Tech

Jensen Huang: The Billionaire Behind Nvidia’s Rise as an AI Chip Powerhouse

On May 25, Nvidia (NVIDIA) released its financial report for the first quarter of fiscal year 2024 (Editor’s Note: Nvidia’s fiscal year is nearly a year earlier than the natural year), and achieved revenue of US$7.192 billion during the period, a year-on-year decrease of 13%. For other companies, this is definitely bad news that will cause investors to sell stocks, but on the day the financial report was released, Nvidia’s stock price soared by nearly 30%, and its market value increased by more than 200 billion US dollars.

The reason why the market is bullish on this stock is because the financial report shows that in the first three months of this year, Nvidia’s AI chip revenue increased by 14% year-on-year. During the earnings call, Chief Financial Officer Colette Kress revealed that expected revenue for the next quarter would be a record $11 billion, well above analysts’ forecast of $7.15 billion.

No one would question Kress’ high-profile prophecy. Everyone has seen that the “big language model” battle of global technology giants triggered by ChatGPT at the beginning of the year has made Nvidia, an “arms dealer”, a lot of money.

ChatGPT is a chat robot launched by OpenAI based on large-scale model technology. Bill Gates called it the most revolutionary technological advancement in the computer field since the graphical interface in 1980. For the first time, humans can ask questions to computers in natural language. The rhythm of “007” tirelessly completes all kinds of tedious desk work for us: writing manuscripts, analyzing reports, revising papers, creating codes, etc. Everyone can perceive what a huge opportunity this is.

Launching large models has become the top priority of every powerful technology company in the world, and they quickly formulate their own large model development strategies—open source, closed source, single-modal, multi-modal, and general-purpose ,vertical.

The “big” in the large model refers to the parameter scale of these models. Although scientists still debate whether the heap parameters can continue to improve the quality of the model, the parameters of the current mainstream commercial models are all above 100 billion. Training such a model corresponds to a huge amount of computation. As an AI start-up company fully supported by Microsoft, OpenAI has the highest priority support of Microsoft Azure cloud-about 25,000 Nvidia GPUs are supporting the training of GPT large models, which is currently one of the largest AI servers in the world. However, in early June this year, OpenAICEO Sam Altman (Sam Altman) revealed that due to the shortage of GPUs, GPT cannot guarantee to provide sufficiently stable services and more powerful functions.

The AI ​​industry dominated by large models has stimulated the blowout of the overall AI server market this year, and the GPU manufactured by Nvidia is the “heart” that supports the computing power of AI servers.

Only from the perspective of commercial realization, Nvidia can be said to be the “biggest beneficiary at this stage” of this global AI technology revolution. The industry’s extreme thirst for GPUs is being transformed into orders from Nvidia. If you can’t buy a single card, you can only buy a module, and if you can’t buy a module, you can only buy a whole server… But even if you are willing to pay a higher premium, there is still no stock. It is reported that the delivery date of modules like the Nvidia A800 is two months after the order is placed.

According to Trend Force’s estimates, Nvidia’s GPUs account for nearly 70% of the entire AI server market. The potential market is vast, and the products occupy an absolute dominant position. Nvidia ushered in the “Davis double-click” moment of performance and market value.

After the announcement of the first quarter financial report, Nvidia’s market value soon exceeded one trillion US dollars for the first time, ranking second only to Apple, Microsoft, Alphabet and Amazon in the US stock market value ranking. This company has also created some other records worth mentioning: it is the world’s first chip company with a market value of more than one trillion US dollars; He’s been in charge for 30 years—every time he’s out in the public eye, he always wears a cool black leather jacket, always looking energetic and ambitious.

Who is Jensen Huang

Huang Renxun was born in Taiwan, China in 1963. When he was 5 years old, his father transferred his family to Thailand because of his job. At the age of 9, he was sent to a relative’s home in Washington State, USA for foster care due to the political turmoil in Thailand. Soon he was sent to a rural boarding school in Kentucky. . There, Huang Renxun lived with students with various problems every day, and had to clean the toilets. After living like this for more than a year, his parents finally came to the United States to reunite with him.

The life experience of wandering around in childhood cultivated Huang Renxun’s ability to understand people. At the same time, he has been full of exploration and adventurous spirit since he was a child. When he was in primary school in Thailand, he once sprinkled combustibles on the swimming pool, ignited them, and plunged into the water to observe the scene of the flames burning on the water.

After graduating from university, Huang Renxun’s first job was engaged in chip design at AMD, but it didn’t take long for him to switch to another graphics chip company, LSI Logic. He stayed in this company for 8 years, did chip design, then did marketing and sales, and finally rose to a management position all the way. This early career experience has trained him to be a generalist.

In Silicon Valley, where entrepreneurial culture prevails, Jensen Huang, who is about to turn 30, also wants to start his own company. He had only one vague idea: to make a more powerful graphics chip. He asked his boss, Wilfred Corrigan, for help. Corrigan didn’t think Huang Renxun’s plan could be realized, but he still introduced his friend, the father of venture capital and the founder of Sequoia Capital, Don Valentine to him.

Huang Renxun tapped into the video game market for his own business plan—a graphics chip with stronger performance. But in the 1980s, not many consumers could afford to play games on a personal computer. Although the investor was not persuaded by him, he still gave him $2 million for Corrigan’s face, and claimed that “if you lose my money, I will kill you.”

In this way, Huang Renxun established Nvidia in 1993. He is a person who is good at creating concepts, and he claims that this is the only “consumer 3D graphics company” in the world at that time.

It makes sense to say this. After all, in the late 1980s, display chips mainly met the customized needs of military enterprises for aerospace simulation. However, as game consoles such as Nintendo SFC and Sega MD swept the American consumer market with bright colors and surprising pictures, more and more chip companies began to develop graphics cards for personal computers.

“Around 1995, there were 50 to 70 companies doing the same thing as us, and this number increased to 300 a few years later.” Huang Renxun once recalled. Facts have proved that his judgment on the market is correct. Games have indeed become the basic disk that Nvidia GPU sales have relied on for a long time. Until 2020, the game business accounted for more than 50% of Nvidia’s revenue. Sony and Blizzard are its important partners in the field of game processors.

From gaming to data center

In 1999, Nvidia invented a new hardware vocabulary called graphics processing unit (GPU).

In the era of PC computers, people are more familiar with another name for GPU: Graphics Card. It is not the computing and control core components of the computer. The central processing unit (CPU) really occupies the core position, and the GPU is responsible for processing some complex graphics display tasks assigned by the CPU. You can think of the CPU as a graduate student who can calculate calculus, and the GPU is equivalent to 100 high school students who can only add, subtract, multiply and divide.

Therefore, for a long time, although Nvidia has a good name in the circle of gamers, it is far from being on the list of major technology companies. Later, in the mobile Internet era, it was even directly squeezed out of the poker table by Qualcomm, ARM and other peers who are more focused on mobile chip development.

Fortunately, with the revolutionary upgrade of the GPU architecture in 2006, Nvidia spent more than ten years cultivating a “data center” outside of games as the company’s second growth curve. Since 2015, the data center business has gradually replaced games. The core position in the revenue structure. In fiscal year 2023, the data center business overtook games for the first time and became Nvidia’s most important source of revenue, accounting for 55.6%.

Nvidia has experienced two “peak” moments in the personal computer industry-in 2011, the annual shipment of PCs reached an all-time high of 364 million units, and in 2016, the annual shipment of smartphones reached an all-time high of 1.47 billion units. Then came the slow decline, and the wave of democratization of computing devices started by IBM’s first personal computer, the 5150, in 1981 was drawing to a close.

Today, although the Internet application layer innovation around personal computers has come to an end, computing may have just begun.

It is worth noting that from the 2016 fiscal year to the present, the compound annual growth rate of Nvidia’s game business is still 18%, indicating that the transfer of its business core is not due to the shrinking scale of traditional business, but the room for growth of new business is far beyond The value of a GPU as a graphics processor.

In fact, Huang Renxun has long seen that if it is only used as a graphics processor, the market demand for GPU will eventually be saturated. In his words, “there are only so many pixels on the screen.” Huang Renxun does not want Nvidia to end up being commoditized—GPU gradually loses its unique value in the full market competition, and eventually becomes a replaceable standardized product.

In a 40-year cycle, the development of computer hardware has once again returned to the era dominated by enterprise customers. In areas such as augmented reality (AR), connected cars, Industry 4.0, big language models, and generative AI, computing power will become a fast pass for enterprises to enter the next technological era. According to a report by the China Academy of Information and Communications Technology, since 2017, the total computing power of global computing equipment has been growing at a rate of more than 30% per year. Nvidia’s next goal is to make the GPU a computing power infrastructure.

“The data center worth $1 trillion in the world is almost entirely composed of GPUs, and it is still growing at a rate of $250 billion a year. These computing powers have not been accelerated.” Huang Renxun said in a recent earnings conference, “As generative AI becomes a major data center load… Nvidia sees an incredible opportunity to restructure data centers around the world.”

Huang Renxun finally led Nvidia to the center of the stage.

Continue to tap the potential of GPU
Let the GPU become popular from the fringe role of the “co” processor, which began with two schemes set by Huang Renxun more than 20 years ago: one is to practice internal skills and improve performance to avoid being overwhelmed by the CPU; Extending the versatility and doing more work makes everyone inseparable from it.

“Customers may say that the graphics card does not need such powerful performance, does not need so many functions, or the price is too expensive…”. In an entrepreneurial class at Stanford University in 2009, Huang Renxun expressed a view that clearly violated the MBA philosophy: “Ignore your customers’ opinions because they don’t know what they need.”

It was at his insistence that Nvidia kept cramming more transistors onto the board. Soon, the CPU chip giant Intel discovered that although the CPU is still the core central brain of the computer, the performance of the new species of GPU has become so tenacious that it cannot be annexed by the CPU. Today, the number of transistors on a high-end GPU is ten times that of a CPU of the same period.

In the 1990s, network cards, sound cards, Bluetooth, etc. all existed as independent chips, but later these functions were integrated into the CPU one by one by Intel. In 1999, Intel finally applied this integration idea to the graphics card, soldering the i740 graphics card on the motherboard and bundling it with the CPU. As long as this last chip is integrated into the CPU, Intel will become the only computing core of the personal computer era.

Relying on its monopoly position in the CPU field, Intel has become the company with the largest integrated GPU shipments in the world. But at this time, Huang Renxun’s strategy of frantically improving GPU performance played a decisive role at the critical moment when Nvidia avoided being completely swallowed by Intel.

Although the integrated graphics card is cheap, it lacks a separate heat sink and independent memory, and its performance is far inferior to that of a discrete graphics card. And there are always a small number of professional workers and gamers who need better performance, and Nvidia has firmly controlled the high-end GPU market.

Releases of new GPUs twice a year in February and July have been a regular cadence for years for Nvidia, to coincide with PC makers shipping to retailers twice a year in April and August. David Kirk, chief scientist at Nvidia, said: “The first step to Nvidia’s success is that we recognize that the pulse of the PC market is regular and predictable.” The emphasis on the production calendar allows Nvidia to continuously maintain The leading edge of technology, and quickly validated in the market.

At the request of Huang Renxun, in the early years, Nvidia chip designers doubled the number of transistors on the GPU every six months, which was three times faster than Moore’s Law. Today, this law proposed by Intel co-founder Gordon Moore is getting closer and closer to the day when it will be overturned-chips can no longer be made smaller.

And Nvidia realized this very early and made adjustments. While making transistors smaller, it improved performance through optimization of chip architecture, so that even if transistors cannot grow, “GPU performance will still double year by year.” , Huang Renxun finally proposed the “Huang’s Law” belonging to Nvidia after 20 years.

Of course, Huang Renxun knows that improving GPU performance can only prevent it from being robbed by the CPU. If you want to become a superstar, GPU must find its own stage. He has always maintained a strong sense of urgency in constantly releasing the potential of the GPU and finding new application scenarios.

In 2001, Nvidia released the first programmable GPU, which allowed developers to create custom visual effects. This was just a small attempt by Nvidia. In 2006, Nvidia’s real killer CUDA architecture came out, which not only further reduces the difficulty of GPU programming, but also supports simultaneous programming of multiple GPUs. That is to say, with the support of the new architecture, the GPU has jumped out of the narrow functional definition of the graphics processor, and has the possibility to participate in more general-purpose computing scenarios.

Moreover, Huang Renxun requires that all Nvidia GPUs, whether low-end or high-end, must be converted to the CUDA architecture. Are you a college student with little cash and can’t afford to rent a server? It doesn’t matter, as long as you buy two NVIDIA GPUs, you can start scientific research calculations. Nvidia cooperates extensively with universities and research institutions. From physics, biology, to climatology and chemistry, Nvidia wants to make various disciplines that require computing power support understand the new capabilities of GPU.

In 2012, Alex Krizhevsky and Ilya Sutskever, two doctoral students of Geoffrey Hinton, known as the “father of deep learning”, used two NVIDIA GPUs to train the AlexNet model for two weeks. The authoritative academic competition – won the championship in the ImageNet competition. Ilya Sutskever later became co-founder and chief scientist of OpenAI.

Through such a path, the seeds sown by Nvidia finally blossomed in the field of artificial intelligence and became the necessary infrastructure for large models.

In fact, as long as there is a demand for large-scale parallel computing, Nvidia will go to the layout. In 2015, NVIDIA launched the NVIDIA DRIVE platform to enter the field of autonomous driving. In the next few years, Huang Renxun once again became a guest at the CES exhibition with the self-driving chip that has been steadily upgraded every year. In 2019, Nvidia launched Omniverse, an open platform dedicated to digital content creation, adding fire to the metaverse concept.

“Nvidia is not just a chip company, but an artificial intelligence software platform provider with a majority of software engineers.” Sheng Linghai, vice president of research at Gartner, told China Business News. Unlike Intel’s idea of ​​vertically connecting chip manufacturing and design, Nvidia hands over manufacturing to TSMC and other foundry companies. It only maintains its identity as a “chip design manufacturer” and focuses more on continuous mining of GPU performance. superior.

CUDA has built a vibrant developer ecosystem for NVIDIA. In the server market, it is considered more efficient to hire a large number of “high school student” GPUs that can only add, subtract, multiply and divide than to use expensive “graduate student” CPUs that can do calculus.

At the COMPUTEX conference held in Taipei in May this year, Huang Renxun released the data center DGX GH200—four such data centers can surpass the computing power of Tianhe-2 (currently the strongest supercomputer in China). In his speech, Huang Renxun repeatedly compared the efficiency of training large models between CPU servers and GPU servers—GPU is ahead in terms of power consumption and efficiency.

Upstream and Downstream Risks
Nvidia is occupying the most favorable position in the current hottest field of artificial intelligence, but the risks of its daily business cannot be ignored.

According to the financial report data, since 2014, Nvidia’s accounts payable cycle has been shortened and the accounts receivable cycle has been lengthened, which reflects that its bargaining power for upstream and downstream has not been significantly enhanced.

At present, there are only three foundries in the world that have mastered the 7nm advanced process technology, namely TSMC, Samsung and Intel. Over the years, Nvidia’s fixed partners have been constantly switching between Samsung and TSMC. In recent years, Samsung has gradually fallen behind in advanced manufacturing technology, and Nvidia has to hand over all high-end products to TSMC, which increases the uncertainty of product delivery and the difficulty of expanding production capacity.

But the good news is that on June 21, Intel announced an organizational restructuring at the investor conference call, and its foundry business will operate independently and be responsible for its own profits and losses in the future. In the past six months, Intel has built factories in many regions around the world to expand production capacity. In the future, if these capacities are put into operation smoothly and meet Nvidia’s process requirements, it will undoubtedly help Nvidia get rid of its dependence on TSMC.

From the downstream perspective, Nvidia’s largest customer base in the data center business—Microsoft, Meta, Google, Amazon and other technology giants and cloud service providers are working hard to promote chip self-development plans to reduce hardware costs, even if they only partially replace Nvidia’s GPU.

According to a research report by Bank of America Corporation (BAC), cloud service providers, major enterprises and consumer Internet companies account for approximately 40%, 30% and 30% of Nvidia’s data center business, respectively. Similar to the upstream situation, for Nvidia, it is in its best interest to maintain the diversity and balance of its customer structure.

Finally, China is still a market where risks and opportunities coexist for Nvidia. On August 31, 2022, the U.S. government asked Nvidia to restrict the export of the latest two generations of flagship GPU computing chips A100 and H100 to China. This ban once caused Nvidia’s stock price to fall by more than 5%. However, Nvidia soon found a solution and launched the A800 with a 30% reduction in communication bandwidth to replace the A100 to meet the regulatory requirements of the US government.

“The Chinese market accounts for more than a quarter of Nvidia’s revenue, and this market is still growing rapidly. If it completely loses the Chinese market, it will be a huge blow to Nvidia.” Sheng Linghai said, “But at present, similar The ban is still in the process of gaming, while the United States is issuing the ban, the Ministry of Commerce is also issuing some special licenses.”

Right now, it has been a month since Nvidia’s market value exceeded one trillion U.S. dollars in early June. Nvidia has basically stood firm at the “trillion” mark.

However, in the not-too-distant past — in October last year, due to the subsidence of the cryptocurrency mining boom, Nvidia’s stock price had just experienced the largest annual decline since the company’s listing, with the lowest bottom of $108 per share. According to analysts’ expectations, global chips have simultaneously entered a cycle of downward demand, and Nvidia will not be immune.

However, at the end of November, the wave of ChatGPT swept across, and the order volume of AI servers skyrocketed, which once again pushed Nvidia’s price-to-sales ratio (PS) to 15 to 25 times. In contrast, AMD’s current price-to-sales ratio is less than 7, and Intel’s price-to-sales ratio is only 2.5.

In the first six months of this year, Nvidia stock has gained $280 overall. Many analysts are envious of Nvidia’s luck, but in Huang Renxun’s view, from graphics processors to CUDA, and then to building a software ecosystem in different fields, Nvidia took a full 30 years to shape itself into a graphics chip. A computing power infrastructure company based on parallel computing, and the explosion brought about by ChatGPT is only inevitable in a series of accidents.

“I think my greatest gift is being able to surround yourself with amazing people and give them the opportunity to do great work and help them achieve more than they thought possible,” Jensen Huang said. “And I’ve held the torch longer than anyone else.” , just because I’m more resilient. Once I get on a path, I can stay on that path for a long time and believe in it for a long time, and that’s what makes me resilient.”

error: Content is protected !!