"Our company has been ten years ahead of our competitors. ”
This is Nvidia CEO Jensen Huang's bold statement at the 2023 DealBook Summit on November 30. So far, at least, Huang has had the emboldenment to make such a statement. According to NVIDIA's financial report for the third quarter of fiscal year 2024, revenue during the reporting period was 181200 million US dollars, a year-on-year increase of 206%;Net profit 924.3 billion US dollars, a year-on-year increase of 1259%, and this performance also enabled Nvidia to surpass TSMC and Intel and become the No. 1 semiconductor company in terms of revenue.
Nvidia CEO Jensen Huang.
At the end of the meeting, however, Huang also showed his anxious side: "I don't wake up with pride and confidence, and I wake up every morning worried that the company will go out of business." "It is true that with ChatGPT sounding the clarion call for the exponential growth of generative AI, Nvidia stands at the top of the semiconductor market value with the strong performance of its own chips and a stable ecosystem, but at the same time, many challenges also need to be overcome by Nvidia.
Nvidia's "** jump".
From stock price, to market value, to revenue, Nvidia, as a Fabless (fabless) chip design company, has completed the counterattack on the old IDM step by step.
On July 8, 2020, Nvidia surpassed the old giant Intel for the first time with a market value of $251.3 billion, becoming the most valuable chip company in the United StatesSince 2023, the stock price has soared, and at the opening of the US stock market on the evening of May 30, Nvidia's stock price rose to 404$25 shares, directly led LinkedIn Weida market value to break through the trillion level, not only became the first chip company in the history of the US stock market to touch $1 trillion, but also became the world's sixth largest market capitalization company, up to now, Nvidia's stock price has doubled, continuing to refresh the competitive pattern of the AI chip market.
It is worth mentioning that in terms of the questionable earning ability in the past, Nvidia has also achieved a "counterattack". As of July 30, 2023, Nvidia disclosed financial results for the second quarter of fiscal 2024 showing that its revenue reached 135100 million US dollars, a year-on-year increase of 101%, surpassing Intel's 12.9 billion US dollars in the same period for the first time, slightly lower than 156 in the same period$800 million TSMC;Net profit increased by 843% year-on-year to 61$8.8 billion, another record high. In the third quarter, the revenue gap between Nvidia and other semiconductor companies further widened, to 181The revenue of 200 million US dollars surpassed TSMC in the same period and became the "most profitable chip company".
Nvidia's Q4 FY2021 to Q3 FY2023 by business unit and total revenue
*: Nvidia went from a market capitalization of more than one trillion to a revenue leader, and Nvidia only took 6 months. In the first meeting in the third quarter, Nvidia said that the current capacity bottleneck has been basically solved, and then the demand will determine the sales volume.
So, will Nvidia be anxious about demand?
Judging from the global pursuit of artificial intelligence and large language models, the demand for computing power, a new type of productivity, is only increasing, which seems to bring more room for growth to NVIDIA, but also challenges.
Competitors continue to emerge
First of all, Intel and AMD are clinging to the data center, providing Gaudi3 and Instinct Mi300X as high-performance AI chips, respectively, even if they can't cross the ecological moat of NVIDIA CUDA in a short time, but they will still give customers more choices. "(Artificial intelligence) is a market that is constantly segmented, although Nvidia is currently the main supplier of AI chips, but as the market matures, artificial intelligence will transition from the training stage to the inference stage. At this time, other leading companies will develop custom chip ASICs optimized for AI workloads, which will build a more competitive environment. Alan Priestley, vice president analyst at Gartner, told reporters.
AMD's AI acceleration chip, the MI300X, will be released on December 7.
Another problem faced by NVIDIA is that major customers are actively developing AI chips suitable for their own products.
Google has been exploring TPU (Tensor Processor) for parallel computing a few years ago, and the TPU V5E, which was launched in August this year, supports models running up to 2 trillion parametersOn November 16, Microsoft announced the AI acceleration chip Azure Maia 100, which together with the CPU Azure Cobalt 100 provides performance support for Microsoft's cloud service platform Azure and its flagship product CoplietAmazon released an upgraded accelerator chip for AI systems, Trainium2, on November 28, and even OpenAI's CEO Ultraman said in a previous interview that "the option of self-developed chips is not ruled out". Perhaps Nvidia is more worried about "not being needed by customers" than being replaced by competitors.
Data**: For customers, self-development may not be to compete with NVIDIA, but to further reduce costs - especially time costs. Since the development of large language models, they have entered a stage where they need to be updated quickly, so time has become an important metric that customers cannot ignore. It is reported that the delivery cycle of Nvidia H100 ranges from 36 weeks to 52 weeks, and compared with the delivery speed of A100 at the beginning of this year, it is only about 4-20 weeksAt the same time, customers also prefer to control the pace of product iteration in their own hands rather than waiting for NVIDIA's "semi-annual" performance upgrade.
The market is subject to change
Another concern for Nvidia on the demand side is the uncertainty of the global market. According to the annual report disclosed by NVIDIA in fiscal year 2023 (February 2022-January 2023), the revenue of the Chinese market accounts for 47% of Nvidia's global revenue. Colette Kress, Nvidia's chief financial officer, said in Q3 commentary: "Our sales in China (and other restricted regions) accounted for 20%-25% in previous quarters and are expected to decline significantly in the fourth quarter. ”
It is understood that Nvidia has launched three chips, HGX H20, L20 PCLE and L2PCLE as solutions for the export licensing regulations of the US Department of Commerce, which are expected to be released in the first half of next year. However, judging from the current specification list, there is still a certain gap between the performance of the H20 and the "previous generation solution" A800, and it is uncertain whether this new batch of chips can bring a large number of orders and increase turnover for NVIDIA.
At the same time, domestic cloud service companies are also calling for a "calm and long-term view" on the development of large models, and actively looking for new ways to efficiently use computing resources. Chairman Robin Li once said at the recent Wuzhen Summit: "If you want to cultivate a large model from scratch, there are high technical barriers and costs, and repeated development will cause a great waste of social resources." In the future, enterprises will use plug-ins or APIs to integrate into the large model ecology of artificial intelligence companies, and build their own industry applications and solutions on this basis is the best solution. ”
It's cold in the heights. Standing at the peak of AI chips, NVIDIA is not only the benchmark and catch-up goal of all competitors, but also tries to avoid the situation of "downhill in all directions", and seeks and even creates new mountains.
We're always in danger" — Nvidia CEO Jensen Huang's mantra is apt to describe Nvidia at any given time. For this company that has suffered setbacks in games, compromised on mobile, and reached the top of AI, it is now a period of coexistence of crisis and opportunity, and actively cooperating with multiple parties may be NVIDIA's next move. As Huang put it, "We try to do business with everybody." ”
Author丨Wang Xinhao, Zhao Yutong (real Xi reporter), editor丨Zhang Xinyimei, editor丨Maria producer丨Lian Xiaodong