한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

It has been a year and a half since the sale of A100 and H100 was banned, but now it is becoming increasingly difficult for Nvidia to sell the castrated versions to China.

Recently, Jeffries analysts said that the United States will conduct an annual review of semiconductor export controls in October, and it is "very likely" that Nvidia H20 will be banned from being sold to China.There are three possible forms of sales ban: a ban on specific products, lowering the upper limit of computing power, and limiting memory capacity.

Nvidia plans to launch a new special version of AI chip for China. This news has not stopped since Huang Renxun officially announced Blackwell in March. The current mainstream news is,Nvidia plans to launch a "castrated version" of B200, B20.

But many small and medium-sized entrepreneurs in China understand:Regardless of the price or the difficulty of obtaining it, Nvidia’s latest high-end AI chip will most likely only be available in the cloud.

AI entrepreneur Jason told Alphabet List (ID: wujicaijing) that his company's previous direction was the AI ​​application layer. In addition to renting A100 and H100 computing power through cloud services, 50 NVIDIA V100 chips and NVIDIA 3090 graphics cards were deployed locally.

This choice is made not only because start-ups want to pursue cost-effectiveness and reduce costs, but also because AI application layer businesses do not actually require extremely high computing power. V100 is a computing card released by NVIDIA in 2017. At that time, the official price of a V100 eight-card server was 1.02 million yuan. Jason's 50 V100s were second-hand goods "picked up at 900 yuan each", and the 3090 graphics cards were bought at a price of 5,000 yuan each.

Starting from the end of 2022, Nvidia's most powerful chips were blocked from export to China due to the United States' upgraded semiconductor export controls. Neither the A100 nor Blackwell's most powerful H100 could be sold officially in China. After that, the United States' export controls on high-end chips became increasingly strict. Nvidia launched one "China-exclusive version" chip after another, from A800, H800, to H20, L20, and L2.

For most entrepreneurs, the special edition is not only inferior to the "original" in performance, but also prohibitively expensive. A seller told alphabet that the price of the H20 eight-card server is around 1.3 million yuan. IT Times reported that the main demand for H20 is still Internet giants such as Baidu, Alibaba, Tencent and ByteDance.

Jason said that H20 is mainly used for reasoning, but "H20 is not as good as 4090".Because the latter is sufficient, and H20 is a commercial chip. "Commercial use has a depreciation rate, and the discount is relatively large. The computer room replaces a batch every few years, and generally does not retain its value. For example, V100 sold for tens of thousands in the past, but now it can only be sold for a few thousand. It has only been five or six years. If it were not for the AI ​​boom, it would be worth at most 500."

More than one industry insider told Alphabet that they prefer to use Nvidia RTX4090. This flagship product, launched in October 2022, was originally marketed as a gaming graphics card, but is also favored by the AI ​​industry and is also affected by US chip export controls.

AI supercomputer supplier Zhejiang Huaxiyun Technology Co., Ltd. also said that "the 4090 is the best one at present", but "it also depends on the configuration, networking, and graphics card". In fact, the 4090 can meet the needs of many customers, and it is also more cost-effective. The staff member also added that the company will install 100 4090s this month.

Jason was not excited about the news that NVIDIA will launch a new special product in China, the "castrated version" of B200, B20, and said directly "it depends on the price-performance ratio." Hua Xiyun said that it has not heard any related news in the industry yet.

Reuters previously reported that Nvidia has cooperated with China's Inspur Information on B20, but the latter has responded that the news is not true.

A

On March 19 this year, Huang Renxun took the stage at the SAP Center in San Jose, California, and delivered a grand keynote speech titled "Witness the Transformative Moment of AI."

Huang rarely brags. Nvidia officially announced the new generation Blackwell architecture and launched the chip B200 and the super chip GB200. "New nuclear bomb" is the description of Nvidia's new products by the outside world. Jim Fan, who was just promoted to Nvidia's research manager at the time, lamented that Moore's Law could no longer hold the company back.

But an experimental exploration from Apple has slightly stopped Nvidia's progress. Huang Renxun may have become the person in the world who least expects Apple's smart phone to be launched.

On July 30th, Beijing time, Apple published a technical paper, which included the following information: the two AI models that support Apple's intelligence are pre-trained on Google's cloud chips.

First, Google's self-developed chip is the TPU tensor processor, which was previously used internally by Google and not sold to the outside world. This time, it snatched away a big customer. Second, Apple was previously reported by the Wall Street Journal in May that it was developing self-developed chips for its data center servers. Now that it is not using self-developed chips, it still did not choose Nvidia. This is enough to embarrass Nvidia.

Apple's "waiting" for Nvidia is a microcosm of Nvidia's current situation. Nvidia's stock price rose 150% in the first six months of this year, but plummeted in July, accounting for four of the eight largest declines in market value.

In addition, Nvidia has been hit by a series of bad news recently: it is rumored that the delivery of its new chip B200 will be delayed by three months or even longer; the US Department of Justice has launched two antitrust investigations against it.

In contrast, there are news about the "China Special Edition" - according to multiple foreign media reports, Nvidia plans to launch the B20, a castrated version of the B200, to China. In addition, Nvidia may also sell servers equipped with the latest chips to China to make up for the performance of the special chips. If the news is true, this will be the first time that Nvidia has launched a server product specifically for the Chinese market.

The above measures also show that despite facing numerous obstacles, Nvidia has not only not given up, but has also made greater efforts for the Chinese market.

End of 2022ChatGPTThe AI ​​wave has been surging for nearly two years, and Nvidia has also leapt from a chip giant to one of the companies with the highest market value in the world. Now that the AI ​​track is gradually returning to rationality, and competitors are slowly forming a siege, China has become a market that Nvidia is increasingly difficult to abandon.

B

For a large number of small and medium-sized start-ups in China's AI field, Nvidia's China-specific chips are not their focus.For China's large Internet companies, the main buyers of special edition chips, choosing NVIDIA is only the best solution at the moment.。

In October 2023, NVIDIA launched the HGX H20, L20 PCle, and L2 PCle tailored for China. Among them, H20 is the highly anticipated "general among dwarfs", the strongest model among castrated versions. However, a month later, with the news that NVIDIA missed the delivery date and postponed the delivery of H20 to the first quarter of the following year, many companies turned to domestic chips.

At that time, China Fund News reported that Baidu had ordered 1,600 Ascend 910B chips from Huawei for 200 servers. Zhou Hongyi also said at the Wuzhen Summit shortly afterwards that 360 purchased more than 1,000 Huawei AI chips, earlier than Baidu.

In the first half of this year, there were reports that companies were waiting and watching, domestic manufacturers such as Huawei were competing, and Nvidia H20 sales in China were weak. In May, there were reports that the price of H20 was reduced. The reasons pointed to two points. One was that H100 was not out of stock and began to reduce prices, so H20 followed suit. The other was that Huawei Ascend 910B competed with it, and the latter had a more advantageous initial price.

In the second half of the year, H20 sales in China seem to have taken a turn for the better. Not long ago, SemiAnalysis, a chip industry consulting firm, predicted that the H20 chip is expected to boost the company's performance in China in the current fiscal year, and it is expected to deliver more than 1 million H20 chips this year. Based on the selling price of $12,000 to $13,000 per chip, H20 alone can contribute more than $12 billion in revenue to Nvidia, a figure that exceeds Nvidia's overall revenue in China in the previous fiscal year. The IT Times report pointed out that the rise in H20 sales is mainly due to Nvidia's ecological advantages and the tight supply of Huawei's Ascend 910B.

This is only a short-lived happiness for Nvidia, as the turnaround gained due to its ecological advantages and tight supply from competitors cannot be said to be stable. The next generation of "China-specific" chips is likely on the way, but Nvidia faces many challenges.

In addition to the rumored "castrated version" of the B200, Nvidia also plans to pair servers with the new version of the chip specifically for China, according to The Information. Nvidia has never launched a server specifically for China before, and this move is to maximize the performance of the special chip with the server to make up for the shortcomings of the "castrated version".

If the news of using servers as a "package" solution is true, it will be a new attempt by Nvidia under the constraints of the sales ban.

The bad news is that the market has once again heard that the delivery of Nvidia's latest chip has been delayed. According to a recent report by The Information, Nvidia has told customers that the release of B200 will be delayed by three months or more, and the mass shipment time may be delayed to the first quarter of next year (originally planned to start mass production in October this year).

The "customers" mentioned here include many technology giants. According to reports, Meta has placed orders worth at least $10 billion. Microsoft has also increased its order size by 20% in recent weeks and plans to provide Meta with 100 million units by the first quarter of next year.OpenAIBe prepared to spend between 55,000 and 65,000 GB200.

There is reason to suspect that the delayed delivery of B200 will also affect Nvidia's pace of launching the castrated version in the Chinese market. The report said that the reason for the delayed delivery was a "design defect" found during production.

C

While Nvidia has yet to confirm the launch of the B20, few doubt that it will happen.

Compared to the first time Nvidia was hit with an export ban on AI chips in 2022, the chip giant is now even more reluctant to give up the Chinese market.

China's huge demand for chips is naturally the primary driving force.In fiscal years 2022 and 2023, mainland China and Hong Kong contributed US$7.111 billion and US$5.785 billion in revenue to NVIDIA, respectively, accounting for 31.7% and 25.9% of its total revenue, respectively.

However, due to the US chip ban, Nvidia is facing the risk of stalling in China. In fiscal 2024, Nvidia's revenue in the Chinese market, including mainland China, Hong Kong and Macau, will stop at US$10 billion, and its share will drop to 16.9%.

In May this year, NVIDIA released its performance report for the first quarter of fiscal year 2025 (ending April 28, 2024). In the data center business, NVIDIA's revenue share from Chinese customers has dropped from 19% in fiscal 2023 to a mid-single-digit percentage (5%) in fiscal 2024.

Huang Renxun is also well aware of the competition launched by China's local chip manufacturers: "Our business in China has indeed declined a lot compared to the past level. Due to technological limitations, competition in China is now more intense. These are facts." A few days later, Huang Renxun mentioned Chinese chip companies again, saying that China has many GPU startups and that people should not underestimate China's ability to catch up in the chip field.

From a certain perspective, the US chip ban has made it more difficult for Chinese AI companies to survive in this wave, but it has also provided space for the development of Chinese chip manufacturers. From Huang Renxun’s perspective, this is undoubtedly dangerous.The clock is ticking, and Nvidia has little time left to break out of the "castration".

This is not all the reasons why NVIDIA is more reluctant to give up the Chinese market. Compared with the end of 2022, when ChatGPT set off a thousand-model war and NVIDIA's sales and stock prices soared, NVIDIA is now facing an increasingly uncertain situation.

This year alone, OpenAI CEO Sam Altman's 7 trillion chip network ambitions were revealed, Microsoft developed a replacement for Nvidia's ConnectX-7 network card to improve the performance of its self-developed chip Maia, and Groq, a technology company founded by Jonathan Ross, the creator of Google's TPU, claimed that its new products could threaten Nvidia.

Apple's embrace of Google is another step forward in Nvidia's nightmare: in addition to developing its own chips, a financially strong technology giant joined the AI ​​war, but did not choose Nvidia.

In addition to competition, Nvidia is also facing increasingly strong regulatory pressure. First, in July, the French Competition Authority confirmed that it was investigating Nvidia's suspected violations of market competition. Then, in August, the US Department of Justice launched two antitrust investigations against Nvidia.

Nvidia's stock price rose 150% in the first six months of this year. However, the Federal Reserve lagged behind in interest rate cuts, and Wall Street also put heavy pressure on technology stocks. Under the influence of multiple factors, Nvidia's stock price "finally" plummeted in July, accounting for four of the eight largest market value declines.

Under such circumstances, Nvidia needs and must maintain its influence in the Chinese market.Although this wave of AI is booming, the road ahead is now shrouded in fog. No one can predict whether Nvidia will hit its ceiling or even fall from the altar as the "curse" of the "AI bubble theory" would have it.

Exactly four years ago, in August 2020, Nvidia released its second quarter financial report for fiscal year 2021, and quarterly data center revenue exceeded the gaming business for the first time. Today, data centers have replaced gaming and become Nvidia's core business. But Huang Renxun has been paving the way for this step for many years.

Today, Nvidia also needs to make plans for the future, maintaining confidence while also keeping a hand. An interesting statistic is that according to Nvidia's first quarter financial report for fiscal year 2025, the automotive business accounted for only 1.2% of total revenue in that quarter, but it was the only business outside of Nvidia's data center business that achieved month-on-month growth. Among Nvidia's officially announced automotive business partners, Chinese automakers and intelligent driving solution providers accounted for more than 80%.

From this perspective, the significance of China's special AI chips may not only be in AI, but also in that Nvidia needs to have a sustained influence in China that can serve as a fulcrum for the future, even if it faces the fate of being castrated again and again.

References:

Silicon Research Lab: "China's revenue share plummets, US chip ban "hurts" Nvidia"

Cyber ​​​​Auto: "Thanks to Chinese car companies, Nvidia has recovered"

Titanium Media: "Behind Nvidia's "roller coaster" stock price, the status of the AI ​​chip giant is not unshakable"

IT Times: "Why is the Nvidia H20, which has "castrated" performance, so popular?"

Semiconductor Industry Observation: "A 3 trillion chip company is also struggling to survive?"