lookalike plants go online. Not only will Intel more than double its chipmaking capacity, but it will also be able to make chips that perform markedly better–Pentiums will bump up from today’s 2.4 gigahertz to nearly five gigahertz–and that cost on average a third less to produce. The same will go for Intel’s other big products, such as flash memory used in cellphones, handhelds, and digital cameras, and for the network and communications chips that the company hopes will power its next round of growth. That is, if Intel can come up with competitive designs–and if it can find customers willing and able to buy them.
In new markets, Intel must usurp well-entrenched rivals. Warns Texas Instruments: “Customers don’t just roll over.”
Grove’s prediction: “You know that saying, ‘The Internet changes everything’? I say, Just wait five years.”
Developing new markets is where Intel might find Moore’s Wall as hard to build as the pyramids. If no one’s buying, the technological advantage gained by the new plants will simply be academic. Wall Street doesn’t assign premiums for good intentions.
That’s not to say there aren’t opportunities for Intel to expand its business in PCs, which today account for 80% of its revenues. Servers built around Intel’s Pentium 4 and Xeon microprocessors are gaining market share thanks to the growing popularity of Linux and the increasing penetration of Windows software into enterprise computing. And Intel’s high-powered new Itanium processor, for specialized servers, which have sold poorly so far because of performance problems, may yet catch on when 90-nanometer versions hit the market. (At up to $4,300 a pop, the Itanium could generate serious revenues.) Then there’s mobile computing. Sales of notebook PCs continue to grow by 15% annually, and Intel has big plans to put even more of its silicon inside each and every one of them.
But communications chips are what Intel thinks will drive its growth. The chips allow notebooks to speak wirelessly to networks, enable cellphones to make calls, and help route web pages, e-mail, and streaming media around the Internet. Intel thinks it can win business by finding a way to marry computing and communication, quite literally on the silicon chips themselves.
Chief technology officer Pat Gelsinger dubs the strategy Radio Free Intel. Simply put, he wants Intel to incorporate, right into many of its processors, radio transceivers that can automatically detect and connect to hot new Wi-Fi wireless networks and even cellphone networks. “How can we beat Texas Instruments or Motorola, companies that have decades more experience than we do in communications technology?” Gelsinger asks. “By changing the rules and defining a new architecture for integrating communications into smart devices. We want to make a radio transceiver something that you expect to be just another feature of just about any device with a microprocessor.”
The most accessible market for Intel’s radio-enhanced processors is mobile PCs. By the end of the year Intel will begin shipping samples of specially designed chip sets for notebooks that include ultra-low-power Pentium processors, graphics chips, and other support circuits, and a built-in ability to attach to a Wi-Fi network. These chip sets will enable a notebook computer to sense and connect with wireless networks as its owner moves around, and even switch from one network to another on the fly. “In mobile computing, to focus on the processor performance as we have in the past would be missing the point,” says Anand Chandrasekher, the vice president in charge of the product line. “The trick is to make all the extra performance that wireless requires invisible, so it just works, and the user can count on it.”
A second big target for the Radio Free Intel initiative involves cellphones and PDAs–markets Intel competes in but doesn’t dominate. This year 400 million cellphones will be sold, and many of them will contain Intel’s flash memory chips. But phones are also getting smarter and beginning to resemble PDAs in their ability to handle address books, calendars, and the like. Meanwhile Intel’s XScale processor is the brains for most PDAs that use Microsoft’s Pocket PC software, and it recently won the support of Palm. It has a shot at becoming an industry standard, much as the Pentium is the standard processor in the PC.
Intel’s grand plan is to couple its XScale chip with flash memory as a way to get more of its chips into cellphones. It also plans to use the same part, attached to a new Wi-Fi chip, to make PDAs more versatile communicators. Ultimately Intel wants to put everything–the communications transceiver for both Wi-Fi and voice cellphone service, the XScale processor, and loads of flash memory–into a single part that would function equally well as the heart and soul of a PDA or a cellphone. Creating that can be achieved only if Intel can make chips with much smaller transistors, and if it can learn how to place radios, logic circuits, and memory in the same chip package without having their electrical signals interfere.
To make headway in cellphones Intel will have to usurp entrenched rivals like TI. The worldwide leader in communications chips, it commands more than 50% of the market for the processors in cellphones. “When another company, even Intel, announces its intentions to enter your space, your customers don’t just roll over,” says a TI spokesman. “Our relationships with cellphone-handset makers are well established.” Plus, while success in the microprocessor business depends largely on continually optimizing a single, complex chip and a few support circuits, big players in the communications chip business make their money by developing and supporting dozens of different kinds of more specialized chips, many of which don’t really need the latest, greatest manufacturing process. Then there’s the problem that profit margins for these chips–unlike those for Pentiums–are slim.
Intel has another major target in mind for its communication chips: the iron deployed by telecom carriers and equipment suppliers. Over the past four years Intel has spent $10 billion acquiring 27 companies that are involved in various aspects of networking technology, from network processors to wireless technology to optical-switching systems for fiber-optic networks.
Now Intel is focused on creating a network processor that would be the telecom world’s equivalent of the Pentium. Network processors are the superfast switching circuits at the core of the gear that routes digital data around the Internet. So far the world has no network processor standard; companies like Lucent, Nortel, and Cisco buy specially designed chips and create their own software tools, each in hope of gaining an edge in switching performance or programmability. The other appeal of this proprietary approach is that customers tend to stick with one supplier once they’ve become familiar with its programming tools.
That kind of diversity has to end, argues Sean Maloney, executive vice president for the Intel Communications Group. “I don’t deny that competition is good, but the telecom industry has never really benefited from Moore’s Law because most of those proprietary network processors weren’t made in large enough volumes to warrant leading-edge manufacturing technology. A standard part made with state-of-the-art manufacturing technology, coupled with a set of standard programming tools that are derived from PC software tools, could finally bring Moore’s Law to telecom.”
Maloney’s group recently announced a high-performance network processor chip called the IXP. Designed to be a key component in routers and digital switches, the IXP is capable of processing up to 6.6 gigabits each second of web pages, e-mails, streaming media, queries, and other Internet traffic–that’s roughly equivalent to the capacity of 4,500 home broadband lines.
There’s a little problem, though. For now, at least, the telecom industry is comatose, and nobody’s buying much gear from anyone. Indeed, revenues for Maloney’s business for the quarter ended Sept. 28 totaled $482 million, or barely 7.4% of Intel’s total sales. Worse, sales were off 17% from the year-earlier quarter and down 10% from the previous quarter. That’s not a very encouraging way to begin broadening Intel’s business or driving growth. “We had higher volume predictions than we actually have now,” concedes CFO Bryant. “‘When will it start growing?’ is more a macroeconomic question than a technology question, but we do believe there will be a return to growth.”
The ebullient Maloney, as might be expected of a man who once ran Intel’s sales and marketing group, thinks there’s a silver lining for Intel in the telecom collapse. Says he: “My telecom customers have laid off 270,000 people in the past two years. A lot of them are engineers who designed their proprietary networking equipment. The telecom companies will never have that much engineering capability again. Our proposition will be to let us take care of developing the core components so that they can concentrate on tailoring gear and services to particular markets. Then they can ride along on the coattails of Moore’s Law just as the computer industry has.”
Such optimism may sound weird coming from a company whose unofficial mission statement has long been “Only the paranoid survive.” But optimism about Moore’s Law runs just as deep in the Intel psyche. Time and again pundits have warned that chipmaking will top out as circuits shrink to the size of atoms. But Sunlin Chou and his peers in the chip industry keep finding new ways to cram even more transistors onto silicon. “Opportunities aren’t dwindling, they’re exploding,” says Chou. Besides, he adds, “in the end, Moore’s Law is a philosophy as well as a strategy. It gives us the confidence to believe in the future.”
Just ask Andy Grove, the Hungarian emigre who, for all his success and notoriety, has seen his share of dark moments. You have to have faith, he says, and extrapolate where technology could lead, much as Gordon Moore did 37 years ago. In his original article, written when telephones still sported rotary dials, and color TV was a novelty, Moore presciently predicted that “integrated circuits will lead to such wonders as home computers, automatic controls for automobiles, and personal portable communications equipment.”
Here’s Andy’s prediction: “You know that saying, ‘The Internet changes everything’? People now are backing away from it, but I say, Just wait five years. Hundreds of billions of dollars we now spend on voice telecommunications will become a freebie–just like [Cisco CEO] John Chambers has said. That’s Moore’s Law at work. The entire entertainment industry will be digitally distributed over broadband networks. [Media companies are] going to tip over, because one of them, with its back to the wall, will make the transition, and the others will have to follow. That’s Moore’s Law at work. Houses will be wireless, broadband will be delivered wirelessly, and home and portable computers and consumer electronics are going to be built to facilitate all of the above. Okay, it hasn’t happened in the first five years; it’s going to take ten. And there will be a lot of pain for some. But it will happen, and we’ll all benefit.”
That’s why, at a time when other chipmakers fear that living up to Moore’s Law might be a bridge too far, Intel still believes it’s the only way to go.
This article first appeared in the November 11, 2002 issue of Fortune.