Disenchanting Jensen Huang
In 1984, when Jensen Huang graduated from Oregon State University, the global semiconductor industry was undergoing its first round of reshuffle. If he had chosen to stay in Taiwan, China, or go to an economy lacking in venture capital, with an imperfect limited - liability system and a lack of tolerance for failure, would NVIDIA have been born?
There are numerous counter - examples in history. The Soviet Union once had the world's top mathematicians and physicists. However, under the planned - economy system, the ideas of these geniuses either remained in academic papers or were directed towards military use, and were never transformed into innovative products in the civilian market. In the 1960s, Soviet scientists proposed the "OGAS" national computer network plan similar to the Internet, but due to the rigid system, it ended up stillborn. In the field of graphics computing, the Soviet Union also had outstanding engineers, but what they faced was not the uncertainty of the market, but the approval forms of the planning committee - there was never an option of "unknown" on those forms.
Another thought - provoking example is Brazil. In the 1970s, the military government implemented "market protectionism" and tried to cultivate the domestic independent computer industry by restricting imports. Brazilian engineers did develop domestically - designed chips and computers. However, due to long - term isolation from the global market, technological iteration stagnated. When the market opened in the 1990s, the Brazilian domestic computer industry almost collapsed instantly. The intelligence of those engineers was locked within the national borders.
A more recent typical case is India. Before the 1990s, India implemented a strict licensing system. At that time, India was not short of smart engineers - the graduates of the Indian Institutes of Technology were all over the world. However, most of the outstanding ones among them chose to go to the United States rather than start businesses locally. Why? Because staying in India meant facing cumbersome approvals, an inefficient bureaucracy, and a lack of venture capital. Even the most brilliant ideas could only be shelved.
In the wasteland of graphics computing, countless equally talented and passionate figures have fallen. NVIDIA was once one of them, but it finally made it through.
In the spring of 1996, NVIDIA was on the verge of bankruptcy, with only enough cash flow to last for 30 days. Facing a seemingly hopeless situation, Jensen Huang made a decision that violated all common sense in business textbooks - how this decision became possible is the institutional code to be dissected later.
However, simply emphasizing the entrepreneur's courage in the face of uncertainty will make us miss a more fundamental question: Why was Jensen Huang brave enough to make such a decision in 1996? If failure meant losing everything and having no chance of recovery, would he still choose to "take a desperate gamble"?
These questions shift our focus from individual heroes to the institutional soil.
If NVIDIA's story is only told as the struggle history of a genius entrepreneur, it obscures the more fundamental things: the spatial agglomeration in Silicon Valley, the institutional arrangement of limited liability, the normalization of in - house R & D in enterprises, and the tolerance of the capital market for long - term investment. These impersonal forces together constitute the infrastructure for innovation to emerge. Jensen Huang's luck lies not in having extraordinary foresight, but in being in an institutional environment where "luck" becomes a high - probability event.
The Gift of Uncertainty and the Premise of Institutions
In 1993, when Jensen Huang and his two co - founders were making repeated calculations in Denny's, there were at least 35 companies fiercely competing in the PC graphics accelerator market. Industry experts warned that "the space was too crowded" and advised against entering. Jensen Huang had no consulting reports, no focus groups, and no historical data to tell him which variable was correct. He just repeatedly adjusted the variables in the spreadsheet and watched the numbers fluctuate around $50 million. This was the threshold he set for himself.
When the company was on the verge of bankruptcy in 1996, he made another decision that was incomprehensible under traditional business logic. At that time, NVIDIA was unable to afford the cost of a new round of tape - outs, but Jensen Huang used the $1 million obtained from Sega to buy a hardware emulator, skipping the physical prototype stage that all chip companies regarded as a necessary step. The engineers verified the design on the emulator at a speed of one frame per second - a normal game requires 30 frames per second. Jensen Huang later said, "It was a 50 - 50 gamble, but we were going to go bankrupt anyway."
When the CUDA project was launched in 2006, the same logic played out again. John Nichols told Jensen Huang based on physical principles that the performance improvement speed of traditional processors would slow down significantly, and parallel computing was the only way out. However, there was no business case at that time to prove this judgment. Jensen Huang's decision - making basis was not how much revenue CUDA could bring that year, but "if this is the right direction, if we don't do it now, it will be too late in the future."
The common feature of these three groups of decisions is that when making choices, there was no historical data or probability distribution for reference. This state is exactly the "uncertainty" strictly distinguished by economist Frank Knight in 1921 - it is different from the "risk" that can be calculated and insured. Risk can be quantified, dispersed, and hedged; uncertainty means that it is impossible to calculate and there is no historical sample for reference. What entrepreneurs face is this pure unknown. In Knight's framework, the function of entrepreneurs is not to calculate risks, but to make "judgments" in the fog of the unknown. And profit is the reward for this kind of judgment that bears uncertainty.
However, without the corresponding institutional support, this kind of "judgment" may turn into a suicidal adventure. The critical moment in 1996 was a natural laboratory to test how institutions work.
The Spatial Carrier and Institutional Root of Creative Recombination
What is the realization mechanism of entrepreneurship? Joseph Schumpeter gave a classic answer in 1911: Innovation, that is, "the recombination of production factors." The premise of recombination is that the factors must be accessible. If entrepreneurs cannot access new technological ideas or find knowledgeable collaborators, recombination cannot occur.
In the 1980s, Paul Romer integrated this idea into the endogenous growth theory in macroeconomics. He pointed out that the long - term driving force of economic growth comes from "the spillover of ideas" - different ideas of different people interact, collide, and recombine with each other. Different from traditional factors such as land and capital, ideas are non - competitive: one person's use of an idea does not prevent others from using it at the same time. This means that ideas can be continuously accumulated and grow exponentially. The most important factor determining growth is the institutional arrangement that can promote the encounter and recombination of ideas.
This leads to the spatial dimension. The collision of ideas requires conditions: reducing the cost of information exchange and increasing the probability of accidental encounters. This is exactly the view argued by Jane Jacobs in The Death and Life of Great American Cities: The core function of cities is to catalyze "new jobs growing out of old jobs" through a high - density and diverse spatial structure. The essence of urban agglomeration is to make the search cost and matching cost of ideas approach zero.
In 1993, NVIDIA's choice to locate its headquarters in Silicon Valley rather than Austin or Portland was not accidental. Silicon Valley gathered the world's densest chip designers, software engineers, venture capitalists, and game developers at that time. Jensen Huang's business partners were from Sun Microsystems, only a few miles away; the first investment came from Sequoia Capital, and the partners were just in the next block; most of the early engineers "jumped ship" from nearby SGI and AMD.
The direct effect of this agglomeration is the low - cost and high - frequency flow of information. In the late 1990s, John Carmack, a programmer at id Software, was looking for a smoother graphics rendering solution for Quake II. He was not satisfied with the fixed - function pipeline of the graphics cards at that time and complained in public on many occasions that the existing hardware could not fully realize the potential of his algorithms. The Jensen Huang team noticed these needs, actively studied Carmack's code, and even invited him to NVIDIA for communication. This high - frequency interaction promoted the birth of the Riva TNT in 1998 - it realized the "dual - texture pixel pipeline" and performed well in games such as Quake II, laying the foundation for NVIDIA to gain a foothold in the 3D graphics market. This is a typical Schumpeterian recombination: software ideas and hardware ideas were quickly matched in a low - friction environment.
An earlier foreshadowing occurred around 2003. Ian Buck, a doctoral student at Stanford University, developed the Brook stream programming language, which for the first time systematically proved that GPUs could be used for general - purpose computing, not just for graphics rendering. His research attracted NVIDIA's attention. After graduating with a doctorate in 2004, Buck was recruited by Jensen Huang into NVIDIA and became the core promoter of the CUDA project. There was no thick confidentiality agreement or long - term due diligence between the two, only a common enthusiasm for technological possibilities.
In 2012, Alex Krizhevsky from the University of Toronto trained AlexNet in his bedroom using two GeForce GTX 580 graphics cards, detonating the AI revolution. Krizhevsky was just a doctoral student at that time. He bought these two graphics cards with his credit card and built a deep - learning system in his bedroom. He had never met Jensen Huang, but the CUDA framework he used, the papers he referred to, and the conferences he submitted to all relied on the dense connection between Silicon Valley and the global scientific research network.
These cases reveal a truth: Jensen Huang's so - called "technological intuition" is actually a conditioned reflex under the impact of high - frequency information flow.
Why does this high - frequency interaction only occur in Silicon Valley and not on Route 128 in Boston, which also has a large number of talented people? The answer lies not in culture, but in law. Section 16600 of the California Business and Professions Code clearly states that, with very few exceptions, any form of non - compete agreement is invalid. In other states such as Massachusetts, employees are often locked by non - compete agreements after leaving their jobs, and talents and ideas are confined within a single enterprise. In California, engineers can change jobs without legal obstacles, and ideas are recombined across society as people move. The high mobility in Silicon Valley is not a "natural trait" but the result of legal discipline.
The Institutional Game of Limited Liability and the Dual Faces of the Capital Market
In the spring of 1996, there was only 30 days' worth of cash left on NVIDIA's books. Jensen Huang did something rare in business history: he told Sega that NVIDIA would stop the NV2 chip project for custom - made chips for Sega and switch to Microsoft's DirectX standard. Jensen Huang proposed that although the cooperation was terminated, as long as NVIDIA could manufacture an NV2 prototype chip that met the specifications, Sega would need to pay $1 million.
Finally, Sega paid the $1 million, and Shoichiro Irimajiri, the then - president of Sega, also made an additional investment of millions of dollars.
This scenario is often told as a victory of Jensen Huang's personal charm or a miracle of trust between business partners. However, from the perspective of institutional economics, it reveals a deeper logic: how the limited - liability system changes the incentive structure of both parties in the game, making seemingly irrational choices a rational equilibrium.
Let's analyze Sega's decision - making. If it didn't pay, NVIDIA would surely go bankrupt. As a creditor, Sega could claim its rights in the bankruptcy liquidation, but the limited - liability system had limited NVIDIA's debt liability to the company's assets. For a startup on the verge of bankruptcy with only 30 days' worth of cash on the books, how many liquidable assets could be left? Even if Sega filed a lawsuit, it could not claim Jensen Huang's personal property - that was the bottom line of the limited - liability system. The expected return of not paying was close to zero. If it paid, Sega's upper limit of loss was locked at $1 million. And the return came from an implicit call option: if NVIDIA survived, there might be continued cooperation in the future; if NVIDIA switched to the Microsoft standard, it might become an important player in the new ecosystem, and Sega, as an early supporter, could share this dividend. The expected return of paying included at least a positive probability.
Limited liability played a dual role here. For entrepreneurs, it isolated personal assets from corporate risks, allowing Jensen Huang to be honest in a desperate situation rather than hiding the truth and delaying time - even if he failed, he would not be out of the game for life. For capital providers, it capped the risk loss, changing Sega's calculation from "whether it would lose all its money" to "whether it was willing to buy a call option." It was this two - way incentive that made cooperation in a desperate situation possible.
This is not a miracle unique to Silicon Valley. Two centuries ago, when James Watt improved the steam engine, Matthew Boulton's capital support was possible because the partnership system at that time had begun to explore a similar risk - isolation mechanism. The mystery of the modern corporate system is not to make failure cost - free, but to make failure priceable, dispersible, and recoverable.
In NVIDIA's growth history, limited liability is just the first line of defense. Behind it is a more complex venture - capital system. In 1993, the startup capital came from Sequoia Capital and Sutter Hill. When Don Valentine of Sequoia Capital decided to invest, he was not satisfied with Jensen Huang's business plan. What he valued was the strong recommendation of his old friend Wilf Corrigan, who was Jensen Huang's former supervisor, and the ecosystem they were in: there were enough engineers to recruit in Silicon Valley, enough game companies to cooperate with, and enough potential customers to validate the products.
The existence of venture capital itself is an institutionalized fault - tolerance mechanism. Traditional bank credit requires collateral and predictable cash flow - these do not exist in startups. Venture capital, on the other hand, aggregates funds from long - term investors such as pension funds through a limited - partnership structure and entrusts professional managers to allocate them to high - uncertainty areas. This system transforms the "uncertainty" in the Knightian sense into a manageable investment portfolio: the failure probability of a single enterprise is extremely high, but through diversified investment, failures can be digested at the system level, and the losses can be covered by the ultimately successful enterprises.
However, venture capital requires a low - cost exit channel, which leads to another function of the capital market: pricing uncertainty and providing patient capital for long - term investment.
In 2006, when Jensen Huang launched the CUDA project, he faced a typical "patient - capital" problem. From 2006 to 2017, NVIDIA invested approximately $10 billion in CUDA in total, and there was almost no significant revenue generated in the long 11 - year period. The doubts on Wall Street never stopped.
Jensen Huang did not choose to engage in a war of words. Instead, he flew to Boston and New York to visit institutional investors one by one. He explained in the meeting rooms again and again why CUDA would ultimately succeed and why giving up at that time would be equivalent to suicide. Finally, he won enough support, and the CUDA project was able to continue. Those investors who shook their heads and left at that time could only remain silent in the face of the trillion - dollar market value later.
This detail reveals the internal tension between the capital market and long - term corporate innovation. The capital market allows enterprises to exchange today's losses for future growth, but this trust has a time limit. What really gave him the confidence to persevere was the strong cash flow from NVIDIA's game business. The monopoly position of GeForce graphics cards provided an "internal patient - capital" source for CUDA for more than 10 years.
In 2016, Google made a one - time purchase of more than 40,000 NVIDIA GPUs. This was the first real large - scale order after 10 years of the launch of CUDA. Since 2017, the data - center business has entered a rapid growth channel. By 2023, when ChatGPT triggered the AI boom, almost all mainstream AI models were running on NVIDIA chips.
Looking back, the success of CUDA is the accumulation of a series of institutional arrangements: Limited liability allowed Jensen Huang to place early bets; venture capital provided the fuel for trial - and - error; the cash flow from the game business provided internal patient capital; and the capital market finally priced the long - term investment.
The Institutionalization of Innovation - From Individual Heroes to Organizational Machines
In 1942, Joseph Schumpeter proposed that the engine of capitalism was shifting from "individual entrepreneurs" to "the R & D departments of large enterprises." He called this "the institutionalization of creative destruction": innovation was no longer an external accidental shock but became a continuous activity within the enterprise.
The CUDA project launched in 2006 is an epitome of this institutional evolution.
The team led by John Nichols continuously optimized the CUDA compiler, libraries, and development tools. Brian Catanzaro started developing cuDNN in 2013, and his performance evaluation was not good at that time. However, Jensen Huang told him that this was the most important project in NVIDIA's 20 - year history,