EN / 中文

NVIDIA B300 GPU Price Surges to 22 Million Yuan as US Export Controls Tighten China Supply

by leifengwang·November 19, 2026

Author | Chen Yuelin, Editor | Bao Yonggang

“B300 spot transaction prices have exceeded CNY 22 million, and the risk of non-performance in overseas procurement is rising. ”

"We don't even dare to place futures orders now." Cheng Lin, who handles GPU procurement in Shenzhen, summarized this year's situation in a single sentence.

From sudden resignation rumors, to reports of his return, to his public expression of willingness to come back—while the board explicitly stated no intention of reinstating him during the earnings call—the news surrounding Jensen Huang’s departure and potential return has reversed several times within just five days.

Multiple sources previously told Leiphone that it would be difficult for Jensen Huang to resume leadership at NVIDIA in the short term.

However, for Chinese computing service providers relying on NVIDIA GPUs, a more immediate concern exists: Even if Jensen Huang returns, will high-end GPUs still be available? And how can they be monetized once acquired?

Domestic channels for purchasing NVIDIA GPUs are narrowing, and renting computing power overseas is not easy either. Cheng Lin’s team attempted to acquire computing resources through self-built data centers and investments in overseas facilities, but construction cycles typically last two to three years. As US regulations on the transfer of advanced GPUs tighten, some projects under construction have begun re-evaluating their strategies.

Another option he considered—Huawei Ascend 950—has reportedly reached full order capacity, with delivery timelines also facing uncertainty.

01、Jensen Huang Wants to Do Business in China, But the US Is Tightening Every Path

"If business generates profit, why doesn’t [Jensen Huang] lobby the US government to ease controls?" Jason Lee, a computing service provider, stated bluntly.

But today, Jensen Huang’s commercial ambitions are colliding head-on with the increasingly stringent US chip export controls targeting China.

Since last year, US lawmakers have advanced four bills concerning the export and transfer of advanced chips, covering export approvals, chip location verification, semiconductor manufacturing equipment export controls for allies, and monitoring remote access to controlled chips via cloud services.

Currently, bipartisan members of Congress are pushing to include these bills in the final text of the Fiscal Year 2027 National Defense Authorization Act (NDAA). Based on current progress, the NDAA could reach floor votes in both chambers as early as December this year.

US regulation of advanced AI chips is extending beyond equipment exports to post-sale transfers, which is why Jason believes Jensen Huang’s departure has limited impact.

"Whether or not Jensen Huang continues as CEO, it will not change the direction of tightening US supply of GPUs to China," Jason pointed out, noting that this reflects US concerns about China’s AI capabilities and underscores the broader technological competition between the two nations.

Semiconductor industry analyst Xiao Du noted frankly that perhaps even Jensen Huang himself did not fully anticipate how rapidly AI has evolved in recent years, transforming from ordinary industrial technology into a strategic capability closely tied to national security.

Many interviewees believe risks were inevitable; they simply did not expect Jensen Huang to be the first target.

In Xiao Du’s view, compared to other chip companies that had incentives to pursue the Chinese market but remained relatively cautious regarding China-related operations, NVIDIA under Jensen Huang’s leadership was more aggressive, even appearing to "pretend ignorance while knowing the rules."

Xiao Du added: "Although many domestic companies participate in high-end GPU supply chains, only a few teams truly control core channels and resources." This means that if US authorities intend to investigate, they can easily trace the relevant interest chains.

Therefore, whether Jensen Huang’s past aggressive tactics in the Chinese market will lead to further accountability ultimately depends on how far the US government intends to push its regulatory efforts.

Under these circumstances, NVIDIA’s own compliance management has clearly intensified.

On one hand, before Jensen Huang’s resignation, foreign media reported that NVIDIA began requiring customers in Southeast Asia seeking cloud partner status to integrate with the Fleet Intelligence system, which collects information such as AI accelerator deployment locations.

On the other hand, NVIDIA recently established a dedicated committee under its board focused on overseeing national security and chip export control matters for the first time.

02、B300 Prices Skyrocket to CNY 22 Million: Profits and Breaches in Gray Channels

Policy shifts have already alerted some computing service providers.

Jason recalled becoming cautious about Southeast Asian GPU transit channels starting in the first half of this year. Facing channel partners claiming they could move GPUs into China via Malaysia, his attitude was clear: First, actually deliver the equipment, then discuss subsequent cooperation.

Subsequent reports of Malaysian customs seizing GPU servers further confirmed the risks of such transactions. In Jason’s view, US oversight of third-country equipment transfers is strengthening, making previously viable transit channels increasingly unreliable.

Yet despite rising cross-border transaction risks, some remain willing to take chances. "No one does unprofitable business; someone always dares to do 'life-threatening' business," commented an executive at an AI chip company.

Xiaoyu, who has been in the computing service business for two years, calculated for Leiphone: The factory price of a B300 is approximately CNY 4 million, actual overseas procurement costs range from CNY 7 to 8 million, and once brought into China, market quotes immediately double.

Where does this huge price gap come from? First, GPU order quotas at overseas data centers themselves hold trading value.

Xiaoyu explained to Leiphone that overseas data centers purchasing high-performance GPUs typically must submit company backgrounds, facility conditions, and end-use details to upstream suppliers, awaiting KYC (Know Your Customer) review.

Passing the audit is merely a prerequisite. Xiaoyu added that the number of GPUs a facility can accommodate also depends on infrastructure conditions such as power capacity.

Furthermore, because the US directly restricts advanced GPU exports to China and strictly monitors flows, goods entering gray channels incur high compliance risks, which translate into premiums.

Jason revealed that before the Jensen Huang controversy, the domestic spot price for a single B300 had already climbed to around CNY 22 million.

Markups at every stage—from the equipment itself to procurement channels to formal entry—not only drive up GPU market prices but have also caused buyers like Xiaoyu to suffer losses.

Previously, his team planned to procure dozens of B300s from overseas and even pre-purchased supporting equipment like switches and network cards, but the supplier ultimately delivered only two units.

Xiaoyu believes the problem lies precisely in the low cost of breach. As market quotes continue to rise, the penalty suppliers pay for abandoning original orders may be far lower than the profits gained from reselling the equipment elsewhere.

03、Pivoting to Overseas Computing Sales: Hitting the CNY 12 Million Threshold and Data Center Operations

Facing increased difficulties in cross-border high-end GPU acquisition and breach risks, Xiaoyu’s team is now focusing more on another business model: keeping GPUs in overseas data centers and serving local and domestic clients through computing power leasing and token-based billing.

However, leaving GPUs overseas does not guarantee easy sales of computing power.

Xiaoyu explained that local AI ecosystems and computing demand in places like Malaysia and Thailand are relatively limited. Some facilities have computing capacities exceeding local absorption capabilities, meaning domestic clients still dominate the demand market.

But this path is no easier than "bringing equipment into China," as the end-users and purposes of overseas computing services also fall under US regulatory scrutiny.

Foreign media previously reported plans by the US to include seven countries—Malaysia, Singapore, Thailand, Indonesia, Norway, Brazil, and South Africa—in advanced AI chip export licensing regimes.

Recently, the US Department of Commerce’s Bureau of Industry and Security (BIS) published final rules listing five countries, excluding Norway and South Africa.

Beyond the reduced list, new rules set a scale threshold for remote computing: If an overseas data center provides more than 12 million TPP (Total Processing Performance) of computing power to enterprises headquartered or whose ultimate parent company is headquartered in restricted areas, the export, re-export, and in-country transfer of related advanced chips require licenses.

The current restricted scope covers Group D:5 of the US Export Administration Regulations (EAR), including Mainland China and Hong Kong, China, as well as Macau, China. However, even if direct clients are not listed, if computing power ultimately flows to restricted enterprises, license requirements may still be triggered.

What scale of computing power does 12 million TPP represent? According to BIS TPP calculation rules, it is equivalent to approximately 95 8-GPU H200 SXM servers (FP8) or 27 8-GPU B300 servers (FP4).

With compliant machines becoming increasingly scarce, ensuring stable operation of dozens or even hundreds of high-end servers requires significant effort in data center construction and operations.

Xiaoyu cited examples of containerized data centers built in partnership with domestic companies in Southeast Asia in the past. Some projects primarily focused on GPU resale, offering limited practical usability.

For such facilities to continue operating, they need stable power supplies, professional maintenance capabilities, and the ability to consistently provide computing services externally to truly keep the machines running.

04、A Card-Short Market: Domestic GPUs May Not Benefit

The US tightening of NVIDIA’s high-end GPU supply to China theoretically creates market space for domestic GPUs, but current positive expectations are largely reflected in stock prices, while industry confidence has not kept pace.

Whether domestic GPUs can meet market demand depends on specific application scenarios.

The training segment remains difficult to penetrate. Multiple interviewees stated that even if NVIDIA GPU prices rise due to supply constraints, large tech firms and model companies are still willing to pay premiums as long as the hardware supports frontier model training.

In contrast, the inference market, where domestic GPUs might find easier traction, may not necessarily be economically viable.

Some interviewees observed that starting with the H-series, although NVIDIA graphics cards have higher unit prices, when calculating based on final inference output, the unit cost may actually be lower than that of some cheaper domestic cards.

Industry analyst Qin Qing explained that profitability in inference businesses heavily relies on token selling prices and the computational costs required to produce those tokens. If domestic cards cannot achieve advantages in HBM costs and computational efficiency, commercialization faces pressure even if per-card procurement prices are lower.

According to Qin Qing, since domestic companies rarely sign contracts directly with foreign manufacturers, procurement costs for HBM in China are at least twice the normal overseas prices.

Meanwhile, domestic HBM supply awaits new capacity releases. However, after research, Xiao Du believes CXMT’s capacity will likely not become a new variable until at least next year.

Pressure extends beyond this; logic chip manufacturing and advanced packaging capacity also plague domestic GPU vendors.

Xiao Du pointed out that chip design companies should theoretically command higher premiums based on architecture, software, and design capabilities. However, before domestic chipmakers can convert these capabilities into competitive advantages, they are already constrained by manufacturing capacity and HBM supply, resulting in a situation where supply chain premiums outweigh design capability premiums.

Coordinated supply chain operations will remain an unavoidable collective test for domestic GPU manufacturers over the coming year.

* All interviewees in this article remain anonymous.