Houlang Forest Research Laboratory|ivy editor, co-ordinator|There is no doubt that NVIDIA is leading a global generative AI revolution with its unparalleled technology and products. Its influence has surpassed any of its competitors and become a leader and trailblazer in this field. When Nvidia rose to prominence again, a Wall Street analyst lamented: "There is a war going on in the field of artificial intelligence, and Nvidia has reached the only arms dealer." ”Nvidia on the altarOn February 21, Eastern time, Nvidia announced its fourth quarter and full-year results for fiscal year 2024 ended January 28. For the quarter, the company achieved revenue of $22.1 billion, up 265% year-over-year. In fiscal 2024, the company's revenue increased by 126% year-over-year to 609$2.2 billion. Nvidia's data center division, with A100 and H100 chips at its core, continued to lead the company's revenue list: revenue soared to $18.4 billion during the period, up 409% year-on-year. In fiscal year 2024, data center revenue reached $47.5 billion, up 217% year-over-year. Data center revenue comes primarily from large cloud service providers, GPU specialists, enterprise software, and consumer internet giants. 
In 2023, Huang announced that Nvidia has given supercomputers with supercomputing power to 50 of the 100 largest companies in the United States. In the tech world, a company's ability to develop large models is even evaluated by the number of GPU chips it has. Technology companies around the world, with and without money, are vying to be the first to "recharge" NVIDIA. Huang also said that accelerated computing and biointelligence have ushered in "explosive" growth. Whether it's a company, an industry, or a country, the demand for [chips] is soaring around the world. 
What does "explosive" mean? Great demand! Yes, but not quite. If it is only a large demand, then the first business can enjoy high income and a gross profit margin that is the same as before. "Explosive" means that supply exceeds demand, it means pricing power and high premiums, it means inflated gross profit margins, and huge benefits. This is an extremely difficult right to obtain. Nvidia, which has been waiting for this moment for many years, naturally refuses to let go easily. In August 2023, Tae Kim of Barron's revealed that according to investment bank Raymond James, the cost of Nvidia's H100 chip is only $3,200, but the selling price is as high as more than $30,000, which means that Nvidia's cost profit margin is as high as 1,000%! Nvidia insiders also confirmed the news and revealed more insider information. He said that Nvidia's H100 chips are in short supply, and many international technology giants are rushing to buy them, and even accept Nvidia's condition of tying in A100 at the same time. With such a mix-and-match set, Nvidia can still earn 8 times the profit margin. Moreover, at present, such a hot situation will continue for two years. So naturally, we saw in the financial report that the net profit in the fourth quarter was 12.3 billion US dollars, an increase of 769% year-on-year. Adjusted gross margin (GAAP) was 767%, the market expectation is 754%。
In the 5 years before OpenAI launched ChatGPT at the end of November 2022, NVIDIA's gross profit margin has been maintained at around 60%+. Odd goods can live on the ground and let Nvidia sit on the ground and raise the yield by 10 percentage points! Of course, this is also in line with the supply and demand relationship of the market economy. How's that, a bit of a violent aesthetic, right? But even so, Microsoft, Amazon, Google, Facebook and other recent earnings conferences have announced plans to increase capital expenditures this year to directly purchase "the kind of AI components provided by Nvidia". The "AI components" they refer to are likely to include Nvidia's upcoming next-generation AI chip, the B100, which is said to be three times faster than the H100. If the B100 is released on schedule, it will further defend NVIDIA's supremacy in the field of AI hardware, and it is also an important guarantee for NVIDIA to maintain profitability. Bank of America analyst Arya believes that the Nvidia B100 will be priced at least 10%-30% higher than the H100, and demand is likely to continue until at least mid-to-late 2025. For the not-too-distant future, at the world's best summit in Dubai in early February, Huang revealed that in the next four to five years, the global data center market will grow from the current $1 trillion to $2 trillion. Nvidia's "program" is to work with software service providers around the world to share this huge business opportunity. Accident and necessityHaving said that, Nvidia is not deliberately hoarding, it has worked hard to cooperate with multiple chip foundries for production, but it is a pity that the elasticity of the ** chain of chips is insufficient, of course, this is also a good embodiment of the so-called "explosive". In fact, even the encounter between Nvidia and the AI industry has a lot of unexpected elements. All along, Huang and his gang have been making game graphics chips. In 1999, Nvidia launched the GeForce series of graphics cards, which sold very successfully with the popularity of the game "Quake". In 2000, Ian Buck, a graduate student studying computer graphics at Stanford University, played "Thor's Hammer" with 8 projectors by chaining together 32 GeForce cards. And just like that, he built an 8K resolution console and exclaimed, "It's beautiful." ”
This perception of beauty coincides with Huang's entrepreneurial focus on image rendering graphics card business. Equally similar is their imagination of beauty. Buck thought, the GPU's parallel computing power is so strong, it shouldn't just be used to play games. So he and his friends founded the "Brook" project, intending to start with GPU support programmability, use GPU for general-purpose computing, and published a series of **. These ** made Huang pick up the treasure of buck and invest a lot of resources to support the Brook project. After all, one more application scenario, one more way for the company to live. In 2004, Ian Buck, who had not yet graduated with a Ph.D., entered Nvidia as an intern and began to create GPGPU models. Here, the former GP of GPGPU refers to General Purpose, that is, the use of GPU's powerful computing power to deal with problems in other fields. The reason for giving up the CPU can be thought of this analogy; A CPU is like a college student, calculus can be calculated. A GPU, like many middle school students, can only do simple addition, subtraction, multiplication and division. But if there is a job that requires calculating addition, subtraction, multiplication and division within 100 millions of times, then dozens of middle school students are obviously beating college students. AI model training requires this kind of calculation. Although middle school students and college students have specializations, the premise is to break down a difficult problem into countless small problems that middle school students can solve. In reality, developers have to be able to easily write ** and use the parallel computing power of the graphics card. Ian Buck, who is currently vice president of NVIDIA, continued his PhD project when he joined NVIDIA, leading the team to launch CUDA, a parallel computing platform that continues to iterate today, in 2006. The vast majority of AI models today are developed using CUDA. 
The idea was very forward-thinking at the time. may be too avant-garde, so the outside world is questioning. Investors' concerns are understandable. In the direction of "scientific computing", at that time, the total market was only a few billion dollars, and in 2008, NVIDIA's revenue exceeded 4 billion US dollars, and its net profit reached 7$9.8 billion. It's not good to do the main game business with huge potential, but it takes a diversified route, and some are not honest enough. And the consequence of investing a lot of manpower and money in CUDA is that Nvidia's profits fell sharply after the financial crisis, and even suffered losses in 2009 and 2010, and the stock price plummeted. For a while, everyone on Wall Street asked Huang why he wanted to do CUDA, and even Wall Street defined the market value of CUDA as "zero". Huang later said in an interview that he only knew he was betting on "accelerated computing" and believed it was the future. If this insistence is right, then a larger and wider market will open the door to him. Such vision and ambition reminds me of Wang Chuanfu, when the battery business was doing well, he set his sights on the larger market, although it was impossible to predict whether and when the electric vehicle market could start, but he persevered. At this time, Nvidia is still separated from the AI industry by a computer competition. God didn't make Lao Huang wait too long. In 2012, the current round of AI wave Xiaohe began to show sharp corners. In that year, Professor Jeffery Hinton, known as the "father of deep learning", and his two students, Alex Krizhevsky and Ilya Sutskeverz, participated in the world's most authoritative computer vision competition, ImageNet, and the deep convolutional neural network Alexnet designed won the championship in one fell swoop. By the way, Ilya Sutskeverz, not the same name, he is now the chief scientist of OpenAI. 
The secret of their championship came from NVIDIA. It turned out that training Alexnet with two NVIDIA GTX580 graphics cards equipped with CUDA took only one week to complete the workload of 14 million **. This process would take several months if it was done with a CPU. This marks a major breakthrough in computing power in the entire history of artificial intelligence. Overnight, the GTX580 became the most important necessary piece of gear for the future. Soon, Google, IBM and other major manufacturers, as well as various university laboratories, began to order GPUs from NVIDIA. Businessman is also an academic Lao Huang, and began to accelerate the creation of GPU hardware designed for AI. You can probably guess the next thing: the year after the computer competition, that is, in 2013, the classification method of Nvidia's annual report revenue business was revised to create a "data center". In the following 11 years, the data center business also grew by leaps and bounds, until the GPU became the "infrastructure" for training large models, and the data center business finally grew into a huge "monster". 
Above we said that GPU unexpectedly met the AI era, but in fact, Nvidia also has the momentum to inevitably go to glory. As for the heroic deeds of Nvidia's victory over AMD and the battle against Intel, we will leave it to the next breakdown. How is Nvidia made? Success is hard to replicate. But we must sum up a trait, Nvidia and Lao Huang have focused beauty, courage and rare flexibility all in. In the 31 years since its establishment, NVIDIA has not changed the track, but has been constantly expanding the application scenarios of the same type of products, and vigorously iterating the latest and strongest products miraculously and quickly. Lao Huang insisted on staying in the pit of GPU and never gave up in the AI scene for 12 years. If the mobile phone chip can't be entered, the mining chip will be studied, and the mining will be out of play to engage in intelligent driving, until the cloud of the large model is cleared. 
Of course, it is obviously not enough to just wait, there must be cheats. What has helped Lao Huang turn the tide many times is the CUDA development environment. With its low cost and compatibility, strong community resources, and a complete ecological chain, NVIDIA has been popularized and promoted by universities through learning, and is now familiar and recommended by many industry and academic professionals, and the moat has been built. If the design and manufacturing of GPU chips is 5 percent, the difficulty of building a mature development environment like CUDA, which can be recognized and widely used by most users, is probably 9 percent. Indispensable is NVIDIA's high degree of flexibility, in the early days of its founding, the company launched a "three teams, two quarters" operating model, which allows the company to launch new products every six months, in line with the graphics market product cycle, and 1-2 R&D cycles ahead of the market. The subtle innovations and high-intensity product output of these operations have allowed NVIDIA to survive tenaciously. "The only player".At this point in the article, rational people will ask one more question, no matter how good Nvidia is, how long can such a carnival last? "Let the bullets fly a little longer", because, this question is difficult to answer exactly. Everyone understands the principle that blessings and misfortunes depend on each other. Nvidia stands on the top of the most beautiful mountain, and it also hides the biggest killer: in addition to Google's TPU, Intel's Falcon Shores next year, AMD's MI300, Microsoft is also cooperating with AMD to jointly develop its own artificial intelligence processor Athena, everyone's purpose is to balance NVIDIA. 
Former companions, comrades-in-arms, major customers and competitors, these identities are constantly switching in small circles. In addition, some chips in professional fields, such as ASIC, Analog, and FPGA chips, are also waiting for opportunities to find more professional AI entry points. In addition, AI development is still in its early stages, and existing Transformer models may not be the optimal solution. In the future, if the development of large models encounters difficulties, or there are other technical paths, GPUs are no longer applicable, and NVIDIA will once again face huge market risks. However, with $60 billion in revenue and at least 5 application scenarios, even if it fails at the forefront of AI again, Nvidia can still stand. Huang has become the longest-standing CEO of Silicon Valley's tech giants, and he has no plans to stop. In an interview with CNBC, he joked that he could fight for another 30 years. Even if the physical body is no longer there, you still have to continue to fight in the form of AI digital humans. Dan Ives, a senior analyst at Wadebush, said that Huang is the godfather of AI, and Nvidia's fourth-quarter results are the most important corporate earnings report in years, and from the perspective of peak demand, this AI party has just begun. In the next two or three years, companies such as AMD may compete, but for now Nvidia is "the only player," Ives said.