젠슨 황의 엔비디아 내 지위는 아직 확정되지 않았다. 그는 돌연 사임했고, 이후 복귀 가능성을 시사하는 보도가 나왔으나, 이사회는 실적 발표 콘퍼런스콜에서 그를 복직시킬 계획이 없다고 밝혔다. 소식통에 따르면 단기간에 그가 엔비디아를 다시 이끌 가능성은 낮으며, 중국 GPU 구매업체들은 선주문이 동결됐다고 전했다.
“Right now, nobody dares to place futures orders.” Cheng Lin, who handles GPU procurement in Shenzhen, summed up this year's situation in a single sentence.
In just five days, news about whether Jensen Huang would stay or go reversed several times: he suddenly resigned, reports of his return followed, he publicly expressed his willingness to come back, and then the board stated clearly during the earnings call that it had no intention of reinstating him.
Multiple sources had earlier told Leiphone that it would be very difficult for Huang to take back control of NVIDIA in the short term.
For Chinese compute service providers whose business depends on NVIDIA GPUs, however, the more immediate question is this: even if Huang returns, can high-end GPUs still be bought? And once bought, how can they be turned into revenue?
The channels for buying NVIDIA GPUs in China are narrowing, and renting compute overseas is not easy either. Cheng's team has tried to secure compute resources by building its own facilities and investing in overseas data centers, but data center construction takes two to three years. As U.S. oversight of the movement of advanced GPUs tightens, some projects under construction have begun to reassess.
The other path Cheng has considered is Huawei's Ascend 950. As far as he knows, its order books are already full, and delivery timing is likewise uncertain.
01
Jensen Huang wants China business; the U.S. is tightening every route
"Doing business can make money. Why doesn't (Huang) push the U.S. government to relax its controls?" said Jason Lee, a compute service provider, bluntly.
Today, however, Huang's commercial goals are colliding head-on with increasingly tight U.S. chip controls on China.
Since last year, members of the U.S. Congress have advanced four bills targeting the export and movement of advanced chips. They cover export licensing, chip location verification, export controls on allied semiconductor manufacturing equipment, and oversight of remote access to controlled chips through cloud services and similar means.
Members of both parties in Congress are now pushing to include these bills in the final text of the National Defense Authorization Act (NDAA) for fiscal year 2027. At the current pace, the NDAA could reach a House floor vote as early as this December.
U.S. regulation of advanced AI chips is extending beyond the point of export to the movement of equipment after it ships. This is why Jason believes Huang's departure will have limited impact.
"Whether or not Huang continues as CEO, it will not change the direction of the U.S. tightening the supply of GPUs to China," Jason said. He noted that this reflects American concern about China's AI capabilities, which is tied to the broader technology competition between the two countries.
Chip industry analyst Xiao Du said that even Huang himself probably did not fully anticipate how fast AI would develop in recent years, rising from an ordinary industrial technology to a strategic capability closely linked to national security.
Many interviewees believed the risk would arrive sooner or later, but none expected Huang to be the first to be "singled out."
In Xiao Du's view, compared with NVIDIA, other chip companies also had reasons to pursue the Chinese market but had been relatively cautious about their China business. Under Huang's leadership, NVIDIA was more aggressive, and at times seemed to "know perfectly well but play dumb."
Xiao Du added: "Although many domestic companies take part in high-end GPU supply, only a handful of teams really control the core channels and resources." This means that if the U.S. decides to investigate, it would not be hard to trace the threads and uncover the chain of interests involved.
Therefore, whether Huang's aggressive approach to the Chinese market will bring further accountability ultimately depends on how far the U.S. government is willing to push its oversight.
Against this backdrop, NVIDIA's own compliance management has clearly been tightened.
On one hand, even before Huang's resignation, foreign media had reported that NVIDIA had begun requiring some Southeast Asian customers seeking cloud partner status to connect to a Fleet Intelligence system capable of collecting information such as where AI accelerators are deployed.
On the other hand, NVIDIA recently set up, for the first time, a dedicated board committee focused on overseeing national security and chip export control matters.
02
A B300 sold for 22 million yuan: the profits and defaults of the gray channels
Changes at the policy level have already alarmed some compute service providers.
Jason recalled that he had been cautious about Southeast Asian GPU transshipment channels since the first half of this year. When dealing with suppliers who claimed they could bring GPUs into China through Malaysia, his position was clear: get the equipment actually delivered first, and only then discuss further cooperation.
Later, reports began to circulate that Malaysian customs had seized GPU servers, which made him even more certain of the risks in such transactions. In Jason's view, U.S. oversight of the movement of equipment through third countries is strengthening, and transshipment routes that once seemed workable are becoming hard to rely on.
But even as the risks of cross-border trade grow, some people are still willing to take chances. "No one does business that loses money, but someone always dares to do business that could cost them their head," one executive at an AI chip company commented.
Xiao Yu, who has worked in the compute service business for two years, calculated the economics for Leiphone: the ex-factory price of a B300 is about 4 million yuan, while the actual overseas procurement cost reaches 7 to 8 million yuan. Once the chip is moved into China, the market quote immediately doubles.
Where does such a huge price gap come from? First, the GPU purchase quota of an overseas data center itself has trading value.
Xiao Yu explained to Leiphone that when overseas data centers buy high-performance GPUs, they typically must submit materials to upstream suppliers covering corporate background, facility conditions, and end use, and then wait for a KYC (Know Your Customer) review.
Passing the review is only a precondition. Xiao Yu added that the number of GPUs a data center can take on also depends on infrastructure conditions such as its own power capacity.
On top of this, because the U.S. directly restricts exports of advanced GPUs to China and strictly investigates their destinations, once the supply enters gray channels, the high compliance risk is converted into a premium.
Jason revealed that before the Huang turmoil, the spot price of a single B300 in China had climbed to around 22 million yuan.
Layer after layer of markups, from the equipment itself to procurement channels and then to formal entry into the country, pushed up GPU market quotes, and also left buyers like Xiao Yu taking a loss.
Earlier, his team had planned to purchase dozens of B300s overseas and had even bought switches, network cards, and other supporting equipment in advance. In the end, the supplier delivered only two units.
Xiao Yu believes the problem lies precisely in how low the cost of default is. As market quotes keep rising, a supplier that abandons an existing order may pay a penalty far lower than what it can earn by reselling the equipment.
03
Turning to overseas compute rentals: the 12-million-TPP threshold and data center operations
As high-end GPUs become harder to move across borders and default risks rise, Xiao Yu's team is now focusing more on another business model: keeping GPUs in overseas data centers and serving local and Chinese customers through compute rental and token-based billing.
However, compute left overseas is not necessarily easy to sell.
Xiao Yu explained that the local AI ecosystems and compute demand in Malaysia, Thailand, and similar places are relatively limited. Some data centers already have compute capacity exceeding what local markets can absorb, which means Chinese customers still dominate demand.
But this route is no easier than moving equipment into China. The end customers and uses of overseas compute services are also coming under U.S. regulatory scrutiny.
Foreign media previously reported that the U.S. planned to bring seven countries, including Malaysia, Singapore, Thailand, Indonesia, Norway, Brazil, and South Africa, under export licensing for advanced AI chips.
Recently, the new rules published by the U.S. Department of Commerce's Bureau of Industry and Security (BIS) ultimately listed five countries; Norway and South Africa were not included.
Beyond the shorter list, the new rules also set a scale threshold for remote compute. If an overseas data center provides more than 12 million TPP (total processing performance) of compute to companies headquartered, or whose ultimate parent is headquartered, in restricted jurisdictions, the export, re-export, and in-country transfer of the relevant advanced chips will require a license.
The restricted scope currently covers Country Group D:5 under the U.S. Export Administration Regulations (EAR), which includes mainland China and Hong Kong, as well as Macau. Even if a direct customer is not on this list, a license requirement may still be triggered if the compute ultimately flows to a restricted company.
How large is 12 million TPP in practical terms? Based on BIS's TPP calculation rules, it is roughly equivalent to about 95 eight-GPU H200 SXM servers (FP8) or 27 eight-GPU B300 servers (FP4).
As compliant-use machines become harder to obtain, keeping dozens or even hundreds of high-end servers running stably is itself difficult, given the construction and operation of the data centers.
Xiao Yu gave an example: in the past, Southeast Asia saw some modular data center projects built in cooperation with Chinese companies, and some of these were mainly centered on GPU resale, with limited real usability.
If such facilities are to keep operating, they need a stable power supply, professional operations and maintenance capabilities, and the ability to keep providing compute services to outside customers, so that the machines can actually run.
04
A market short on cards: domestic GPUs may not get a share
Tighter U.S. restrictions on NVIDIA's high-end GPUs in China theoretically open up market space for domestic GPUs. For now, however, the positive expectations are reflected more in stock prices than in industry confidence.
Whether domestic GPUs can capture this demand depends on the specific application scenario.
Training remains hard to crack. Several interviewees said that even as NVIDIA GPU prices rise due to supply constraints, large technology companies and model developers are still willing to pay a premium as long as the chips can support frontier model training.
By comparison, the inference market, which domestic GPUs find easier to target, may not add up economically.
One interviewee observed that since the H series, NVIDIA cards have been more expensive per unit, but when calculated by final inference output, their unit cost can actually be lower than that of some cheaper domestic cards.
Industry analyst Qin Qing explained that the profitability of inference businesses depends heavily on token prices and the compute costs required to produce those tokens. If domestic cards cannot achieve an advantage in HBM cost and computing efficiency, they will face commercial pressure even with lower procurement prices per card.
According to Qin Qing, because domestic companies can rarely sign directly with overseas original manufacturers, the procurement cost of some domestic HBM is at least twice the normal overseas purchase price.
At the same time, domestic HBM supply is still waiting for new capacity to come online. But after his research, Xiao Du believes that CXMT's capacity will not become a new variable until next year at the earliest.
The pressure does not stop there. Logic chip manufacturing and advanced packaging capacity are also troubling domestic GPU makers.
Xiao Du pointed out that chip design companies should, in theory, be able to earn higher premiums through their architecture, software, and design capabilities. But before those capabilities can be turned into market competitiveness, domestic chip companies have been held back by manufacturing capacity and HBM supply, producing a situation in which "supply chain premiums outweigh design capability."
Coordinated operation of the supply chain will remain a collective test that domestic GPU makers cannot avoid in the coming year.
*All interviewees in this article are anonymous.