What happened
Reuters, citing two sources familiar with the matter, reported on 5 November that roughly 600,000 Nvidia H200 AI chips are stored in Hong Kong and cannot be moved into mainland China. The United States has lifted its export ban on the H200 for approved Chinese customers, but Chinese regulators have not granted permission for bulk shipments. Small, broker‑mediated deliveries and limited approvals for training‑only use have been allowed; major internet firms such as Alibaba, ByteDance and Tencent each received about 10,000 units under those narrow licences. The bulk of the chips remain in limbo, creating a bottleneck for Chinese AI developers who rely on Nvidia’s high‑performance hardware for large‑scale model training.
On 5 November, Reuters reported that about 600,000 Nvidia H200 AI accelerators are currently piled up in Hong Kong, unable to enter mainland China. The H200, while not Nvidia’s newest Blackwell‑era chip, remains a top‑tier processor for training large language models and other data‑intensive AI workloads.
The United States lifted its export ban on the H200 after the 2023 trade talks, allowing sales to Chinese customers who obtain a licence from the U.S. Commerce Department. However, Chinese authorities have not granted large‑scale import approvals, citing a policy that prioritises domestic chip manufacturers and seeks to limit reliance on foreign AI hardware for workloads.
Limited licences have been issued to major Chinese internet firms for training‑only use, with each company receiving roughly 10,000 units. Smaller quantities have entered via brokers or indirect channels, but the bulk of the chips remain stored in Hong Kong awaiting clearance.
The stalled imports have created a secondary market where servers equipped with eight H200 chips are listed for more than 6 million yuan, a price roughly three times higher than a year earlier. Similar pressure is seen for Nvidia’s next‑generation Blackwell chips, which command even higher premiums.
Source details: biz.chosun.com
Why it matters
The impasse underscores a growing tension between China’s need for world‑leading AI hardware and its policy goal of fostering a domestic chip ecosystem. Nvidia’s H200 remains far more powerful than any Chinese‑made alternative, so Chinese AI firms must either wait for official import clearance or pay premium prices on the gray market. Reuters noted that a server equipped with eight H200s was listed for more than 6 million yuan in October, roughly three times the price a year earlier. Higher hardware costs could slow the rollout of new AI services, increase operating expenses for firms like Alibaba and ByteDance, and pressure the Chinese government to either relax import rules or accelerate its own chip development. The situation also illustrates how overlapping U.S. export controls and Chinese self‑reliance measures can create supply‑chain friction, potentially reshaping the global AI hardware market.
China’s AI sector depends heavily on high‑performance GPUs for training state‑of‑the‑art models. Without access to Nvidia’s H200, domestic firms must either accept slower training cycles or pay inflated prices on the gray market, raising overall development costs.
The policy tension highlights the broader strategic dilemma: allowing imports of superior foreign chips accelerates AI progress but undermines the government’s goal of building a self‑sufficient chip industry. If domestic alternatives cannot match Nvidia’s performance soon, the restriction could hamper China’s competitiveness in AI services and research.
The price surge for H200‑based servers signals a tightening supply chain that could ripple through downstream AI applications, from recommendation engines to large language models. Higher hardware costs may be passed to consumers or reduce profit margins for Chinese tech giants.
The situation also serves as a case study of how overlapping export controls (U.S.) and import restrictions (China) can create unintended bottlenecks, potentially prompting other countries to reassess their own AI hardware policies.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Analysts will be watching whether Beijing expands the limited‑use licences for H200 chips beyond training, how quickly domestic alternatives can close the performance gap, and whether the gray‑market price premium stabilises or spikes further. Future policy announcements from the Ministry of Industry and Information Technology, as well as any new U.S. licensing decisions, could alter the supply dynamics. Additionally, the pace of Chinese chip projects such as the home‑grown “Kunpeng” and “Ascend” families will be a key indicator of whether the country can reduce its dependence on Nvidia hardware in the medium term.
Any official statement from China’s Ministry of Industry and Information Technology expanding the scope of approved H200 imports, especially for workloads.
Updates from the U.S. Commerce Department on licensing criteria for Nvidia’s AI chips, which could either ease or tighten the flow of hardware to China.
Progress reports on Chinese domestic AI chip projects (e.g., Huawei’s Ascend, Alibaba’s Pingtouge) that aim to close the performance gap with Nvidia’s offerings.
Trends in gray‑market pricing for H200 and Blackwell servers, which will indicate whether supply constraints are easing or worsening.