Semi survival

By Michael Kanellos
Staff Writer, CNET News.com
January 22, 2003, 4:00 a.m. PT

Survival in the semiconductor market will soon depend on having friends in high places.

The exploding costs of fabrication facilities, or fabs, combined with technical hurdles surrounding the design of the next generation of chips, are creating a bifurcation that will change the face of the industry, executives and analysts say.

On one side will be Intel and a select few that can afford their own fab plants--which will cost between $2.5 billion and $3 billion to build in 2003 and $6 billion by 2007--and perform basic research on transistor design or new chip materials. These new fabs will process wafers with 300-millimeter diameters, larger and more complex to make than today's 200-millimeter variety.

On the other side will be everyone else. They will have to share fabs, pool research, buy technology or rely more heavily on outside foundries, which in turn will have to seek help.

"It's becoming almost impossible for the smaller semiconductor companies--smaller being anything less than an Intel or a Texas Instruments or an NEC--to afford these 300-millimeter fabs on (their) own nickel," said Brian Halla, chief executive of National Semiconductor. "What you are going to see is more and more partnerships."

Historically, companies spend 20 percent to 30 percent of revenue on capital expenditures, meaning that a fab can be justified only if it costs about one-third or less of annual revenue. In 2002, only Intel and Samsung, with about $23 billion and $9 billion in annual sales, respectively, could spring for new fabs.

		Related story A fab construction job Rebuilding an IBM chipmaking facility proves no easy task.

Other large manufacturers will need outside help to make their chips. TI and Infineon, the German memory maker, have said they will outsource some production to Shanghai's Semiconductor Manufacturing International. Motorola, ST Microelectronics (the world's third-largest chipmaker), Philips and Taiwan Semiconductor Manufacturing Co. (TSMC), are collaborating on a wide variety of ventures: how to build 300-millimeter factories, designs for new types of memory and techniques for making 90-nanometer chips, among other projects.

"It takes a lot more horsepower to get these technology road maps to hit on target," said Alex Pepe, vice president and director of strategy for technology and manufacturing in Motorola's Semiconductor Products Sector, which is increasing its partnerships to cut costs.

Motorola's Austin, Texas-based chip division is in many ways emblematic of the trend. In 1998, the company operated 25 fabs and 15 sites for testing and assembling chips. By the end of the year, Motorola will have eight fabs and two test centers. Any future factories will be built jointly with others, while outsourcing manufacturing--around 20 percent now--will likely increase.

Cooperative research and development will also become norm, Pepe said. Motorola will increasingly license its technology to start-ups, large companies and even equipment makers, for financial gain.

In this new era, Intel could gain an even greater opportunity to undercut opponents or enter new markets because of its unique self-sufficiency. Asian foundries such as TSMC, meanwhile, will grow in power, and IBM may become the de facto R&D department for an array of semiconductor companies. Other chipmakers may simply fade away.

"There is going to have to be some consolidation in the industry," said AMD Chief Executive Hector Ruiz, whose company obtained manufacturing techniques from IBM and will likely co-own its next factory with another company. "Outside of Intel, I don't know who could do it (alone). There has to be a coalition."

Rule of the Rock
The harsh realities come partly as the byproduct of Rock's Law. Named after venture capitalist Arthur Rock, this dictum holds that the cost of building a semiconductor fab will double every four years, according to Dan Hutcheson, CEO of VLSI Research, which tracks the semiconductor-equipment industry.

In 1985, a fab that processed wafers with 5-inch diameters cost about $100 million, Hutcheson said, and the average size of components on the chip measured a micron, or 1,000 nanometers. The $3 billion megafabs coming online now process wafers with diameters of 300 millimeters, or 12 inches, and will contain components that average around 90 nanometers. (A nanometer is a billionth of a meter.)

Hutcheson said the cafeteria at one of these new fabs could cost more than an entire plant of years past. "The costs are getting outrageous. That is driving some fundamental structural change," he said.

The $3 billion price tag is largely soaked up by equipment. Lithography machines used for "drawing" circuit lines can cost $17 million or more each and weigh 40,000 pounds. Just getting equipment in the door can take months.

The payoff comes in vastly improved efficiency: New, expensive fabs produce far more chips at a lower cost. The semiconductor industry accounted for about $22 billion in revenue in 1985, when about 2,700 fabs were in operation. In 2002, the industry amassed $139 billion in revenue out of 600 fabs.

Still, greater productivity creates risks of equal size. If a company builds a $3 billion megafab, it needs to find customers for all the cheaper, smaller chips it can suddenly make in far larger volumes. If that doesn't happen, operating expenses will send the company into the red.

Aside from Intel and Samsung, according to a recent survey from iSuppli, every other semiconductor manufacturer in the world pulled in $6 billion or less in revenue, making them candidates to seek outside help to make chips. Many are still trying to unload factories and real estate acquired in the last boom.

"What the industry is faced with is very significant capital costs, very high revenue requirements, and very basic changes in process technology," said Sumit Sadana, director of strategy for IBM Microelectronics. "The financial health of most of the companies in the industry is not pristine, to put it nicely."

Fabs vs. foundries
So why own a fab? To control your own destiny. Products can be delayed, advanced or tweaked at will, while production can be monitored constantly.

"You get to (determine) your own capacity. You don't have to wait in line. You don't have to pay someone's margin, and you absolutely have the lowest costs," Intel President Paul Otellini said.

The Santa Clara, Calif.-based company believes this strategy will pay off this year when it starts making chips on the 90-nanometer process. The same 90-nanometer factories will be capable of manufacturing microprocessors, communications chips, and subsequently flash memory and other products. Intel will also incorporate communications functions into processors. In other words, whatever chips it wants to mix and match, it can.

Concurrently, by using the same production lines for a wider variety of chips, the company can insulate itself from owning idle, excess capacity, thus reducing costs.

Foundry customers, by contrast, have to pay for their services while shuttling back and forth to Taiwan. In addition, foundries have at times had more difficulty than autonomous manufacturers in adopting new technologies or manufacturing methods, Hutcheson and others say. TSMC's move from the 180-nanometer chipmaking process to the 130-nanometer process proved difficult and contributed to delays for products from Transmeta and others.

Semiconductor designers can also lose their edge by relying too much on foundries. "I have not yet seen anyone who has taken a 0.13-micron process (from a foundry) and done well," said Syed Ali, CEO of Cavium Networks, which designs processors that perform security functions in chips. Many chip start-ups, he said, are staffed with engineers who "have never seen a transistor under a microscope."

Cavium outsources manufacturing to TSMC but places heavy emphasis on circuit design. The company also employs a number of old-timers: Many at Cavium, Ali added, worked on Digital's semiconductor division in the preceding decade.

Nevertheless, foundries are rapidly growing in sophistication and importance. Collectively, they will grow by 31 percent annually and likely account for 40 percent to 50 percent of the world's semiconductor output by 2010, said Chuck Byers, director of worldwide marketing for Taiwan Semiconductor. He said TSMC will start making 90-nanometer chips around the same time as Intel, adding that neither company is immune from the difficulty of shrinking chips.

Part of the reason the gap is closing between fabs and foundries is that semiconductor-equipment makers have shifted from selling basic equipment to delivering more nearly complete solutions that can be put on production lines fairly rapidly.

"You (just) don't buy the hardware anymore. You get a recipe for the chemistry and then you fine-tune the equipment to yield what you want," AMD's Ruiz said. "It is like a machine gun...Today, the equipment guys are very powerful."

Driven by research
Even if manufacturing parity can be achieved, however, smaller companies are going to face a problem because of the changing nature of chip design. For more than 30 years, designers have increased the performance of their chips by shrinking and simultaneously doubling the number of transistors on their products--a la Moore's Law, which dictates that the number of transistors on a chip doubles every two years.

But the shrinking of transistors is hitting a wall, and to keep the gains from Moore's Law on track, designers are being forced to come up with creative solutions: strained silicon, silicon-on-insulator, and double- and triple-gate transistors.

"The foundries are entering a precarious time right now. While they see all these companies going asset-light, they face a large R&D effort in process technology to stay a step within Intel and IBM," Big Blue's Sadana said. "You will see fewer and fewer companies majoring in process innovation."

		Special report China awakes to new destiny The chip industry and others are finding gold in the emerging economy.

In IBM's deal with Chartered Semiconductor, the two companies will work together to develop processes for 90-nanometer and 65-nanometer manufacturing, a move that will effectively open Big Blue's labs to Chartered and its customers. Similarly, Chartered will be able to postpone opening a 300-millimeter fab while IBM will be able to fill its plant with new customers because the factories of the two companies will be equivalent.

Sony will also rely on IBM to provide intellectual property and manufacturing expertise for the "cell" processor that will likely go into the next version of PlayStation. Sony with Toshiba designed its own chip for PlayStation 2.

The catch: IBM charges significantly for such arrangements. AMD, for example, could pay hundreds of millions under its deal with Big Blue, according to one source. And TSMC says the R&D centers of the world are ignoring a bigger factor: Companies can't afford to integrate these complex technologies into their chips. Some TSMC customers refrained on moving to 130-nanometer production because the cost estimates didn't justify the move.

"We have reached a point as an industry where our ability to innovate now exceeds our ability to profitably capitalize on that innovation," Byers said. "This is all very sexy stuff, but it is going to be expensive and someone is going to have to pay for it."

Larger, more established companies with revenue between $3 billion and $5 billion will develop hybrid models. National Semiconductor, for instance, designs its own chips but then has TSMC draw the manufacturing model and build prototypes. Subsequently, National takes the chips back and performs mass manufacturing.

"In the olden days, it used to be intellectual property drove everything," Halla said, adding that National will outsource some, but never all, of its chips. "Now, you've got to do your manufacturing with a great deal of finesse as well."

To that end, the Philips-ST-Motorola research alliance somewhat resembles a U.N. peacekeeping mission. The common research facility is managed by three executives--one from each company--and each critical engineering project has key technologists from each company.

Partnerships, though, aren't always easy. In 2002, AMD entered a multiyear agreement to develop manufacturing technology with United Microelectronics and co-build a fab in Singapore with the company. This month, AMD said it ended the technology deal, struck the new one with IBM, and is looking at new partners for building a fab.

Consolidation will likely be a fact of life for many small to midsized companies. Hutcheson, though, noted that history seems to indicate that the odds won't discourage new companies from trying to break into the business.

"Semiconductors are like the airline industry," he said. "They are sexy, and investors are always willing to lose money on the airline industry."