The Fabless Revolution

In the spring of 1996, Jensen Huang sat in his office in Sunnyvale and counted, line by line, how long his company could stay alive. Nvidia had been incorporated for three years. It had taken roughly twenty million dollars from Sequoia Capital and Sutter Hill, hired a hundred engineers, shipped one product, watched that product fail in the market, and nearly burned through everything it had on the second. The first chip, the NV1, had been a clever device that solved the wrong problem; the second, the NV2, was a custom design intended for Sega’s next console, and after a year of work Huang had concluded the architecture was technically inferior to what Microsoft’s new Direct3D standard would require. He could deliver a product he believed was already obsolete and take Sega’s money, or he could tell the customer the truth and watch payroll run out within a month. He chose the second. Without the kindness of Sega’s American CEO, who agreed to convert the canceled contract into a five-million-dollar equity investment, Nvidia would have died that quarter. Huang had said later that the call to walk away from Sega gave him six months to live.

The thing that was not obvious in the room, in 1996, was that the entire structural premise of his company depended on a different bet, made elsewhere, by other people. Nvidia did not own a fab. It had never owned a fab. It had no plan to own a fab. Whether the next chip lived or died turned not on a production line that Nvidia controlled, but on whether some other company on the far side of the Pacific could be persuaded to manufacture for him on terms that let a startup with one month of cash survive its own redesigns. The wave of companies that defined the next thirty years of American semiconductors, the ones that would in time include the most valuable corporation in the world, were all making a version of this same wager. They had decided to be designers only. They were betting their lives that someone, somewhere, would build their chips.

The bet had been placed first, more than a decade earlier, by a small company in Milpitas that almost no one outside the industry remembered. Chips and Technologies was incorporated in December 1984 by Gordon Campbell, a former Intel marketing executive, and Diosdado Banatao, a Filipino-born engineer who had already designed the first single-chip CMOS Ethernet controller at Seeq Technology. They wanted to build something specific: a small set of integrated circuits that could replace the nineteen separate proprietary chips IBM used in its Enhanced Graphics Adapter card. If they could compress the EGA into four chips, the makers of IBM-compatible PCs in Taiwan and Texas could buy a kit and stop paying IBM’s premium for the original silicon. The market was waiting. The problem was that no American semiconductor maker would manufacture for them.

The reason was not technical. The reason was that every fab Campbell visited was owned by a company that competed with him. Intel had a fab. Texas Instruments had fabs. Motorola had fabs. National had fabs. AMD had fabs. None of them was going to put Chips and Technologies’ designs ahead of their own. A merchant fab in the United States, operating purely as a contract manufacturer, did not exist. Campbell and Banatao went looking in Japan and ended up partnering with two second-tier Japanese makers, Toshiba and Seiko, that had spare wafer capacity and were willing to sell it. The first EGA chipset shipped in September 1985. By the November COMDEX, more than half a dozen board makers had built EGA-compatible cards on Chips and Technologies’ silicon. The company went public on NASDAQ in 1986, twenty-two months after incorporation, and by the late 1980s was earning roughly a hundred million dollars a year from a building that contained no manufacturing equipment. Banatao would later be called the father of the fabless semiconductor industry. The cleaner description was that he had built a real chip company without ever pouring a foundation for a fab, and he had proved the model could pay.

A handful of contemporaries had reached the same conclusion. Suhas Patil, an Indian-born computer architect, had founded a small startup in Salt Lake City in 1981 that, in 1984, reincorporated in California as Cirrus Logic and moved south to Fremont, designing mass-storage controllers and later graphics and audio chips that, like Chips and Technologies’, were manufactured by someone else. By the end of the 1980s, perhaps a dozen small American firms were designing chips and contracting their fabrication out, mostly in Japan, occasionally at Singapore’s Chartered Semiconductor or at the new Taiwanese foundry Morris Chang had founded in 1987. They did not yet have a name as a category. In 1994 a marketing executive named Jodi Shelton organized about forty-five of them into a trade association called the Fabless Semiconductor Association. Until then, the integrated device manufacturers had treated the design-only firms as a fringe model that worked for niche products. The FSA’s argument was that the fringe was the future.

The companies that would carry the model into geopolitical importance were already in motion by the time the FSA was incorporated.

In July 1985, two former MIT and UCLA professors, Irwin Jacobs and Andrew Viterbi, sat with five colleagues from their previous venture, Linkabit, in the den of Jacobs’s house in San Diego, and incorporated a new firm. They called it Qualcomm, a contraction of Quality Communications. Jacobs had written one of the field’s standard textbooks on digital communications. Viterbi had given his name to the algorithm that decoded convolutional codes. Their first product was OmniTRACS, a two-way satellite messaging system for trucking fleets that needed to track their tractors across the continental highway network. OmniTRACS paid the bills. It was not what the company had been founded to build. What Jacobs had concluded, over the course of his Linkabit years, was that the way the world’s cellular networks were going to be designed in the coming decade was wrong. The early-1990s standard for second-generation mobile, called Time Division Multiple Access, sliced a frequency channel into time slots and gave each user a turn. Jacobs believed a different scheme called Code Division Multiple Access, originally a military spread-spectrum technique, would deliver four to six times the capacity per cell tower if it could be made to work in commercial silicon at consumer prices.

Almost no one in the industry believed him. The Cellular Telephone Industries Association rejected CDMA in 1989 in favor of TDMA. Ericsson treated Qualcomm’s claims as marketing hyperbole. Qualcomm engineers built a working CDMA prototype in San Diego, drove it around in a fleet of vans, and handed reporters and regulators printouts of the spectral efficiency numbers. By 1993, the holdouts in North America had been convinced; CDMA was adopted by a parallel standards body as IS-95. The first commercial CDMA system ran in Hong Kong in 1995, with Korea and the United States following within a year. By 1997, more than half the cellular subscribers in North America were on CDMA networks. Ericsson and Qualcomm settled their patent disputes in 1999 in a cross-licensing agreement that handed Qualcomm’s infrastructure manufacturing arm to Ericsson. In December 1999, the company sold its handset manufacturing business to Kyocera. The remaining Qualcomm was a designer of two things: the patent portfolio that any CDMA-compatible carrier had to license, and the silicon that ran inside every CDMA handset, a family of chips its engineers had labeled the Mobile Station Modem, or MSM.

The MSM was, by the late 1990s, the closest thing the wireless industry had to a single point of dependence. A Korean phone maker selling into Verizon’s CDMA network bought MSM chips from Qualcomm. A Japanese phone maker selling into KDDI’s CDMA network bought MSM chips from Qualcomm. The chips were designed in San Diego and fabricated by partner foundries, principally TSMC, which by 2004 was running Qualcomm’s first ninety-nanometer MSM parts on its low-power process line. Qualcomm had not chosen to be fabless out of doctrine. It had become fabless because spinning off the manufacturing arms in 1999 left the design business as the part that scaled, and because TSMC by then could deliver every node Qualcomm wanted on the schedule Qualcomm needed.

In Los Angeles, a slightly different story was unfolding around the same time. In 1985, Henry Samueli, a young UCLA electrical engineering professor, had set up a multidisciplinary research program in his department to design integrated circuits for digital broadband. The premise was that the analog modems the cable and telephone industries used could be replaced by digital signal processing on a single chip. One of his Ph.D. students was Henry Nicholas, a former TRW engineer who had returned to UCLA after industry. In August 1991, Samueli and Nicholas put down five thousand dollars each, rented an office in Westwood, and incorporated Broadcom Corporation. Their first product was a chip set for the cable industry that allowed a coaxial line to carry digital data at megabit speeds, the silicon foundation of the cable modem. By the end of the decade, Broadcom’s chips were inside almost every cable modem, every wireless router, and a majority of the network interface cards in PCs and servers. The company went public in 1998. None of its products were fabricated in a building Broadcom owned.

The Samueli-Nicholas pattern, professor and student leaving the lab to commercialize the silicon they had been writing papers about, was repeating across the western United States in the late 1980s and early 1990s. The structural fact was the same in every case. None of these companies could have existed without a foundry willing to build for them. The interdependence was reciprocal and tightening with every new node.

By the time Jensen Huang began counting payroll in his Sunnyvale office in 1996, that interdependence was about to become his survival strategy.

Huang had not started in graphics. Born in Tainan in 1963 and brought to the United States as a child, he had grown up in Oregon, attended Oregon State as an undergraduate, taken a master’s at Stanford, and spent the 1980s designing chips for LSI Logic and AMD. By the early 1990s he had concluded that the most underserved problem in personal computing was 3D graphics for consumers, a market then divided among a half-dozen workstation vendors and almost nonexistent for the IBM PC. In April 1993, he met two engineers, Chris Malachowsky and Curtis Priem, at a Denny’s roadside restaurant on Berryessa Road in East San Jose, the same chain where Huang had washed dishes as a teenager. Malachowsky had spent his career at Hewlett-Packard and then at Sun Microsystems, where he and Priem had collaborated on Sun’s GX graphics products. Priem had built his reputation a decade earlier, at a small Vermont firm, by architecting IBM’s Professional Graphics Adapter, the first true graphics processor for personal computers. The three founders chartered the company on April 5, 1993, with forty thousand dollars between them. They named it Nvidia, from the Latin invidia, which meant envy. They hired no fabricators because they had no intention of building any.

The first product, the NV1, shipped in 1995 as part of a board called the Diamond Edge 3D. It tried to do too many things at once. It rendered graphics, processed audio, and spoke to a Sega-style joystick, all from a single chip that combined functions in a way that the emerging standards did not anticipate. Nvidia had bet, with Sega’s encouragement, that the future of 3D rendering would describe shapes as quadrilaterals rather than as triangles. Microsoft’s Direct3D, released in 1996, settled the question against them. Diamond Multimedia shipped roughly a quarter of a million NV1 boards into retail channels. Almost all of them came back. Inside the company, by the spring of 1996, half the staff had been laid off, the NV2 was failing to track Microsoft’s evolving API, and the runway was down to a single payroll cycle.

What saved Nvidia first was the candor with Sega and the equity check that followed. What saved Nvidia second was a redesign that abandoned almost everything the founders had originally believed. The third product, the RIVA 128, would speak triangles. It would render only graphics, with no audio side-trip and no proprietary input scheme. It would target Microsoft’s Direct3D 5 directly. And it would be cheap, fast, and built on a process node that none of the company’s competitors expected to see at consumer prices. The team produced the design in nine months and put it into fabrication at SGS-Thomson, the European IDM that had agreed to take Nvidia’s wafers when no one else would. The RIVA 128 began shipping in August 1997. By the end of the year Nvidia had sold roughly a million units. Within four months, the cash crunch that had nearly killed the company in 1996 was over.

The next decision was the one that would matter for thirty years.

Huang had not been satisfied with the SGS-Thomson relationship. The yields had been mediocre. The cycle times had been too long. Nvidia’s venture investors, by then including Sequoia’s Don Valentine, were pushing for a better foundry, and the only foundry in the world running the volumes and the process technology that Nvidia needed at the price Nvidia could pay was the one Morris Chang had founded in Taiwan ten years earlier. Nvidia had tried to interest TSMC’s San Jose office and gotten nowhere. In 1997, Huang sat down and wrote a letter to Chang directly, a four-paragraph plea that explained who Nvidia was, why it needed TSMC, what the next product would be, and why the customer relationship was worth Chang’s personal attention. Huang sent it to TSMC’s Hsinchu post box and waited. The letter, by Chang’s later account, made its way to his desk and made him both curious and slightly irritated. He called Huang back. They met. By 1998 a multi-year supply agreement was in place, and the variant of the RIVA 128 known as the ZX, along with every Nvidia product that followed, was being fabricated in Taiwan.

What that letter purchased was not just capacity. It was a working relationship between a fabless designer in California and a foundry in Taiwan that would, over the following decades, deepen into something closer to a strategic union than to a vendor agreement. TSMC’s process engineers learned the quirks of Nvidia’s designs. Nvidia’s designers wrote their layouts to take advantage of TSMC’s specific design rules. Each new node was rolled out in concert. By 1999 the relationship was sufficiently robust that Nvidia’s fourth-generation product, called the GeForce 256, could ship on a 220-nanometer process at a fixed schedule with predictable yields. The GeForce 256, announced on August 31, 1999 and released October 11, was the chip Nvidia called the world’s first GPU, a marketing term Huang’s team had defined as a single-chip processor with integrated transform, lighting, triangle setup, and rendering, capable of processing at least ten million polygons per second. Whether the term GPU had appeared in earlier literature was, by then, beside the point. Nvidia owned it. The category had a name, the company had a moat, and the foundry that had said yes to Huang’s letter had a customer it would still be supplying when the chips it manufactured for Nvidia, two decades later, were running every large language model in the world.

The pattern that the fabless companies were establishing in the late 1990s rewrote the economics of the industry on every axis that mattered. The capital expenditure profile was the most visible change. An integrated device manufacturer in 1995 spent something on the order of twenty to thirty cents of every revenue dollar on new fabs and equipment, plus another fifteen to twenty cents on R&D split across both process technology and product design. A fabless company spent only on product design. By the early 2000s, financial analysts routinely noted that fabless firms posted gross margins five to ten percentage points higher than IDMs in the same product categories, because the absence of fab depreciation outweighed the foundry’s manufacturing markup. Nvidia and Qualcomm, by the late 2000s, were running gross margins in the high fifties or low sixties on commodity markets where Intel’s competitors in memory had margins in the twenties.

The product cycle was the second change. A vertically integrated firm scheduled its launches around its own fab. A fabless firm could schedule launches around TSMC’s whole roadmap. The multi-project wafer services that the foundries provided, descended from the MOSIS shuttles of the 1980s, meant that even a small fabless firm could prototype a design at a fraction of the cost of a full mask set. A startup with five million dollars and ten engineers could produce a working SoC, a feat that in 1985 would have required hundreds of millions of dollars and a manufacturing line. The barrier to entry in chip design dropped and the diversity of designs exploded.

The third change was less visible at the time and would matter most. Every fabless firm had, by definition, a manufacturing dependency that was not a polite vendor relationship but an existential one. If TSMC stopped shipping wafers to Nvidia for ninety days, Nvidia stopped having products to sell. The competitive advantage the model generated was purchased with a structural exposure to a single foundry. Through the late 1990s and into the 2000s, when the foundry industry was diversified across TSMC, UMC, Chartered, IBM Microelectronics, and a half-dozen others, that exposure looked manageable. As one foundry pulled ahead, node by node, it was about to become specific.

Not every American chip company had taken the fabless path by the late 1990s. Intel still ran its own fabs and would for a long time. AMD, whose founder Jerry Sanders had made owning manufacturing into a personal creed, would not spin off its fabs as GlobalFoundries until 2009 and would not move its leading-edge wafers to TSMC until well after that. By the late 2010s, every major American designer of leading-edge logic chips, with the partial exception of Intel, was a customer of a foundry on the western Pacific.

The cumulative effect, by the early 2020s, was an American semiconductor industry that designed the most valuable silicon on Earth and manufactured almost none of it. Nvidia’s data-center GPUs, the chips on which the artificial-intelligence boom of the 2020s would run, were designed in Santa Clara and fabricated in Hsinchu. Qualcomm’s Snapdragon mobile platforms were designed in San Diego and fabricated in Hsinchu. Broadcom’s networking switches were designed in San Jose and Irvine and fabricated in Hsinchu. The pattern was so consistent that it had ceased to be a pattern and had become the structure of the industry.

That had not been the goal of the men in the Denny’s booth in 1993, or of Campbell and Banatao in 1984, or of Jacobs and Viterbi in 1985, or of Samueli and Nicholas in 1991. None of them had set out to make the United States dependent on a single foundry on a contested island. They had set out to design chips without paying for fabs. They had taken the seam that Mead and Conway had argued could be cleanly cut, and they had cut it. What they had not yet seen was the second-order consequence of their freedom. The factory floor where their chips were made was no longer in the same country as the design office that drew them. The decision to make it that way had been a thousand small decisions, made for a thousand small reasons of margin and speed and survival, by founders who could not have imagined, on the days they signed their incorporation papers, the strategic question their choices would one day pose.

The question would arrive in due course. For now, in the autumn of 1999, the GeForce 256 was shipping out of TSMC’s fabs in volume, and Jensen Huang’s company was on its way to becoming, briefly, the most successful technology IPO of the year.