Nvidia CEO Jensen Huang (C) speaks to media after a fireside chat event during the China International Supply Chain Expo (CISCE) in Beijing on July 17, 2025.
JADE GAO/AFP via Getty Images
Small batches of Nvidia’s H200 artificial intelligence processors entered mainland China in recent weeks, with ByteDance and Tencent each receiving roughly 10,000 units — the first meaningful deliveries since President Trump approved their export in December 2025. The arrival ends more than eight months of delivery paralysis, during which Washington had cleared the sale while Beijing sat on its hands. But 10,000 chips against a legal ceiling of 75,000 per buyer represents barely 13% of what each company is authorized to receive — and Beijing’s National Development and Reform Commission is now the gatekeeper controlling whether the other 87% ever arrives.
Every future H200 purchase still requires individual NDRC sign-off through an application process that forces buyers to disclose how many chips they need and justify why they cannot use domestic alternatives. The structural result: Beijing gets ongoing visibility into which AI labs are training on Nvidia hardware, at what scale, and for what purpose — turning what looks like a market opening into an ongoing industrial-policy lever. Chinese regulators have also told companies they may ship H200 chips to Hong Kong, which sits outside mainland China’s customs border. But that route comes with its own dead end: Hong Kong’s data-center infrastructure cannot absorb the volumes that approved buyers could theoretically receive. “It’s a dilemma. Everyone needs the chips but struggles to find a way to use them in Hong Kong,” a person familiar with the situation told the Financial Times.
Several other major Chinese technology groups are expected to receive NDRC approval for shipments of comparable size in the coming weeks, according to the FT report.
Why 10,000 Chips Is a Bigger Deal Than It Sounds — and a Smaller Deal Than It Looks
The numbers carry two stories simultaneously, and readers who hear only one of them will misunderstand the event.
The bigger-than-it-sounds story: these chips are going specifically toward frontier AI model training, the most computationally demanding phase of AI development — and the one phase for which Chinese companies cannot substitute domestic hardware. Nvidia designed the H200 around 141 gigabytes of HBM3e memory running at 4.8 terabytes per second, nearly double the memory capacity of the H100 and 43% faster. That memory bandwidth is what enables training runs on models with 100 billion or more parameters without excessive overhead from moving data between chips. China’s leading domestically produced AI accelerator, Huawei’s Ascend 950PR, performs competitively for inference — serving live model outputs to users — but Chinese AI developers continue to depend on Nvidia’s Hopper architecture for the training phase, where the H200’s memory subsystem provides a performance advantage domestic chips have not yet fully matched. ByteDance and Tencent can run production AI services on Ascend chips. They need H200s to train the next generation of models that will power those services.
The smaller-than-it-looks story: 10,000 H200 chips, while meaningful, is a modest cluster by frontier-training standards. Nvidia has approximately 500,000 H200 chips staged in inventory primarily for Chinese customers, but every chip in that warehouse requires its own NDRC sign-off before it can move. The legal framework permits each approved buyer to acquire up to 75,000 H200 units. Against the 10,000 delivered, the pipeline is 87% full and locked. Beijing’s preference, meanwhile, is for approved companies to keep the majority of their licensed chip allowances outside mainland China — specifically to avoid undercutting Huawei and other domestic chipmakers competing for the same data-center infrastructure spend.
There is a compounding technical bottleneck that policy alone cannot resolve. The H200 is manufactured on TSMC’s 4-nanometer process and assembled using TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging, which bonds the GPU die and its HBM3e memory stacks onto a silicon interposer. CoWoS capacity is a binding constraint separate from TSMC’s standard wafer fabrication and requires years to expand — a hard ceiling on how quickly H200 production can scale up to meet demand regardless of what regulators in Washington or Beijing decide. Chinese firms collectively placed orders for more than two million H200 chips earlier this year against a total Nvidia inventory of roughly 700,000 units, a mismatch that forced Nvidia to approach TSMC about restarting Hopper-generation production runs.
Why China’s CUDA Dependency Explains the Demand
A question that the delivery numbers alone do not answer: why do ByteDance and Tencent need H200s when Huawei’s Ascend 950PR is commercially available in China and demonstrably capable of training frontier models? DeepSeek proved in April 2026 that the Ascend can run frontier-scale training; its V4 model was the first Chinese frontier model explicitly optimized for Huawei hardware.
The answer lives in Nvidia’s CUDA software ecosystem — a platform built over 17 years that underpins every major AI training framework including PyTorch, TensorFlow, and JAX. Transitioning a large AI research team’s entire training infrastructure from CUDA to Huawei’s competing CANN framework involves re-engineering codebases, revalidating numerical reproducibility, and retraining engineering teams in new tooling. ByteDance and Tencent have institutional knowledge, optimized codebases, and production infrastructure built on Nvidia’s software stack. They can run inference on Ascend — that requires serving existing trained weights, which is less CUDA-dependent. They cannot trivially move training to a different hardware platform without significant engineering cost and potential quality risk. The 10,000 H200 chips are almost certainly being deployed to the training clusters where switching costs are highest.
This dynamic is precisely what prior congressional critics warned about. Former deputy national security advisor Matt Pottinger warned Congress that H200 sales would provide genuine uplift to Chinese frontier AI development. Rep. Gregory Meeks, the senior Democrat on the House Foreign Affairs Committee, called the export policy a strategic mistake. Bureau of Industry and Security chief Jeffrey Kessler had characterized H200 shipments as “trivial” as recently as July 14 congressional testimony — the deliveries reported August 19 represent a meaningful uptick from that baseline, though still a fraction of the licensed volumes.
A Policy Opening Eight Months in the Making
The H200’s path to China was the product of deliberate diplomacy and a specific revenue architecture. President Trump announced the policy reversal on December 8, 2025, establishing a framework that permitted Nvidia to sell H200 chips to approved Chinese buyers on the condition that 25% of revenues from those sales would be remitted to the US Treasury, and that chips would physically route through US territory for security screening before reaching Chinese recipients.
The Bureau of Industry and Security formalized the framework on January 15, 2026, shifting H200 exports to China from a “presumption of denial” to case-by-case review — a meaningful policy liberalization that nonetheless kept the approval requirement intact. Washington cleared approximately ten Chinese companies to purchase H200 chips under the new rules, including Alibaba, Tencent, ByteDance, JD.com, and electronics distributors Lenovo and Foxconn. Each approved buyer was capped at 75,000 units.
Despite those approvals, no chips moved for months. Beijing instructed Chinese tech companies to hold back orders and direct their AI infrastructure investment toward Huawei’s Ascend line, in part because the routing-through-US-territory requirement raised suspicion among Chinese regulators about possible tampering, and in part because Beijing’s domestic-chip mandate was actively competing for the same procurement budget. During the May 2026 Trump-Xi Beijing summit, Jensen Huang attended as part of the US delegation — and left with zero orders confirmed and zero chips shipped. As recently as the July 14 congressional hearing, Kessler confirmed shipments remained “trivial.”
The August 19 FT report marks the first credible confirmation that the policy stalemate has begun to break. Lenovo and other Nvidia partners told Chinese customers the previous week that they could resume placing orders for H200-equipped servers — though each order still requires separate NDRC sign-off.
The Hong Kong Complication
Beijing’s preferred deployment geography for the licensed H200 chips illustrates the contradictions at the center of its AI hardware policy. Chinese regulators have told approved companies they can ship the chips to Hong Kong — a separate customs territory from mainland China — and deploy them there. US export licenses cover both mainland China and Hong Kong destinations, giving companies an additional legal route.
But Hong Kong’s data-center infrastructure cannot absorb the volumes those licenses theoretically permit. The territory lacks power capacity and physical rack space to house tens of thousands of H200 chips at scale, and power supply expansion faces the same multi-year lead times as any large-scale infrastructure build-out. The practical result: companies are legally authorized to buy more chips than Hong Kong can house, while being actively steered away from mainland data centers where their infrastructure actually sits.
This reflects Beijing’s core dilemma in domestic chip policy. Chinese companies including Alibaba, Tencent, ByteDance, and Baidu have invested heavily in developing their own AI accelerator chips — ByteDance’s custom ASICs, Tencent’s inference processors, Alibaba’s T-Head line — alongside purchases of Huawei’s Ascend series. Domestic chip producers claimed approximately 50% sourcing mandates for data centers by August 2025, with state-funded projects entirely barred from foreign accelerators by November 2025. A third-party survey by TrendForce published August 10 projected that domestic Chinese solutions are on track to capture nearly 90% of China’s hardware market by 2026, combining domestically produced GPUs from Biren, Cambricon, and Huawei alongside custom ASICs from the major internet platforms.
Beijing allowing 10,000 H200 chips into mainland China is a pragmatic concession that domestic hardware cannot yet fully cover frontier model training at the scale ByteDance and Tencent require. Letting those same chips flow in at full licensed volume — 75,000 apiece — would signal something closer to a policy retreat.
What It Means for Nvidia and the AI Competition
For Nvidia, the deliveries offer tangible if partial relief. The company held approximately 500,000 H200 chips in inventory for Chinese customers throughout the first half of 2026, representing substantial tied-up working capital. Nvidia shares edged higher in pre-market trading on August 19 following the FT report, breaking a three-day losing streak.
The longer trajectory remains complicated. Before export controls tightened, China generated $17.1 billion in annual Nvidia revenue and hosted roughly 95% of the country’s advanced AI chip market. Nvidia’s own FY2026 annual filing described the company as “effectively foreclosed from competing” in China’s data center computing market. The H200 window is now open at a trickle, but Nvidia’s most advanced hardware — the Blackwell-generation GPUs — remains banned from Chinese export, and May 2026 Commerce Department guidance closed the Blackwell loophole that had allowed Chinese-owned subsidiaries outside mainland China to acquire Blackwell chips without triggering controls.
The geopolitical context is intensifying on both sides. Chinese President Xi Jinping launched the World AI Cooperation Organisation in July 2026, a Beijing-led multilateral body promoting China’s open-weight AI technology as an alternative to US-dominated AI frameworks. Washington simultaneously prepared to pressure 35 countries to align with the US-led AI framework, and the Pentagon added Alibaba, Baidu, and BYD to its list of companies linked to Beijing’s military capabilities.
The chips that arrived in Beijing’s data centers over recent weeks are real, the first of their kind, and genuinely meaningful for the frontier training workloads they were designed for. They are also a small fraction of a much larger story about who controls the valve — and Beijing, not Washington, is the entity currently deciding how much to turn it.
Frequently Asked QuestionsWhy did Beijing hold back H200 deliveries for eight months after Washington approved the sale?
Beijing’s hesitation reflected several overlapping pressures: distrust of the US requirement that chips physically route through American territory before reaching Chinese buyers, concern about potential surveillance or tampering; a domestic industrial policy goal of steering Chinese AI companies toward Huawei’s Ascend chips rather than Nvidia hardware; and the NDRC’s establishment of a case-by-case review process that required companies to justify foreign chip purchases against domestic alternatives. The practical effect was that US approval created a gate Washington controlled, and Beijing’s NDRC controlled a second gate on the other side.
What does the H200 do that Chinese domestic chips cannot?
The H200 is specifically valuable for frontier AI model training — the computationally intensive phase where a model’s parameters are adjusted across trillions of weight values. Its 141 gigabytes of HBM3e memory at 4.8 terabytes per second bandwidth allows large language models (100 billion parameters or more) to train efficiently without the excessive inter-chip communication overhead that smaller memory footprints impose. Huawei’s Ascend 950PR and similar domestic chips perform competitively for inference — serving existing trained models to live users. But China’s domestic chip ecosystem has not yet produced a chip that fully matches the H200’s training performance at scale, which is why ByteDance and Tencent specifically sought this hardware rather than expanding their existing Ascend clusters.
Why can’t Chinese companies just deploy their H200 chips in Hong Kong?
Beijing’s preferred solution — routing most of the licensed chips to Hong Kong rather than mainland China — runs into an infrastructure wall. Hong Kong lacks sufficient data-center capacity and power supply to absorb the volumes that approved buyers could theoretically receive under their US licenses. Building that capacity would require the same multi-year construction timelines as any large-scale data center project, and power grid constraints cannot be resolved quickly. At current volumes, the few thousand chips already in Hong Kong represent a workable arrangement. At 75,000 chips per buyer spread across 10 approved companies, Hong Kong’s infrastructure cannot scale to the task.
Will Nvidia’s more advanced Blackwell chips ever reach China?
Under current US export policy, Blackwell-generation GPUs remain prohibited for export to China. A May 2026 Commerce Department guidance also closed a loophole that had allowed Chinese-owned subsidiaries incorporated outside China to acquire Blackwell chips without triggering controls. Bipartisan legislation — the AI OVERWATCH Act — proposed codifying a two-year Blackwell ban in statute and requiring congressional approval for any future AI chip export licenses to China, though it remained pending as of August 2026. Nvidia is separately reported to be designing downgraded versions of its next-generation chips that could comply with export controls, a pattern it has followed since the original 2022 restrictions — though the policy environment for any new China-specific product remains volatile.