Beijing just handed ByteDance and Tencent 10,000 Nvidia H200 chips each, and the message is hard to miss: when China's own AI giants get a real choice, they still pick Nvidia over Huawei.
Nvidia's H200 chips have started arriving on the Chinese mainland in small shipments, with ByteDance and Tencent each receiving around 10,000 units, according to reporting from the Financial Times published August 18, 2026. Alibaba has also won approval to buy the chips, Reuters reported, citing people familiar with the decision. It's a notable reversal. For much of the past year, Beijing had informally steered its biggest tech firms away from foreign silicon, pushing them toward Huawei's Ascend line instead. Now the same government is letting its national champions bring in Nvidia hardware by the thousands.
The timing isn't a coincidence. The U.S. Commerce Department eased its own restrictions on H200 exports to China on January 14, 2026, moving from a presumption of denial to case-by-case review. Nvidia has to hand over 15% of the revenue from each Chinese H200 sale to the federal government, and it must certify the shipments won't pull supply away from U.S. customers. That's the American half of the thaw. China's approval for ByteDance, Tencent and Alibaba is the other half, and TrendForce has reported that total approvals could land below 200,000 chips, less than half of what Chinese buyers originally requested. Beijing is opening the door, not throwing it wide.
There's also a catch buried in the geography. Chinese regulators have told some companies they can ship H200 units to Hong Kong instead of the mainland, since Hong Kong sits outside mainland customs jurisdiction. That sounds like a workaround until you check the numbers: Hong Kong's data center capacity and power supply are both limited, so stacking large volumes of high-end AI chips there isn't simple. For now, the chips landing directly in Shenzhen and Beijing, in batches of 10,000, are the real story.
Huawei still can't close the gap #
Huawei has been racing to catch up. The company is aiming to roughly double output of its Ascend AI chips in 2026, with plans to push toward 1.6 million dies for the Ascend line. But a production target and a shipped, working chip are two different things. Huawei's Ascend 950PR is built on SMIC's N+2 process, roughly equivalent to a 7-nanometer node and two to three generations behind the 3-nanometer production Nvidia uses for its newest chips. Industry estimates put SMIC's yield on those large AI dies at around 50 to 60%, compared with the 80 to 90% yields TSMC typically hits. Huawei's own roadmap doesn't help the case either: its next-generation Ascend 950PR and 950DT chips are expected to underperform the current Ascend 910C, according to the Council on Foreign Relations.
Cadence's ChipStack AI Agent Can Now Design Chips Without Engineers Cadence Design Systems unveiled ChipStack, a Level 5 autonomous AI agent that runs chip design and verification loops with minimal human input, built with Nvidia and shown at Computex 2026. The company's Q2 FY26 revenue hit $1.584 billion, up 24.2%, with backlog at a record $8.1 billion and a raised full-year outlook of $6.26 to $6.34 billion. - AI chip design tool replaces traditional chip engineers - Cadence ChipStack autonomous agent Level 5 verification loop
Memory is a bottleneck too. Estimates on Huawei's HBM output for 2026 range from roughly 2 million stacks, enough for about 250,000 Ascend 910C chips, up to a more optimistic 7 million stacks that could support around 600,000 chips after packaging losses. Either way, that's a fraction of what Nvidia can ship in a single approved order to one customer.
What the choice actually says #
Here's the part worth sitting with. ByteDance and Tencent are not scrappy startups forced to take whatever chips they can get. They're the two companies with the deepest pockets and the strongest political incentive to prove China's chip independence works. When Beijing finally let them choose, they chose Nvidia, and they chose it at the exact volume regulators would allow. That's not a small signal about performance. It's confirmation, straight from the buyers with the most to gain from praising Huawei publicly, that domestic silicon still can't do the job at scale.
For Nvidia, the upside is real but bounded. Reuters and the Financial Times both frame these shipments as small relative to the appetite Chinese firms have shown, and the sub-200,000-chip cap TrendForce reported means this isn't a return to pre-restriction volumes. China's government still wants Huawei to win eventually, and it's still capable of tightening the tap again if Ascend yields improve or if Washington's mood shifts. What changed this week is narrower but still meaningful: for the first time in months, the two biggest forces in this trade war, Washington's export desk and Beijing's industrial planners, are loosening their grip at the same time. Also read: From Database to AI Data Platform: OceanBase's Roadmap to Rival Databricks • OpenAI Halted Training After Its Own AI Model Hacked Hugging Face • Cadence's ChipStack AI Agent Can Now Design Chips Without Engineers