Review | The only stock that has increased 500 times after the dot-com bubble! Why Nvidia?

Looking Back at NVIDIA's Success Story

Looking back over the past 20 years, NVIDIA's success was not accidental:

In 2002, NVIDIA's collaboration with Microsoft on the Xbox graphics chip ended in failure, facing a survival crisis and its first-ever performance decline in history.

Through unremitting efforts, NVIDIA eventually won the order for Sony PS3 and reached a cross-licensing agreement with Intel. This critical diversification choice also put the company back on the growth track.

In 2007, NVIDIA's revenue reached $4.1 billion, tripled compared to 2002, and the company no longer relied on a single major customer, successfully turning the crisis into an opportunity.

In 2000, NVIDIA completed its first acquisition, acquiring 3Dfx, the largest graphics processor manufacturer in the mid-1990s, for $70 million in cash and 1 million shares of company stock. In 2019, NVIDIA acquired Mellanox for $6.9 billion, boosting the company's fastest-growing data center business.

NVIDIA founder Huang Renxun appointed David Kirk as chief scientist to promote the CUDA computing framework, which was a core turning point: developers no longer had to use complex machine language for programming, but could directly use the GPU to compute programs written in high-level languages through the CUDA framework.

When the big data explosion and artificial intelligence (AI) emerged, traditional CPUs began to struggle. Behind the changes in NVIDIA and Intel's market capitalization is the alternation of two computing eras, as well as the reversal of the fortunes of the two companies.

Currently, GPUs provide 56% of the total computing power in the world's top 500 supercomputers. At the 2020 GTC conference, NVIDIA released the DGX A100 based on a new architecture, breaking 16 AI performance records in one go, and the speed was 4.2 times faster than the previous generation of products. In the past five years, under NVIDIA's leadership, GPU performance has increased 25 times.

NVIDIA's path to expanding AI technology applications has not stopped: in April 2023, NVIDIA interpreted the company's new Auto DMP open-source method.

Reinforcement learning methods have shown great potential in the field of chip design, and the DREAM Place uses a GPU-accelerated algorithm calculated by the PyTorch framework to calculate wire length and density gradients. The global layout achieved more than 30 times acceleration of the original process, which can greatly improve the effectiveness of GPU-accelerated layout and the combination of multi-objective parameter optimization of artificial intelligence/machine learning, and achieve new exploration in modern chip design processes.

The Absolute Leader in the AI Era

After the collapse of the internet stocks in 2000, the market demand for NVIDIA was not affected, but continued to make rapid progress in product technology research and development, especially in the gaming graphics card field, occupying the market core with 15 generations of updated products in the past 15 years (JPR released the independent graphics card market share report for the third quarter of 2022: NVIDIA accounted for 88%, AMD accounted for 8%, and Intel accounted for 4%).

In the AI era of the 21st century, NVIDIA has become the core hardware supplier.

NVIDIA supplies AI chips with GPU as the core (fully utilizing the basic structure of GPU multi-core parallel computing, which can support parallel computing of large amounts of data and has higher floating-point operation capability). From Tesla, Fermi, Kepler, Maxwell, Pascal, Volta to the latest Turing, the computing power of the microarchitecture is getting higher and higher, and the architecture is also leading the market.

In the 2023 AIGC wave, NVIDIA became the global leader in computing chips.

In this AI competition, as of the end of 2022, Meta was the first to use NVIDIA A100 chips, with 21,400 A100 GPUs; the second was European defense contractor Leonardo, and the third was Tesla, with 7,360 A100s.

Public clouds, private clouds, and national high-performance computing have the largest demand, and private enterprises have more high-end GPUs than national supercomputers. The owners of higher-end GPUs, such as NVIDIA's Eos supercomputer and the EU's MareNostrum 5 supercomputer, mainly include 4,608 H100s and 4,500 H100s, respectively.