A Denny's, the family restaurant chain, on Berryessa Road in Silicon Valley. In 1993, three friends, Jensen Huang, Chris Malachowsky and Curtis Priem, gathered in one of its booths. They had come to talk through an idea: a chip that would draw realistic 3D graphics on a personal computer.

"We would hang out here for hours and talk about what NVIDIA would be," Huang later told a local TV station. On why they picked the place, his answer was candid: "It had all the coffee you could drink and no one could chase you out."

Thirty years later, on September 26, 2023, Huang returned to the diner and received a commemorative plaque from Denny's CEO Kelli Valade. The local station, NBC Bay Area, introduced the NVIDIA of that day as a company "now worth a trillion dollars." In May of the same year, invited to speak at the commencement of National Taiwan University, Huang had described his starting point this way:

"We founded NVIDIA to create accelerated computing. Our first application was 3D graphics for PC gaming."

As a founding story told by today's champion, the words are neatly composed. Open the prospectus (424B4) that the company filed with the U.S. Securities and Exchange Commission (SEC) for its listing in January 1999, however, and you find a plainer sentence.

"The Company's objective is to be the leading supplier of high performance 3D graphics processors for PCs."

An empire of computing was not in view from the start. What the filings of the same era record is a semiconductor startup that tried to make its own approach the standard and lost to the industry's standard, saw major customers bought by rivals, received an order from the SEC over its accounting, and still kept putting out the next chip.


The business plan that was never finished

Around the time they were working out the idea at Denny's, Malachowsky and Priem were engineers who had worked on the design of "GX" graphics at Sun Microsystems, and Huang was a director of the CoreWare business unit at LSI Logic. Before that, from 1983 to 1985, Huang had worked on microprocessor design at AMD (Advanced Micro Devices).

According to the financial statements in its registration statement for the listing (Form S-1), NVIDIA's inception was April 5, 1993, and the company was incorporated in California. The company's old official timeline preserves a different date as well: "January, 1993," with the entry "NVIDIA is founded."

When they raised their first money, they had no finished business plan in hand. At the company's own conference, GTC, in 2018, Huang recalled, "I didn't know how to write a business plan and never did finish one." On the same stage, Malachowsky told how, at a meeting with venture capitalists (VCs), Huang said, "I worked on it last night, but didn't finish it," and the VCs replied, "That's OK, we wouldn't have believed it anyway." In a separate interview, Huang has said the door opened when Wilf Corrigan of LSI Logic called Don Valentine of Sequoia Capital and told him to "invest in this kid."

In June 1993, Tench Coxe of Sutter Hill Ventures and Mark Stevens of Sequoia joined the board. That same year, 1993, the company issued 4,303,000 shares of Series A preferred stock at $0.50 a share (about $2.15 million). Then, in November 1993, one more person joined the board: Harvey C. Jones Jr., who had led Synopsys, an electronic design automation (EDA) company.

Thirty-three years later, on November 12, 2026, the company's announcement of Huang's resignation stated that the board had appointed this same Jones as its "first-ever Chair of the Board."


A chip where "90 percent [didn't] work," and Sega's $5 million

In 1993 and 1994, NVIDIA had no revenue. After a stretch devoted entirely to development, its first product, the "NV1," was introduced in May 1995. According to the prospectus, the NV1 was a multimedia chip that handled 3D graphics, video and audio on a single part, and it was targeted primarily at the game console market.

On the podcast Acquired in 2023, Huang said NVIDIA's first business strategy was to be "a game console inside the PC." The technology it chose, unlike the industry mainstream, was an approach built on "forward texture mapping" and curves.

The bet did not pay off. At GTC in 2018, Huang described the NV1 this way: "We taped it out," he said, "it came back, and 90 percent [didn't] work. But we were so excited this was back, we took it to Computex." The company showed the same demo at the same trade show the following year, he said, and nearly went out of business. In Huang's words, a year and a half after the founding, "the technology was wrong, the market strategy was wrong, architecture was wrong, and the execution wasn't good."

The industry's standards, too, went a different way from the NV1. In September 1995, four months after the NV1 was introduced, Microsoft released the first version of DirectX, its game development kit for Windows. The prospectus is candid about what happened:

"The NV1 was developed in the absence of industry standards with the goal of establishing the Company's proprietary NV technology as a 3D graphics standard. By the end of 1996, the PC industry had broadly adopted Microsoft's Direct3D and SGI's OpenGL 3D APIs."

As Huang explained at National Taiwan University, Microsoft's Direct3D was an approach based on "inverse texture mapping and triangles." Revenue in 1995 was $1.18 million, and the net loss was $6.38 million. The prospectus says that revenue from the NV1 fell sharply and that the company stopped selling it in the first quarter of 1996.

Closely tied to NVIDIA in this period was Japan's Sega Enterprises.

In his 2023 speech at National Taiwan University, Huang looked back on a decision he made over the Sega contract:

"I contacted the CEO of Sega and explained that our invention was the wrong approach, that Sega should find another partner, and that we could not complete the contract and the console. We had to stop. But I needed Sega to pay us in whole or NVIDIA would be out of business. I was embarrassed to ask. The CEO of Sega, to his credit and my amazement, agreed. His understanding and generosity gave us six months to live."

In a July 2026 article, the gaming outlet VGC wrote that Shoichiro Irimajiri, then Sega's president, agreed to invest $5 million in NVIDIA, and that Huang has recalled this as what kept the company afloat. The same article also quotes Huang in 2025: "$5 million was a big amount of money for Sega at the time. I told him that if they invested that money in us, it was likely to be lost."

Trace the notes to the S-1 filed in March 1998, however, and the chronology of the deal looks a little different.

The S-1 describes a transaction with an unnamed "third party." According to it, in May 1995 NVIDIA entered into a five-year strategic alliance agreement with this third party and began developing the "NV2," to be incorporated into the third party's products. In addition, "as part of this agreement," the same third party purchased 750,000 shares of Series C preferred stock for $5 million in July 1995.

The S-1 does not name the third party. But the parties to the investors' rights agreement attached to the same S-1 include "Sega Enterprises, Ltd." (in Ohta-ku, Tokyo; "Attn: Mr. Tsuneo Naito, Managing Director"), and the old official timeline reads, "July, 1995: SEGA partnership is established." Huang himself has said, "We won a contract with Sega to build their game console." From these three points, the third party can be strongly presumed to be Sega.

As the S-1 tells it, the purchase of $5 million in stock was part of the same transaction as the NV2 development agreement. It did not begin as a rescue investment made after the company had started to sink. The S-1 separately records that the third party made "nonrefundable payments," which NVIDIA recorded as reductions to research and development expense: $2 million in 1995 and $3 million in 1996. And it concludes, "The third party revised its product development plans, and the Company terminated the development of this particular technology in 1996."

Huang's "pay us in whole" can also be read as referring not to the investment but to the payments under this agreement. The July 1995 investment ($5 million) and the agreement's payments in 1995–1996 ($5 million in all) may have run together in the later telling.

On the point that the relationship with Sega kept the company alive, Huang's account has been consistent. On July 15, 2026, Huang met Irimajiri again at the former Sega Akihabara arcade (now GiGO Akihabara 3) in Tokyo's Akihabara district, and the two companies announced a new partnership. There, Huang said, "Sega, Mr. Irimajiri, Mr. Suzuki, your friendship, support, and belief in us mean a great deal to me."


A 33-company scrum and a tape-out without a prototype

As of December 31, 1996, NVIDIA had 42 employees. The hosts of the podcast Acquired have said that the company laid off 70% of its staff during this period. This is the time Huang called "six months to live" in his National Taiwan University speech.

NVIDIA moved away from the curved-surface approach and switched to developing a chip that was based on polygons (triangles) and supported DirectX. It also chose an unusual way of getting there. As Huang explained on Acquired, the normal process (build a chip, write the software, fix the bugs and build it again) would not have been fast enough. "If we only had six months and you get to tape out just one time, then obviously you're going to tape out a perfect chip." So the company bought an emulator from a firm called ICOS, on the very day ICOS was shutting down for lack of customers, and ran its entire software stack on it before making real silicon.

According to Huang, the software team sat in the lab waiting for Windows to paint the screen. Drawing a single frame took about an hour, Huang recalled. "Everything that we could have tested, we tested in advance."

The result was the "RIVA 128," announced in April 1997, with commercial shipments beginning that August. Huang himself admits that not everything went right. Asked on Acquired about supporting only 8 of DirectX's 32 blend modes, he answered, "We should have implemented all 32."

The PC 3D chip market of the day was crowded. The chip-industry newsletter Microprocessor Report (June 23, 1997, issue) listed the companies that had announced 3D chips for PCs or planned to develop them, and counted "33 companies." In that list, alongside 3dfx, ATI, S3, Matrox and Intel, NVIDIA was a single line: "Nvidia/SGS-Thomson (RIVA 128)." The newsletter wrote that the cost of developing a new 3D product was approaching the cost of developing a high-end RISC CPU, or $10 million to $30 million.

The RIVA 128 sold. Huang later called it "a raging success." Revenue grew more than sevenfold, from $3.91 million in 1996 to $29.07 million in 1997. According to the S-1, nearly all of the company's product revenue in 1997 came from a single product, the RIVA 128.

Commercial shipments of its successor, the "RIVA TNT," began in July 1998, and by the time of the prospectus, the company's main manufacturing partner was Taiwan's TSMC. The prospectus says that its customers, the PC makers, brought out new configurations about twice a year, in step with spring and fall design cycles. Follow the product dates listed in the old official timeline, and from around this time NVIDIA was releasing new products every six to nine months.

The road to the listing was not smooth. NVIDIA first filed its S-1 on March 6, 1998. According to the prospectus, Intel began shipping its "i740" 3D graphics chip that same March. In the spring of 1998, demand fell at many PC makers, and the drop in orders, together with lower yields on the RIVA 128ZX and RIVA TNT, pulled down results for the quarter ended July 26, 1998. The company posted a loss for that quarter.

It took about 10 months from the first S-1 filing to the listing. At GTC in 2018, Huang described this period: "We went from days away from going public to days away from going out of business."

In January 1999, NVIDIA issued its prospectus dated January 21 and listed on NASDAQ (ticker: NVDA). The offering price was $12.00 a share; it sold 3.5 million shares, for a total of $42 million.


Declaring the "GPU" and losing big customers

The listing prospectus already carried one warning. Its rival 3dfx had agreed to acquire STB Systems, a board maker and one of NVIDIA's largest customers. The prospectus expected that sales to STB would fall significantly and that STB might no longer be a significant customer.

STB and Diamond Multimedia were NVIDIA's two largest customers, accounting for 63% and 31% of its revenue, respectively, in 1997. 3dfx completed its acquisition of STB in May 1999, and in October of that year, the rival S3 completed its acquisition of Diamond. According to the following year's 10-K, sales to Diamond fell from 24% of total revenue in the second quarter of fiscal 2000 to just 3% in the fourth quarter. The prospectus had listed concentration on a small number of customers as one of its main risks.

On August 31 of that year, 1999, NVIDIA announced the "GeForce 256."

NVIDIA itself called the chip a "GPU (Graphics Processing Unit)." It put the circuits for coordinate transformation and lighting calculations (hardware T&L) on a single chip together with the rendering circuits; it was made on a 0.22-micron process and had 23 million transistors. In its annual report (Form 10-K) filed in March 2000, the company wrote, "We are recognized for developing the world's first GPU."

Some coverage at the time was cool toward the label. The British technology news site The Register wrote in an article on August 31, 1999:

"Nvidia stressed that the chip is what it calls a 'graphics processing unit (GPU)'. That's a term that can be applied to almost every 3D graphics accelerator punched out since the Voodoo 1."

The article pointed out that S3's "Savage 2000," announced the previous day, also offered the same features to a greater or lesser extent. It went on to say that the GPU label was meant to show NVIDIA as a supplier of chips, not boards. 3dfx had bought STB, and S3 was in the process of buying Diamond; both had moved over to making boards themselves. In the article's reading, NVIDIA was trying to signal that, for the time being, it would not compete with the board makers.

3dfx, which had sued NVIDIA for patent infringement on September 21, 1998, agreed on December 15, 2000, to sell the assets of its graphics business to NVIDIA. The cash consideration was $70 million ($15 million of which had already been lent to 3dfx at signing), and NVIDIA was to deliver 1 million shares of its stock, subject to conditions, once 3dfx's liquidation was complete. The acquisition closed on April 18, 2001, and the patent suit was dismissed on the 26th of the same month.


The Xbox strings, an SEC cease-and-desist order and the fall to "21%"

On March 5, 2000, NVIDIA signed an agreement with Microsoft to develop and sell chips for Microsoft's "Xbox" game console and to license technology. In April 2000, Microsoft paid $200 million as an advance against chip purchases and other items. The agreement had one more provision: if anyone offered to buy 30% or more of NVIDIA's outstanding shares, Microsoft would have "first and last rights of refusal" to buy the stock. The 10-K says this provision could delay or prevent a change in control of the company.

The two companies had been in discussions over the pricing and volumes of the Xbox chips, but on April 23, 2002, Microsoft submitted the matter to binding arbitration. Microsoft asked that NVIDIA supply chips "in whatever quantities are ordered by Microsoft," alleged violations of the agreement and sought damages, and also sought lower prices. NVIDIA countered, "We do not believe we have a requirements obligation to Microsoft." It was on February 6, 2003, that the two companies settled the pricing issue and agreed to cooperate on future cost reductions.

On April 29, 2002, six days after the arbitration filing, NVIDIA announced that it would restate its results for fiscal 2000, fiscal 2001 and the first three quarters of fiscal 2002. According to the announcement, the restatement was the result of an investigation led by the independent audit committee, an investigation that had begun at the request of the SEC staff.

On September 11, 2003, the SEC issued a cease-and-desist order against NVIDIA (an order to stop a violation and not to repeat it; Release No. 34-48480). According to the order's findings, in the quarter ended April 30, 2000, NVIDIA agreed with a supplier to have its costs reduced by $3.3 million, in exchange for paying the amount back through higher prices on future products. According to the order, the arrangement was split into two separate agreements, and neither agreement referred to the other. In the quarterly review, NVIDIA showed its auditors the agreement supporting the cost reduction and did not show them the agreement on the price increases.

The order quotes an email from the chief financial officer (CFO) at the time (the order does not give a name).

"[We] can not sign this or have this in print. Will wipe out the credit in Q1."

According to the order, as a result, the quarter's gross profit was overstated by 6.4%, net income by 15.3% and diluted earnings per share by 14.8%. NVIDIA consented to the order without admitting or denying its findings. The only thing the order imposed was the cease-and-desist itself.

Later, on August 10, 2006, the company announced that its audit committee was reviewing its practices for granting stock options. The audit committee concluded that incorrect measurement dates had been used for grants in certain past periods, and the company recorded additional stock-based compensation expense.

On the product side, too, the stumbles continued. In November 2002, NVIDIA introduced the "GeForce FX 5800 Ultra," its first GPU made on a 0.13-micron process. A 10-K of around the same time wrote that other companies in the industry had suffered reduced yields and delayed shipments in migrating to new manufacturing processes, and admitted, "We have experienced and may continue to experience similar difficulties."

Meanwhile, Canada's ATI Technologies, starting with the "RADEON 9700," used the name "VPU (Visual Processing Unit)" for its chips, and claimed to be the first company to introduce processors compatible with Microsoft's DirectX 9.0. According to The Register on May 24, 2003, the benchmark software company Futuremark said that NVIDIA's drivers contained eight instances of "cheating" that raised scores by as much as 24.1%. In early June, Futuremark changed its wording, saying that "Nvidia's driver design is an application specific optimization and not a cheat," but did not restore the original scores. An NVIDIA spokesman said, "There are application-specific optimizations in every CPU and every GPU." On August 14, 2003, Microsoft announced that it had signed an agreement with one of NVIDIA's competitors to develop technology for the next Xbox (the partner was ATI).

The slump shows in the numbers as well. NVIDIA's own 10-K for fiscal 2006, citing estimates by the research firm Mercury Research, shows that the company's share of what it calls the "Performance DirectX 9.0-compatible graphics" segment was 21% in the fourth quarter of 2003. Gross margin also fell, from 38% in fiscal 2002 to 29% in fiscal 2004. The 10-K attributes this mainly to the GeForce FX's gross margins being lower than those of earlier GeForce products.

The comeback was quick. According to the same 10-K, the launch of the GeForce 6800 and 6600 brought the company's share of the same segment back up to 67% in the fourth quarter of 2004. It rose to 79% with the GeForce 7800, released in June 2005, and to 85% with the GeForce 7900, released in June 2006 (all of these figures are for this segment alone).


A Stanford paper and congratulations for a rival

While NVIDIA was winning back the graphics fight, a different experiment was under way in university labs.

The groundwork for it was the "GeForce3," which NVIDIA announced in February 2001. The chip let developers rewrite, by programming, the circuits that calculate the colors of vertices and pixels (shaders), and NVIDIA's old timeline calls it "the industry's first programmable GPU."

In 2004, at SIGGRAPH, the international conference on computer graphics, Ian Buck of Stanford University and his colleagues presented a paper: "Brook for GPUs: Stream Computing on Graphics Hardware." The work added simple data-parallel constructs to the C language so that the GPU could be used as a stream-processing coprocessor, and across five applications it showed speeds of up to seven times those of running on a CPU.

The paper contains a sentence that is surprising in light of what later became common knowledge. What it compared were ATI's "Radeon X800 XT Platinum" and a pre-release NVIDIA "GeForce 6800."

"We observe that, for these applications, ATI generally performs better than NVIDIA."

The paper speculated that the reason might be ATI's higher floating-point texture bandwidth, recording 1.2 billion floats per second for NVIDIA against 4.5 billion per second for ATI. The paper's acknowledgments list ATI, IBM, NVIDIA and Sony, in addition to DARPA.

According to the technology site bit-tech on October 3, 2006, ATI announced its commitment to general-purpose GPU computing under the name "Stream Computing," and Stanford University, partnering with ATI, released a new GPU-enabled version of the distributed computing client "Folding@home" (which analyzes how proteins fold). The article reported that Stanford's tests showed 15 times the performance of a CPU running alone, and wrote that teams using ATI cards should be able to finish their work faster. In general-purpose GPU computing in this period, ATI was ahead in some respects.

On July 24, 2006, the CPU maker AMD announced that it had agreed to acquire ATI Technologies. According to AMD, the deal was valued at $20.47 per share to ATI shareholders, or about $5.4 billion in total.

At the time, AMD was also a partner of NVIDIA's. According to NVIDIA's 10-K for fiscal 2007, the company was the largest supplier of AMD64 chipsets, with a 53% share. The same 10-K wrote that before acquiring ATI, AMD had not sold its own GPUs but relied on companies such as ATI or NVIDIA, and noted that if AMD and Intel pushed ahead with platform products, NVIDIA might be unable to sell GPUs to either company. Among the aims of the acquisition that AMD listed in its investor materials was the development of chips integrating the CPU and the GPU. In the October announcement that the deal had closed, this plan was presented under the name "Fusion."

On October 25, 2006, AMD announced the completion of its acquisition of ATI. Among the comments from business partners included in the announcement was one from Dan Vivoli, executive vice president of NVIDIA.

"NVIDIA and AMD have innovated together for over a decade. As a result, NVIDIA GeForce and nForce processors set the standard for AMD systems. AMD's open platform strategy enables continued partnership aimed at providing customers with a wide choice of industry leading solutions."

It was a cordial greeting.

The next month, November 2006, along with its new "GeForce 8" GPU, NVIDIA began rolling out a technology to the industry. In the 10-K's words, it was "a technology and compiler that, for the first time, allows for programmers to write C language applications for GPUs." Its name was "CUDA."

(Continued in Part 2: "The House That GeForce Built")

References