Chinese AI: open weights, export controls, and the domestic stack

For much of 2023–2024, Western coverage treated Chinese AI as a follower story: strong engineering, thinner frontier models, and a hard ceiling from US export controls on advanced GPUs. That frame cracked when DeepSeek's open-weight releases matched — and sometimes undercut on cost — systems the market had priced as US-only. The buzz since then is less about one chatbot and more about a full stack: models, apps, cloud, and chips that no longer assume Nvidia as the default.

Why DeepSeek mattered

DeepSeek did not invent Chinese large models. Alibaba's Qwen line, ByteDance's Doubao, Baidu's Ernie, Moonshot, Zhipu, MiniMax, and others were already shipping. What DeepSeek changed was the narrative abroad: competitive reasoning and coding models, aggressive open-weight releases, and training/inference economics that made US CapEx stories look less inevitable. Once that proof existed, every Chinese lab's roadmap — and every Western CIO's "do we need a China strategy?" slide — got rewritten.

Open weights as strategy

US frontier labs mostly keep top models closed. China's public AI culture leans harder on open or open-weight releases: weights on Hugging Face and ModelScope, Apache-friendly licenses, and rapid forks. That is not charity. It seeds domestic developers, fills enterprise fine-tunes, and exports influence where ChatGPT is blocked or expensive. It also creates a measurement problem for the West: capability is no longer gated only by API access; it ships as downloadable artifacts.

Apps and distribution

Models alone do not win China. Distribution does. ByteDance can put AI into Douyin and enterprise suites; Alibaba into cloud and commerce; Tencent into WeChat surfaces; Baidu into search. Consumer assistants compete inside platforms Western products cannot easily enter. The result is a parallel product universe: familiar patterns (chat, agents, coding copilots, video) with different defaults for language, payment, and compliance.

Export controls and the Ascend bet

US rules have narrowed what Nvidia can sell into China. Beijing has pushed champions toward domestic silicon. Huawei's Ascend line became the practical alternative for large clusters. In 2026, DeepSeek's V4 line was adapted for Ascend rather than treated as an Nvidia-only artifact — a signal that "frontier on domestic chips" is now a product claim, not only a policy slogan. Cloud giants followed with large Ascend procurement. Supply is still tight: the same controls that block Nvidia also constrain China's access to the tools and memory needed to manufacture advanced AI chips at scale.

Everyone wants their own silicon

Huawei is not alone. Alibaba and Baidu have been building AI chips; DeepSeek has been reported to explore its own inference silicon to reduce reliance on both Nvidia and Huawei. That mirrors Western moves (custom inference ASICs at hyperscalers) with a sharper geopolitical edge: control the stack or accept a permanent bottleneck.

What the buzz gets right — and wrong

Right: China is not a bit player. Open-weight models travel globally; domestic apps lock in local users; hardware substitution is real enough to matter for planning. Wrong: that US labs have "lost" or that Ascend already equals the best Nvidia training clusters. Capability gaps, manufacturing limits, and content/regulatory constraints still shape what ships and where. The accurate read is bipolar competition with messy interdependence — talent, papers, and open code still cross borders even when GPUs do not.

The point

China's AI push is about building a full stack it controls, from models and apps to cloud and chips, despite US export limits. In practice, follow Chinese open-weight models as closely as US APIs, expect enterprise buyers in Asia to compare Qwen, Doubao, and DeepSeek with GPT and Claude, and treat chip export rules as a real business risk. Success now depends on who can train, serve, and sell models under the rules that actually apply to them.