After eight months of waiting, Nvidia's H200 has finally reached the Chinese mainland. ByteDance and Tencent each took delivery of roughly 10,000 accelerators in the days before August 18, the first meaningful movement of the chips into the country since Washington cleared their export in January. It reads like a re-opening, and that is exactly the trap. Measured against what was licensed, the deliveries amount to barely 13 percent of the two firms' quotas, most of the hardware is parked in Hong Kong rather than serving anything, and Beijing is still the real gatekeeper. Head to head with Huawei's surging Ascend line, this is not much of a contest: Nvidia won the right to sell, but it cannot turn that right into chips that actually run.
The headline is true, barely
Ten thousand units sounds decisive, and it is how the story is being sold. But it is 13 percent of the 75,000-unit cap each firm was granted under the January rule. Roughly ten Chinese companies cleared that first licensing round, Alibaba and Tencent and ByteDance and JD among them, and only two have shown up in the delivery reports. Even against the 400,000-plus units the big buyers were collectively approved for, the hardware actually on the mainland amounts to about 2.5 percent of the order book.
Location matters more than the number. Most of each firm's licensed allowance must stay outside the mainland, largely in Hong Kong, which sits beyond China's customs border. That lets Chinese companies reach the compute over cross-border network links without technically importing it, and the Financial Times says Beijing wants it held there to protect its domestic chipmakers. On top of that sits a second, discretionary checkpoint: the National Development and Reform Commission approves each server purchase individually. Washington opened its door in January; eight months later, Beijing's own machinery decides how much actually clears.
Why Beijing is its own best bottleneck
The subtle part is that this may not be a stubborn rule at all. It might be a rationing strategy. Keep the fast Nvidia chip technically legal but functionally scarce until Huawei's Ascend has closed the gap.
Run the power math and the strategy looks almost effortless. An H200 draws up to 700 watts; eight in an HGX node with hosts and networking land around 10 kilowatts per server. One firm deploying its full 75,000-unit allocation would need about 9,375 servers, or roughly 94 megawatts of IT load before cooling overhead.
Hong Kong's entire installed data center base is only about 581 megawatts. A single company maxing out its quota would need more than a sixth of everything the territory has ever built, competing for colocation and grid power against every other tenant. The relief valve is years out: the Northern Metropolis Sandy Ridge cluster is not targeting operations until 2029. Two gates, one of them a power grid that cannot expand on any useful timeline, have become a more effective ceiling than the license cap ever was.
The scorecard: H200 versus Ascend
Set the two contenders side by side and the verdict stops being subtle.
| Metric | Nvidia H200 | Huawei Ascend | Edge |
|---|---|---|---|
| Production | China-configured output halted March 2026; capacity shifted to Vera Rubin | 950PR in mass production since Q1; training-focused 950DT due Q4 2026 | Ascend |
| Access | NDRC case-by-case approval; most units parked in Hong Kong | Domestic silicon, no export barrier | Ascend |
| 2026 China share | Roughly 40% in 2025, tracking toward about 8% | Climbing toward roughly 50% (Bernstein) | Ascend |
| Frontier software | Depends on a chip that cannot run at scale | DeepSeek V4, a 1.6 trillion-parameter model, shipped day-one on Ascend | Ascend |
| Revenue outlook | Guidance still excludes China data center compute | Roughly $12 billion in AI chip revenue expected this year | Ascend |
The verdict is lopsided, and not just because of one shipment. Bernstein projects Nvidia's share of China's AI chip market falling from about 40 percent in 2025 to roughly 8 percent by the end of 2026, with Huawei taking around half. TrendForce goes further, putting domestic silicon on track for nearly 90 percent of the country's AI hardware market by year end. The story is not a re-opening; it is a passing of the baton that had already been handed over.
None of this is locked in, and a comparative read should never miss the open questions:
- Whether shipped volume scales past 13 percent of quota, which would signal Beijing loosening its own gate rather than just Washington's.
- Whether Hong Kong's grid builds out in time, or the licensed capacity stays paper for another three years.
- Whether Huawei's 950DT training chip ships on schedule in Q4, closing the one gap where H200 might still compete on merit.
For now, the practical answer for anyone buying compute is the same one the market has already reached: where power exists and the queue clears, the imported H200 is still a fine chip. But in China, the machine that decides whether that chip ever runs belongs to the seller's rival. Nvidia has the hardware, the account, and the permission. It just cannot get the lights on.
Comments