China Quietly Lets Alibaba, ByteDance, and DeepSeek Buy Nvidia H200 Chips Again
Beijing is reversing its own domestic-purchase restrictions on Nvidia H200 GPUs, reopening a channel it closed in response to US export controls. Nvidia shares climbed on the news.
Chinese regulators have told Alibaba, ByteDance, and DeepSeek they can now buy Nvidia H200 AI chips — reversing a domestic-purchase restriction Beijing itself imposed, not one Washington lifted.
What changed
Bloomberg reported July 8, citing sources familiar with the matter, that Chinese officials are allowing a capped number of H200 purchases: under 200,000 units total, less than half of what the three companies had requested. Each firm must justify the quantity and specific use case to get approval — this isn’t an open floodgate, it’s a rationed exception.
The irony is pointed. US export controls restrict which Nvidia chips can legally reach China. But Beijing had layered its own restriction on top, pushing domestic firms toward homegrown silicon from Huawei and others as part of a self-sufficiency push. That policy is now bending under pressure from AI labs and cloud providers who say domestic chips can’t yet match H200 throughput for frontier-scale training runs.
Market reaction
Nvidia shares moved on the report, with gains in the 1.5%–3.7% range depending on which trading session and outlet you check — the spread reflects normal noise in same-day reporting, not conflicting facts. Investors read it as a reopening of a revenue channel that had been closed by Chinese policy rather than by US sanctions, which is a meaningfully different risk profile for Nvidia’s China exposure going forward.
Why it matters
DeepSeek’s inclusion on the approved list is notable. The company has spent the past two years demonstrating that constrained-compute training can still produce competitive models — LongCat and DeepSeek’s own releases have leaned on efficiency gains to compensate for chip access. A green light to buy H200s at scale suggests even efficiency-focused Chinese labs see hardware access as the harder constraint to engineer around, not the software layer.
For Alibaba and ByteDance, the calculus is more straightforward: cloud AI workloads and internal model training both benefit directly from H200 throughput, and neither company wants to bet a frontier roadmap entirely on domestic chips that are still catching up on memory bandwidth and interconnect.
The cap matters as much as the approval. Under 200,000 units split three ways is a fraction of what any one of these companies would need for a genuinely frontier-scale cluster. This reads less like liberalization and more like Beijing managing a controlled release valve while its domestic chip industry — Huawei’s Ascend line and newer entrants like Biren — continues to close the gap.