Beijing can ban Nvidia purchases faster than it can replace CUDA

China's Cyberspace Administration issued guidance in mid-September 2025 barring major technology firms from buying Nvidia AI chips and pushing them toward domestic parts, with ByteDance and Alibaba told to suspend acquisitions including the RTX Pro 6000D ^[1][2]. The reason this matters to anyone planning compute capacity is that the procurement rule arrived before a substitute existed, and the missing piece is code rather than wafers.

Huawei's Ascend line is the most prominent domestic contender and has made real progress, according to a Crypto Briefing account of the situation ^[3]. But engineers and developers working with the chips consistently describe hardware that trails Nvidia on raw performance and, more critically, on software maturity ^[4]. That ordering is the whole story. CUDA has been built up over nearly two decades, with thousands of optimized libraries, tools and developer integrations layered on top of it ^[5]. A fab can eventually close a performance gap. Nobody closes a twenty-year library gap on a procurement timetable.

What developers are doing instead looks like triage. The reported workaround is optimizing inference workloads to run more efficiently on lower-powered or mid-range local hardware, with DeepSeek, Baidu and Alibaba among the firms adapting operations to what is domestically available ^[6][7]. That is real engineering, and DeepSeek's emergence as a globally competitive model trained on constrained hardware showed that a hardware disadvantage is not automatically fatal ^[14]. It is also not a substitute for a toolchain.

The demand side has been made captive by policy. State-funded data center projects have been retooled to require domestically produced chips, and the requirement was applied retroactively to builds already under way ^[11]. That guarantees Chinese suppliers a market even where their products are not yet competitive on performance ^[12]. Beijing is effectively running two tracks at once, a hard political directive to cut foreign dependence and a development program meant to make domestic chips good enough to carry it, and track one is moving faster than track two ^[13].

The clearest evidence that this is a demand-side decision rather than a supply constraint sits on Nvidia's side of the ledger. After Washington cleared the H200 for sale to China, no Chinese firm bought one, and Nvidia reported zero revenue from H200 sales to China as of mid-2026 ^[8][9]. Beijing's restrictions have proved comprehensive enough to block purchases even where the US approved them ^[17]. That is roughly nine months from directive to a confirmed zero ^[15], in a market that was historically one of Nvidia's largest for data center hardware before export controls tightened in 2022 and 2023 ^[10]. The lesson for Nvidia is narrow and expensive: designing a compliant part for China is not producing sales under the current political climate ^[18].

Three things to watch. Huawei's Ascend roadmap, which developers currently making do with available hardware are tracking for anything meaningfully better ^[16]. Whether the inference-optimization work being done at DeepSeek, Baidu and Alibaba hardens into a portable software layer or stays a per-company workaround ^[6][7]. And whether H200 revenue from China stays at zero, since that number, not any export licence, is the actual measure of who controls Chinese chip demand ^[9][17].