God will also be subjugated in China.

Golden Horn Finance

2024-05-29 17:18Published in Guangdong financial field creators

原创首发 | 金角财经（ID: F-Jinjiao)

Author | Zelda

Nvidia, which became a god, was forced to cut prices by Huawei?

According to relevant foreign media reports, Nvidia recently lowered the price of H20 artificial intelligence chips for the Chinese market, and in some cases, the price of H20 chips is more than 10% lower than that of Huawei Ascend 910B.

You know, Nvidia just released a record single-quarter revenue earnings report a few days ago, and its performance has significantly outperformed market expectations for four consecutive fiscal quarters;

At the previous GPU technology conference, NVIDIA released a new generation of AI chip architecture Blackwell, with 2.5-5 times the performance of the previous generation Hopper architecture, which has the momentum of further distance from competitors.

If Huawei is far ahead in China, and Nvidia is leading in the world, then why does Nvidia still care so much about Huawei that it does not hesitate to lower its price?

Nvidia achieves God

In recent times, Nvidia has been in a lot of limelight.

On May 28, local time, Nvidia's share price exceeded $1,000, and its market value soared by $187 billion (about 1.35 trillion yuan) overnight, reaching $2.81 trillion, comparable to France's GDP and only $100 billion worse than Apple. Huang is expected to surpass Elon Musk as the new richest man in the world next year.

Nvidia's latest financial report shows that revenue in the first quarter was $26 billion, a year-on-year increase of 262% and a quarter-on-quarter increase of 18%, setting a new record for single-quarter revenue; Net profit was US$14.881 billion, up 21% quarter-on-quarter and 628% year-on-year. In the previous four consecutive fiscal quarters, Nvidia's performance significantly outperformed market expectations.

"The next industrial revolution has begun." "AI will dramatically improve productivity in nearly every industry and help companies become more cost-effective and energy efficient while expanding revenue opportunities," said Nvidia CEO Jensen Huang. ”

At present, in the field of large model training, there is almost no alternative to NVIDIA's high-end GPUs, and the supply has exceeded demand for a long time.

According to data provided by Akshara Bassi, a senior analyst at market research agency Counterpoint, in 2023, Nvidia will win 85% of the market share in the overall AI chip market, including publicly sold products such as GPUs and self-developed accelerators from cloud service providers such as Google and Amazon, while its biggest competitors in the open market, Intel and AMD, will have a market share of only 2% and 1%, respectively.

What's even more terrifying is that after being really far ahead, Nvidia also has the momentum of letting competitors chase farther and farther in technology iteration.

On March 18, NVIDIA's annual GPU Technology Conference returned to offline for the first time in five years, with the number of offline attendees doubling from about 8,000 in 2019 to 16,000, and the number of online attendees reaching about 300,000.

"In the past, only Apple's press conference could have this kind of style." A technology industry person based in Silicon Valley lamented that Nvidia is now the "top stream" in Silicon Valley.

Nvidia really lived up to this popularity. On the same day, NVIDIA released a new generation of AI chip architecture Blackwell, which has 2.5-5 times the performance of the previous generation Hopper architecture.

According to public information, Blackwell uses TSMC's 5nm family of customized 4NP processes, in which the B100 GPU has 104 billion transistors, and the B200 is integrated by two B100 dies through inter-chip interconnection technology, bringing the number of transistors to 208 billion.

As a comparison, the number of H100 transistors released by Nvidia two years ago and still the main product today is 80 billion; The performance of related products launched by Intel and AMD is still benchmarking with H100.

In addition, compared with the previous architecture iteration that focused on the performance increase of a single GPU, Blackwell's focus is on the "super chip" and platform system level, with more emphasis on the combination effect, and the upgrade strategy at the software level, in an effort to build an "AI foundry" serving all walks of life.

Huang said that using the previous-generation Hopper architecture to train a large model with 1.8 trillion parameters in 90 days would require 8,000 GPUs and consume 15 megawatts of power, while Blackwell would require only 2,000 GPUs and reduce power consumption to 4 megawatts.

Subsequently, at the recent earnings conference, Huang said that the AI chips using the Blackwell architecture have been produced on a small scale, and are expected to be delivered in the second quarter, and large-scale delivery will be achieved in the third quarter, and the data center can run on new products in the fourth quarter, and new products can bring a lot of revenue this year.

He also emphasized that the current market demand for AI computing power is in short supply, and the demand for products with the previous generation of Hopper architecture will increase.

God is in China, and he is also going to lower the price

However, God's life in China was not easy.

Affected by the escalation of U.S. export controls, in the data center business, the proportion of Nvidia's revenue from Chinese customers decreased from 19% in fiscal 2023 to about 5% in fiscal 2024.

Overall, however, the Chinese market still accounts for 17% of Nvidia's total revenue. According to a report by CCID Consulting, China's share of the global AI industry is expected to exceed 30% by 2035.

Nvidia is clearly not going to give up on the Chinese market, and nearly 1 million H20 chips will be shipped to China in the second half of this year, and as Huang said, it will continue to serve Chinese customers to the best of its ability.

In August 2022, affected by the U.S. government's chip control policy, NVIDIA's most advanced GPUs could not be supplied to customers in Chinese mainland, and could only "castrate" the bandwidth rate to develop "special versions" H800 and A800 for the Chinese market; More than a year later, the U.S. Department of Commerce upgraded export controls, and even shipments of "special version" products were halted.

At present, NVIDIA's main product for the Chinese market is the H20 chip. However, a number of people in the chip industry said that the computing power of Nvidia's current products shipped to China has shrunk seriously, and even the strongest H20 among them has greatly reduced its cost performance.

Market analysts believe that due to the limited technology, it is difficult to open the gap between H20 and Chinese competitors, so the demand has also been constrained to a certain extent. According to sources, sales of the H20 were not smooth at first, and they encountered obstacles, mainly because some Chinese customers felt that the product was "not high-end enough".

Due to the lack of obvious competitiveness of the "special reduced version" of the product, Nvidia also suffered from being "voted with its feet" by the market.

According to related media reports, Nvidia recently lowered the price of H20 AI chips for the Chinese market, and in some cases, the price of H20 chips is more than 10% lower than that of Huawei's Ascend 910B.

It is understood that China's server dealers currently sell H20 server single card prices of about 100,000 yuan, and eight-card server single card prices of 1.1 million yuan to 1.3 million yuan. In comparison, the price of a single card sold by distributors is more than 120,000 yuan, while the price of a single card for an eight-card server is 1.3 million yuan to 1.5 million yuan.

Some market participants pointed out that Huawei only started to compete with Nvidia last year, and the former will significantly increase shipments of its Ascend 910B chip this year, and the chip performs better than the H20 in some key indicators.

In addition, in the past six months, only five countries or state-affiliated buyers have expressed interest in buying H20 chips, while the number of buyers of Huawei 910B chips has reached more than a dozen in the same period. Some Chinese tech giants have already placed orders, with news that Alibaba has ordered more than 30,000 H20 chips.

Analysts believe that the involvement in the price war highlights the challenges facing Nvidia's Chinese business amid U.S. sanctions on AI chip exports and increased competition, casting a shadow over the company's 17% revenue in fiscal 2024.

Hebe Chen, market analyst at IG, said: "Nvidia is taking a delicate line trying to find a balance between preserving the Chinese market and dealing with tensions in the United States. In the long run, Nvidia is certainly also preparing for the worst-case scenario. ”

The rise of Huawei

In its latest annual report released in February, Nvidia listed Huawei as one of its main competitors in the AI chip market for the first time. In the past, only Intel and AMD were on this list of rivals for a long time.

Specifically, it is a fear of Huawei's Ascend chips.

Huawei's Ascend 910 was released in August 2019 and was initially built on TSMC's 7nm process. Since then, the 910B has increased its memory capacity and upgraded Huawei's self-developed interface similar to NVIDIA's NVLink, significantly improving its large model training capabilities.

Wu Zihao, CEO of Ronghe Consulting, a semiconductor industry consulting firm, predicts that Huawei is expected to launch the Ascend 910C in the second half of 2024, and the interconnection capability will be the most important improvement in this iteration. If Huawei can successfully launch the subsequent Ascend 920 series in 2025, it will further surpass the H20 in terms of computing power and interconnection capabilities.

At present, technology companies including Baidu, iFLYTEK, and Meituan have begun to purchase Huawei chips. Unlike most large models in China that are trained on NVIDIA GPUs, iFLYTEK's Spark large models are even based on the Ascend 910B.

Liu Qingfeng, chairman of iFLYTEK, previously bluntly said that the biggest pressure on using Huawei chips is that the task that could have been completed in one month with the NVIDIA solution may take three months to complete due to the migration of a series of development tools, which requires re-verification of the results of the algorithm. "But this step has to be taken."

In terms of specific shipments, Wu Zihao predicts that Huawei's supply capacity will ship about 400,000 units in 2024; As a comparison, Ascend will only ship more than 100,000 units in 2023. A brokerage analyst who focuses on the electronics industry is more optimistic, believing that Ascend's guaranteed shipments in 2024 will be 400,000 copies, with a high probability of rushing to 500,000 copies.

Of course, the development of Huawei chips is still in its infancy, and there are still many problems that need to be solved.

A semiconductor analyst said that due to the imperfect software ecosystem, the current high willingness to use Huawei chips is mainly information and innovation customers, who tend to focus more on autonomy and controllability rather than performance and user experience.

Wu Zihao also said that by the end of 2023, the attitude of major Internet companies towards Huawei Ascend is still relatively vague. "The interconnection capability is still a bit poor, the key is the lack of engineers in the relevant ecosystem, and the migration and driving of ecological software require people, and Huawei needs to send people to support it."

In a report released in March, market research firm Isaiah also pointed out that Huawei's limited number of engineers is one of the constraints it faces when selling the Ascend 910 series products. Enterprises that can make full use of Huawei's Ascend AI chips to meet their large model training needs must rely on Huawei's team to help, and the migration can take 9 to 12 months, depending on the complexity of the large model, and at least 3 to 5 months for engineer training.

Although Huawei's Ascend software and ecosystem still have shortcomings, it is clear that the Chinese market will gradually tilt towards Huawei in anticipation of the United States' intention to relax chip controls, thus giving the latter enough soil for development.

"H20 can be seen as a model under the current US ban framework, and Nvidia, AMD and Intel will not be better than H20 in the future. On the one hand, H20, which is likely not to improve in the next two or three years, and on the other hand, Ascend, which continues to improve in performance and the software ecosystem is gradually growing, how will major domestic manufacturers choose? ”

However, Nvidia is not a vegetarian either, and in the face of Huawei's menacing threat, there are still many back-ups.

On the one hand, the supply of Huawei's Ascend chips is limited, and in anticipation of the shortage of computing power due to the rapid development of AI large models, NVIDIA still has room to exert its strength with the help of the CUDA ecosystem. A senior technical officer of a major domestic Internet company said that many colleagues in his company are unwilling to invest time in adapting to other platforms other than CUDA. "The cost of switching is relatively high, a lot of things have to be rewritten, and there is a lack of mature supporting tools."

What's more, taking 10,000 steps back, Nvidia can also exchange prices for the market. Some industry insiders pointed out that although the cost of H20 is the same as that of H100, there is still a lot of room for price reduction due to the higher gross profit. "Pricing depends on the competitive landscape and how much Nvidia wants to win these projects."

Nvidia's price cut also further confirms that the tech giant wants to trade price for the market. However, being able to force Nvidia to cut prices to meet the competition does prove to some extent that Huawei is far ahead.

Resources:

Caixin "Nvidia's Revenue and Profit in the First Fiscal Quarter Soared, and Its Stock Price Has Risen by Nearly 90% Since 2024"

Caixin "NVIDIA's Next Step: AI Foundry"

Hexun.com's "Breaking Big News! Involving China, Nvidia H20 chip price reduction: the price is more than 10% lower than that of Huawei Ascend 910B》

Yicai "Nvidia H20 chip price reduction rumors are spreading, Huang Jenxun once said that "the Chinese market is fiercely competitive"

View original image 125K

•