Nvidia is hoping to ship tens of thousands of its advanced H200 artificial intelligence chips to China before the Lunar New Year in mid-February, Reuters reported on Tuesday, citing sources familiar with the matter.
It could mark a tentative reopening of a market it once dominated but was forced to abandon under U.S. export controls.
The company is aiming for an initial delivery of between 40,000 and 80,000 H200 chips, largely sourced from existing inventory and current production capacity, according to the sources. Nvidia is also planning to ramp up H200 output in the second quarter of next year to meet potential demand from Chinese customers, the sources added.
Nvidia declined to comment as of the press time.
Whether the plan materialises remains highly uncertain. As of now, there is no public indication that Chinese authorities have approved procurement plans related to the H200, while the U.S. government has only just begun a formal interagency review of potential exports.
Still, even the prospect of renewed shipments has sent ripples through China’s technology sector, global semiconductor markets and Washington’s policy circles, underscoring the strategic weight Nvidia’s chips carry in the global AI race.
Until recently, Nvidia was the undisputed backbone of China’s AI boom. Before Washington tightened export controls, the company commanded an estimated 95% share of China’s AI accelerator market, supplying chips that powered everything from internet giants’ large language models to government-backed computing centres.
That dominance collapsed after the United States imposed sweeping restrictions on advanced chip exports to China, a policy initiated under former President Donald Trump and expanded by subsequent administrations. Nvidia gradually withdrew its most powerful products from the Chinese market, replacing them with downgraded, China-specific versions such as the H20, which were designed to comply with U.S. rules but offered weaker performance.
Then came additional headwinds. Chinese regulators raised concerns over potential “security backdoors” in foreign AI chips, summoning Nvidia executives in July for talks over alleged vulnerabilities in the H20. Nvidia denied the accusations, saying it had never embedded backdoors or surveillance functions in its products, but the episode further clouded its prospects in China.
By Nvidia CEO Jensen Huang’s own admission, the company’s AI chip market share in China fell “from 95% to effectively zero.”
The situation shifted abruptly on Dec. 8, when Trump said in a social media post that, while safeguarding U.S. national security, he would allow Nvidia to ship H200 chips to China — provided that 25% of the resulting sales revenue was paid to the U.S. government.
The proposal amounts to a hefty levy, but it represented a major policy signal. More importantly for Nvidia, the approved product was upgraded from the H20 to the far more capable H200, a flagship chip in the Hopper family that remains competitive globally.
Nvidia later said that supplying H200 chips to commercial customers was “a positive move,” though it did not comment on the specifics of any China-bound shipments.
Last week, Reuters reported that the Trump administration had formally launched a review process for H200 exports, with the U.S. Commerce Department submitting export licence applications to the State Department, the Department of Energy and the Department of Defense. While the process is opaque, it is expected to take around 30 days, and Trump is understood to retain the final decision-making authority.
Opposition and scrutiny
Despite signs of momentum, the path forward is fraught with political risk.
Several U.S. lawmakers have voiced opposition to allowing advanced AI chips into China. Democratic Senator Elizabeth Warren and Representative Gregory Meeks called on the Commerce Department to disclose more details about any export licences granted to Nvidia, arguing that the chips could be diverted to military uses and that Washington must consider the reaction of U.S. allies.
There have also been reports that U.S. officials discussed requiring Nvidia to install tracking mechanisms on chips exported to China — a measure that could reignite security concerns on the Chinese side and revive allegations of hidden surveillance capabilities.
For Chinese regulators, such conditions would be highly sensitive. The July probe into Nvidia’s H20 chips remains unresolved publicly, and any suggestion of monitoring or control features could complicate approvals for new imports.
Despite the uncertainty, several signals suggest Beijing is not closing the door.
After Trump’s comments earlier this month, Chinese Foreign Ministry spokesperson Guo Jiakun said China had “taken note of relevant reports” and reiterated that Beijing supports mutually beneficial cooperation with the United States.
More broadly, Chinese policymakers have repeatedly stressed that while the country is accelerating efforts to localise critical technologies, it remains open to importing advanced foreign products that comply with laws and regulations.
That message was reinforced last week by the high-profile visit of AMD Chair and CEO Lisa Su to China, where she met customers, partners and senior officials. AMD has said its MI308 AI chip has already received a U.S. export licence and that it is prepared to pay tariffs to ship products to China.
China’s Ministry of Industry and Information Technology and Ministry of Commerce both publicly disclosed meetings with Su, with officials expressing support for deeper cooperation and signalling that foreign chipmakers are still welcome in the market.
Who would buy?
If Nvidia does manage to ship H200 chips, the first buyers are expected to be China’s largest technology firms, including Alibaba, ByteDance and Tencent, according to people familiar with the matter.
These companies are locked in intense competition over large language models, AI infrastructure and cloud services, where training performance remains critical. While China has made rapid progress in deploying domestically produced inference chips for government-backed projects such as intelligent computing centres and the Xinchuang initiative, training cutting-edge models at scale still favours Nvidia’s ecosystem.
Several Chinese firms have already held preliminary discussions with Nvidia about potential purchases since Trump’s announcement, the sources said. Under earlier arrangements, companies seeking H20 chips were required to submit detailed explanations to regulators, a process that is likely to be repeated — or intensified — for H200s.
Strong demand is also evident in less transparent channels. Industry sources say some Chinese firms have sought access to Nvidia chips via third countries such as Japan and Singapore, or have shifted training workloads overseas to data centres equipped with Nvidia hardware.
What’s at stake financially?
By 2025, Nvidia’s main U.S. shipments will have largely transitioned to its next-generation Blackwell architecture, meaning H200 production capacity is expected to shrink and pricing could become volatile.
Based on industry estimates of $20,000 to $40,000 per H200 card, and factoring in the proposed 25% revenue share payable to the U.S. government, the first batch of shipments to China could generate between $1 billion and $4 billion in gross sales.
That would be a sharp contrast to Nvidia’s recent performance in China. In its latest quarterly results, the company disclosed just $50 million in Hopper-series sales to the Chinese market, with no major data centre orders. Its revenue guidance for the current quarter does not include any contribution from China data centres.
At its peak, revenue from Greater China reached $17.1 billion in fiscal 2024. While Nvidia’s retreat has opened space for domestic competitors, analysts still expect China’s AI chip market to grow by around 30% year on year, underscoring the scale of the prize.
Does H200 still matter?
Technically, the H200 retains a clear edge over most Chinese alternatives, particularly in large-scale model training. Its advantages in memory capacity, bandwidth and interconnect performance remain significant, and its compatibility with Nvidia’s CUDA software ecosystem reduces migration costs for developers.
Most of China’s early large language models were built on Nvidia hardware, deeply embedding CUDA into workflows. Switching to domestic chips often requires costly rewrites and optimisation.
That advantage, however, is narrowing.
Domestic GPU makers have made aggressive strides. On Dec. 20, Moore Threads — often dubbed China’s first publicly listed GPU company — unveiled its next-generation “Huagang” architecture, claiming major gains in computing density and energy efficiency.
The company said its new “Hushan” training and inference chip surpasses Nvidia’s Hopper series across multiple metrics, including floating-point performance and memory bandwidth. Moore Threads founder and CEO Zhang Jianzhong, a former Nvidia global vice president, said users migrating from Hopper would see results “only better, not worse.”
Moore Threads and peers are also racing to build software ecosystems such as MUSA to challenge Nvidia’s CUDA dominance — a strategic front Nvidia itself has long recognised as critical.
Jensen Huang has previously acknowledged that exporting lower-end chips to China was unlikely to succeed long term, and has questioned whether China would still accept H200s once domestic alternatives mature.
Nvidia is not alone in eyeing a return. AMD and Intel are also seeking ways back into the Chinese market under the same restrictive policy environment.
AMD, the world’s second-largest GPU supplier, derived about 24% of its revenue from China in 2024, but its China business has cooled sharply since mid-year. With export approval for the MI308 and an apparent willingness to absorb tariffs, AMD appears to be moving quickly.
Intel, though less competitive in high-end AI accelerators, has also signalled interest in maintaining its China footprint, particularly in data centre CPUs and edge computing.
For now, Nvidia’s H200 remains the focal point — a test case for whether commercial logic can still carve out space amid geopolitical rivalry.
As one industry executive put it: “Even if the door opens, it will only open a crack. But for companies building the next generation of AI, even a crack matters.”
