Nvidia's H200 finally lands in China at about 1% of the order book

Nvidia's H200 has reached Chinese buyers: ByteDance and Tencent have each received about 10,000 of the accelerators in recent weeks, according to the Financial Times, with other Chinese groups expected to secure similar volumes shortly ^[1][2]. Measured against the licences Washington granted in May and the orders Chinese firms placed in January, the shipments are small enough to say something useful about where the bottleneck in Chinese AI capacity actually sits ^[3][4]. The arithmetic is unflattering. Two companies at 10,000 units each is roughly 20,000 chips ^[1]. Chinese buyers had reportedly ordered around two million H200s in January, so the delivered volume is on the order of 1% of that book ^[4][2]. When the Commerce Department cleared about ten Chinese firms to buy the part in May, each licence carried a 75,000-unit cap and not a single chip had shipped ^[3]. Twenty thousand units is therefore about 13% of what those two firms alone were permitted to take, and under 3% of the ceiling across all ten licences ^[3][4]. Before this, eight months of export licences moved almost nothing ^[6]. The cap itself is not settled. The FT puts the US ceiling at 100,000 chips per company, and Engadget's account of the same report repeats that figure ^[5][c5b]. The number reported by Bloomberg in March and echoed in coverage of the May approvals was 75,000, and no source has documented a revision ^[5]. Until Commerce says otherwise, the higher figure belongs to one paper. What has been binding is not manufacturing capacity but administrative permission, and per the reporting the tighter grip is Beijing's. China blocked H200s at customs in January, then conditionally cleared ByteDance, Alibaba and Tencent for more than 400,000 units combined a fortnight later, before Washington narrowed the field again in May ^[7]. Commerce Secretary Howard Lutnick said in May that "the Chinese central government has not let them ... buy the chips, because they're trying to keep their investment focused on their own domestic industry" ^[14], a position consistent with the $295bn domestic AI data centre plan Beijing drafted in June ^[15]. The sources also diverge on where the silicon is allowed to sit. The FT, as summarised by TNW, reports that Beijing has told companies to keep the hardware outside the mainland to protect domestic chipmakers and is letting firms route processors to Hong Kong, which lies outside the mainland customs border ^[8]. Engadget's reading of the same report says the chips were allowed into the mainland, with authorities wanting most of the order kept out and shipped to Hong Kong instead, where the firms do not currently operate data centres ^[18]. Hong Kong already handles more than half of China's chip imports, some $124bn between January and May ^[9], and mainland engineering teams can reach Hong Kong compute over cross-border links, which keeps the hardware offshore while making it usable ^[10]. Commerce has tried to pre-empt that once: a 31 May clarification extended the licensing requirement to Chinese and Macau-headquartered entities wherever they operate ^[12]. The Bureau of Industry and Security is now reviewing how Chinese firms rent access to hardware they may not own outright ^[11], and Nvidia cut more than half its Asian customers from an internal approved-buyer list in July ^[13]. Watch three things: whether the next tranche of approved firms lands in the same 10,000-unit increments ^[2], whether BIS turns its rental review into a rule ^[11], and whether Commerce confirms 75,000 or 100,000 as the per-company cap ^[5]. Also worth pricing: the 25% revenue share to the US government that Trump attached to H200 sales in December, up from the 15% Nvidia and AMD accepted on H20 and MI308 the previous August ^[16]. Jeffrey Kessler told the House Foreign Affairs Committee on 14 July that "very few" H200s had reached China or Hong Kong, a volume he called trivial ^[17].