Nvidia H200 Approved for China, Beijing Won't Accept Delivery

Washington gave Nvidia exactly what it asked for. In December 2025, the Trump administration reversed years of export controls and cleared H200 chip sales to around ten Chinese firms, including Alibaba, Tencent, ByteDance, and JD.com, each authorised to buy up to 75,000 units. Lenovo and Foxconn were approved as distributors. Five months later, not a single chip has shipped. The blockage is not in Washington. It is in Beijing. US rules require the chips to be used inside China. Beijing has separately told its tech companies to focus procurement on domestic suppliers and limit Nvidia use to overseas operations. Those two requirements cannot both be satisfied at the same time, so nothing moves. There is an additional layer. Trump negotiated a structure under which 25 percent of H200 sale revenue flows to the US government, requiring the chips to physically pass through US territory before reaching China. Beijing has responded with concern about potential tampering. The arrangement was designed as a legal workaround for export fee restrictions, but it gave Beijing a credible reason to treat the chips as a security risk. Jensen Huang flew to Beijing with Trump, added to the delegation at the last minute after Trump saw press coverage that he had not been invited. He left without a deal. US Trade Representative Jamieson Greer confirmed that semiconductor controls were not substantively on the bilateral agenda. What the summit produced was optics. What it confirmed is that the stalemate is structural, not diplomatic. The more significant development happened in the weeks around the summit. DeepSeek released V4, its most powerful model yet, with 1.6 trillion parameters and a 1 million token context window. It is the first major Chinese frontier model optimised for Huawei's Ascend chips in training, not just when running the finished model. DeepSeek gave early hardware access to Chinese chipmakers only, skipping Nvidia and AMD entirely. The market responded immediately. Alibaba, ByteDance, and Tencent placed large orders for Huawei's Ascend 950PR chip, which entered mass production in March. Huawei now projects AI chip revenue of roughly $12 billion in 2026, up 60 percent from $7.5 billion in 2025, based on orders already secured. Huawei is on track to hold around 60 percent of China's AI chip market by year-end. Nvidia's position in that same market is now zero, in Huang's own characterisation. Before export controls tightened, Nvidia held roughly 95 percent of China's advanced chip market and China contributed around 13 percent of total company revenue. Nvidia's current guidance assumes zero recovery from China this year. Huawei's chips are not yet equal to Nvidia's best. The Ascend 910C lags behind the H200 in raw performance, and Huawei's current timeline does not put a chip matching the H200 into production until late 2027. But DeepSeek V4's results suggest the gap is manageable for inference workloads, which is where most real-world AI applications actually run. For training large models, the gap is still meaningful. Beijing is making a calculated bet. Lock the domestic stack in now, accept a performance disadvantage for a few years, and come out with a self-sufficient supply chain that the US cannot disrupt. DeepSeek V4 is the first hard evidence that frontier AI models can be trained on Chinese hardware. It is not a proof of parity. It is a proof of viability. For any business that buys AI services from Chinese platforms, or competes with Chinese companies that do, this matters. The AI tools being built in China are increasingly running on a different hardware stack, optimised differently, and insulated from US supply chain decisions. That split will widen before it narrows. The H200 deal being frozen is not a temporary standoff waiting for diplomacy to fix it. It is Beijing choosing which direction its AI industry faces.