"I would never want to found NVIDIA a second time!" Jensen Huang personally recounts the price behind his 33 years of entrepreneurship.
33 years ago, Jensen Huang, then 30, co-founded NVIDIA with two partners. 33 years later, he left the company he built from scratch. Just a few days after his resignation, voices in Silicon Valley began to call for him to return to NVIDIA.
On November 12, NVIDIA announced that Jensen Huang had stepped down from his positions as CEO, President and member of the board of directors. Since then, an online petition in support of his return has collected more than 170,000 signatures, with tech industry figures including Elon Musk publicly voicing their support. According to a Financial Times report on November 15, Jensen Huang has discussed the possibility of returning to the CEO position of NVIDIA with his close confidants, and some Silicon Valley executives are also planning to put pressure on the board of directors.
But just half a year ago, when someone asked Jensen Huang if he would still found NVIDIA if he could go back to the age of 30, his answer was: "Absolutely not."
In April this year, Jensen Huang gave this unexpected answer in an interview with Guy Raz, host of How I Built This. The program was originally released in audio format in May. On November 12, the day Jensen Huang announced his departure, the program re-launched the 87-minute full video of the interview. In the new introduction, Guy Raz stated that the content of this conversation did not involve Jensen Huang's recent turmoil, but hoped it would allow the public to see a more complete side of Jensen Huang.
During the interview, Jensen Huang recalled the days when NVIDIA was on the verge of collapse: the first generation of chips failed due to a wrong technical route, so he had to fly to Japan to admit to SEGA that the company could not fulfill the contract, and asked the other party to convert the remaining $5 million contract payment into an investment; to bet on CUDA, NVIDIA invested huge sums of money for many years, but never saw clear commercial returns. He also talked about those irreplaceable family moments, and why he has spent almost his entire life trying to forget what happened the day before.
Now, this old interview has taken on new meaning because of Jensen Huang's sudden departure. A man who once devoted 33 years of his life to NVIDIA and openly said he would never want to go through the entrepreneurial process again is now standing at the crossroads of whether to return to the company.
1
When he founded NVIDIA at the age of 30, he never thought that the startup would fail
In 1993, Jensen Huang decided to leave LSI Logic and co-founded NVIDIA with Chris Malachowsky and Curtis Priem. That year, he was about to turn 30, already married with two young children, and had a well-paid, stable job with promising prospects.
Before starting his own business, Jensen Huang's career development seemed quite smooth. After graduating from Oregon State University, he joined AMD and then LSI Logic, a company that mainly helped computer manufacturers design and produce custom chips, with clients including Silicon Valley enterprises such as Sun Microsystems and Silicon Graphics. Through cooperation with these companies, he got in touch with many outstanding computer architects, and gradually realized that technological leadership does not necessarily mean commercial success, and strategic choices are equally important. He later described this professional experience as a series of good luck.
It was Chris and Curtis, whom he met at Sun Microsystems, who first came up with the idea of starting a business. The two wanted to develop graphics chips specifically for computer games, but failed to get support from their company, so they decided to start their own business and invited Jensen Huang to manage the new company. At first, Jensen Huang had no intention of leaving LSI Logic. But Chris and Curtis never gave up inviting him. The three repeatedly discussed the market opportunities of computer graphics technology, and Jensen Huang was gradually attracted to the idea.
What they were really optimistic about was the personal computer revolution.
At that time, the computer graphics effects shown in the movie Jurassic Park impressed many people, but generating these images required expensive professional computers. Jensen Huang recalled that a computer that could produce similar images at that time might cost 1 million US dollars, and what they wanted to do was to allow ordinary consumers to obtain similar 3D graphics processing capabilities for a few hundred dollars. If this technology could be integrated into the personal computer architecture, a brand new market could be created.
This was not an opportunity only they had discovered. The host mentioned in the interview that about 60 companies were trying similar things at that time. Jensen Huang had a stable job and children to raise, and faced a highly competitive market, but these factors did not seem to make him hesitate much.
The host asked him if he had ever worried about the failure of the startup.
"No." Jensen Huang replied. A moment later, he smiled and admitted: "Actually, I should have thought about it back then."
In Jensen Huang's memory, NVIDIA had already become very clear in his mind at that time. When an idea is repeatedly thought about, imagined, and even fascinating, it will gradually become as real as reality. He believed that he and his two partners could build a successful company, and the worst outcome was nothing more than finding a new job after failure.
"The thought that the startup would fail never crossed my mind, not even once." Only later did he realize what the seemingly uncomplicated decision back then really meant.
2
Almost all of the first batch of chips were returned, and NVIDIA was once on the verge of bankruptcy
The first major setback after NVIDIA's founding came from the technical solution the team initially chose.
The first chip, NV1, almost pushed the newly established company to bankruptcy. To bring expensive professional 3D graphics processing capabilities into the consumer market, NVIDIA redesigned a set of graphics processing algorithms. After more than two years of R&D, NV1 was finally launched, and about 250,000 chips were sold initially. However, almost all of these products were returned in the end.
Jensen Huang later admitted that they chose the wrong technical route from the very beginning. NV1 adopted Forward Texture Mapping, while Inverse Texture Mapping gradually became the mainstream at that time. This meant that even if NV1 itself could complete graphics processing, it was difficult to be compatible with the mainstream software interface that the industry was forming. As Microsoft prepared to launch DirectX in Windows 95, NVIDIA found that its products were incompatible with this new graphics API, while other competitors were developing products along another route.
"We sold 250,000 units, but the algorithm didn't actually work. So later, all 250,000 products were returned." Jensen Huang recalled that this was not an ordinary product mistake, but a technical disaster.
The trouble was that NVIDIA had spent two and a half years investing in this architecture, and the subsequent NV2 and NV3 were also built on the same technical foundation. The company had previously signed a contract worth about 12 million US dollars with Japanese game manufacturer SEGA, planning to provide graphics chips for its game consoles. If it continued to fulfill the contract, NVIDIA could at least survive for a period of time, but that also meant it would have to spend about two more years on the technical architecture where problems had already been found. By then, dozens of competitors in the market might have been far ahead.
Divisions emerged within the company. Some advocated continuing to fulfill the contract to ensure the company's survival first; others believed that since the technical direction was known to be wrong, it should be abandoned immediately. Jensen Huang finally chose the latter. He flew to Japan, confessed to the person in charge of SEGA that the technical solution NVIDIA had promised was no longer suitable for SEGA, and suggested that the other party find a new partner.
But abandoning the contract also meant that NVIDIA would lose a sum of funds to maintain its operations. Without this money, the company might go bankrupt soon. Therefore, after explaining the situation to SEGA, Jensen Huang made another request: hoping that the other party would convert the remaining contract payment of about 5 million US dollars into an investment in NVIDIA, giving the company a chance to readjust its technical direction.
The person in charge of SEGA was skeptical. At that time, NVIDIA faced dozens of competitors, had decided to abandon the technical architecture it had invested more than two years in, and did not yet have a clear new business plan. Jensen Huang did not try to depict a definite future. He admitted that he did not know what the complete business plan would be next, but two things had to be done: keep the company alive, and get the technology back on the right track.
In the end, SEGA agreed to convert the remaining contract payment into an investment. This sum of about 5 million US dollars prevented NVIDIA from going bankrupt immediately, but the company's original business plan had to be adjusted accordingly. Previously, NVIDIA kept recruiting employees based on the expectation that NV1, NV2 and NV3 would succeed one after another, and the team size had expanded to about 250 people. Now that the original technical route was abandoned, the R&D and business plans formulated around these generations of products could hardly continue to be implemented, and the company's funds were very limited. To cut expenses and maintain operations, Jensen Huang had to carry out large-scale layoffs, and finally laid off about two-thirds of the employees. He admitted in the interview that it was an extremely difficult decision.
After the personnel adjustment, NVIDIA still had to solve more tricky technical problems. Although the team decided to abandon the original graphics processing architecture, all the R&D work in the previous two years was carried out around that route, and it was not easy to redesign a chip that met the mainstream technical standards in a short time. Jensen Huang recalled that he thought at that time, since dozens of other companies in the market could adopt the correct technical solution, why couldn't NVIDIA? So he went to an electronics store called Fry's Electronics and bought three technical books introducing OpenGL programming and graphics processing pipelines. These books were written by engineers from Silicon Graphics, explaining in detail how OpenGL works. He gave the books to three engineers respectively, asking them to use them as a reference to start designing new chips.
However, even if a new technical direction was found, there was not much time left for NVIDIA. At that time, it usually took one and a half to two years from the completion of chip design to the production of prototypes, through software testing, defect modification and final mass production. This process often required multiple design and manufacturing iterations, but NVIDIA only had enough funds to keep the company running for about six months. According to the traditional process, the company would run out of funds before the chips were manufactured.
Jensen Huang therefore began to look for ways to shorten the R&D cycle. He heard that a soon-to-be-bankrupt company had developed a hardware emulator that could simulate the running behavior of a chip before it was actually manufactured, allowing engineers to run software in advance and verify the design. This meant that NVIDIA had the opportunity to skip the time-consuming physical prototype verification part of traditional development, and find problems as much as possible before tape-out. But buying this device was not cheap. Jensen Huang eventually spent half of the company's remaining funds to buy this neglected machine from the creditors.
The company was almost out of money, but he further shortened the time the funds could last. The host asked him if anyone opposed such a risky decision. Jensen Huang said he did not remember anyone explicitly opposing it, nor did he remember how many people's opinions he had solicited.
"That was the only solution we could think of at the time."
In the end, NVIDIA used the emulator to advance the development and verification of RIVA 128, and the product reached a level sufficient to keep the company operating. Years later, when Jensen Huang recalled that experience, he still remembered that the company could go bankrupt at any time, but he did not remember the team discussing this every day. What he remembered was that the engineers were very calm and focused, solving the problems that had to be solved in front of them step by step.
The crisis finally passed, but NVIDIA did not enter a stage of stable development from then on. A few years later, Jensen Huang decided to invest a lot of money in developing a technology that could hardly generate revenue in the short term. This time, he could not set a clear six-month deadline for the team as he did when developing RIVA 128.
3
Betting on profits to promote CUDA: NVIDIA waited for many years before the AI turning point arrived
NVIDIA initially focused entirely on the video game market, but as GPU performance continued to improve, the team gradually realized that these chips could not only generate images, but also undertake other computing tasks.
The initial exploration still came from games. Jensen Huang found that for a virtual world to be truly immersive, exquisite pictures are far from enough. Water needs to flow, leaves need to swing with the wind, and explosions should also conform to the laws of physics. These effects involve calculations such as fluid dynamics and particle physics, and the GPU itself already has powerful parallel processing capabilities. From this, NVIDIA began to study whether the GPU could no longer be limited to graphics rendering, but become a more general-purpose computing device.
At the same time, universities and scientific research institutions also began to try new applications. Jensen Huang mentioned that doctors at Massachusetts General Hospital once used NVIDIA's graphics chips for CT image reconstruction, and researchers at Stanford University were exploring other parallel computing uses of GPUs. These efforts showed him a possibility: tasks that could only be completed by expensive supercomputers in the past might be run on the GPU of an ordinary computer.
Around 2006, NVIDIA launched CUDA, enabling developers to use GPUs for general-purpose parallel computing. But after the technology was released, the market did not respond immediately.
Jensen Huang found that promoting a new computing architecture faces the typical "chicken or egg first" problem. Developers are unwilling to write software for platforms with a limited number of users, and consumers will not buy hardware for platforms that lack applications. Without a large enough installed base of devices, it is difficult to attract developers; without developers, there will be no killer applications that can prove the value of the technology. Even with its potential, CUDA could be trapped in this cycle for a long time.
NVIDIA finally chose a costly route: integrating CUDA into GeForce graphics cards for the game market, and using the shipment volume of game graphics cards to bring general-purpose parallel computing capabilities to as many personal computers as possible.
At that time, the vast majority of consumers bought GeForce just to play games, and did not need the general-purpose computing capabilities provided by CUDA. But NVIDIA still had to bear the corresponding costs of development, hardware support and ecological construction, and could not charge extra fees from gamers for this. Jensen Huang recalled that this further deteriorated the already pressured gross profit margin.
However, only when enough computers have CUDA capabilities, will developers have a reason to spend time learning it and writing software for it. So NVIDIA decided to bear the cost first and build up the market.
The company then launched a promotion program called "CUDA Everywhere". Employees traveled to universities, research institutions and enterprises around the world to introduce CUDA to scientists and software developers, and help them try to migrate their original programs to GPUs. From seismic data processing and medical image reconstruction to molecular dynamics and other scientific computing, NVIDIA kept looking for applications that could take advantage of parallel computing capabilities. Jensen Huang himself often participated in these promotion activities.
These applications gradually proved the technical value of CUDA, but for a long time could not support its commercial investment. Most of the early users came from scientific research institutions, with limited procurement volume, and it was still the game graphics card business that really maintained the company's operations.
From 2006 to 2012, NVIDIA experienced a stock price decline and a long-term slump. The company also incurred hundreds of millions of dollars in related expenses due to processor product issues. As the CEO of a listed company, Jensen Huang needed to face investors' doubts and had to explain the company's operating conditions on quarterly earnings calls.
Jensen Huang did not know when it would achieve commercial success. He recalled that at that time, he often felt that things would get better in another year. After the second year passed, he told himself that maybe things would change next year. This state lasted for many years.
The host asked him if he had ever worried that NVIDIA might enter this field too early, and eventually create a revolutionary technology without waiting for the day when it would really become popular.
"Of course I thought about it, and I thought about it often." Jensen Huang replied.
The team continued to contact image processing software developers, and also helped researchers migrate their programs to GPUs. The application scenarios of CUDA continued to expand, but no one knew which application would eventually bring a large enough market.
In 2012, an important breakthrough emerged in the field of deep learning. AlexNet achieved a significant leading performance in the ImageNet computer vision competition, and its training process used the parallel computing capabilities of the GPU. The GPU general-purpose computing ecosystem accumulated over many years finally encountered a new application direction that could bring large-scale demand.
Jensen Huang recalled that during that period, multiple research teams contacted NVIDIA, hoping to get help using CUDA to accelerate deep learning training. If traditional CPUs were used, such tasks might require a large number of processors; while GPUs have many computing units that can work at the same time, researchers can use a small number of GPUs to complete training at lower cost and faster speed.
Before these researchers contacted NVIDIA, Jensen Huang actually did not know much about deep learning. What really caught his attention was the huge performance improvement AlexNet achieved compared with previous algorithms.
He began to study why this technology worked, what other problems it could solve, and how it might change the future computer architecture. After repeated discussions with multiple teams, Jensen Huang gradually established AI as an important development direction for NVIDIA. Around 2013, Jensen Huang once sent an email to the whole company, clearly stating that NVIDIA was transforming into an AI company.
Looking back at this history, NVIDIA did not design CUDA from the very beginning for today's AI market. It