Triage of scarce compute: Beijing is instructing western provinces (Inner Mongolia, Gansu, Qingyang) to handle training workloads, while new clusters in population hubs like Wuhu (Yangtze Delta), Guizhou (Guangzhou), and Qingyang (Chengdu/Chongqing) focus on inference, where proximity to users matters.

Massive spend: Wuhu’s “Data Island” alone = RMB 270B (~$37B), hosting Huawei, China Telecom, China Unicom, and China Mobile. Subsidies cover up to 30% of AI chip procurement, more generous than other regions.

Workaround strategy: U.S. export controls prevent Chinese access to Nvidia’s most advanced GPUs (Blackwell series). China’s stopgaps include (1) domestic chips from Huawei & Cambricon, (2) black-market imports of restricted Nvidia GPUs, and (3) networking smaller clusters together to mimic hyperscale capacity.

Technology pivot: Huawei is spearheading “UB-Mesh,” a networking solution that links disparate sites and claims to double LLM training efficiency by smarter load allocation across clusters.

Strengths of U.S.: Hyperscalers (Meta, Google, Microsoft, xAI) are deploying tens of thousands of cutting-edge Nvidia GPUs, with projects like Stargate (Texas) targeting 400,000 processors in a single mega-complex. U.S. advantage = scale + supply chain + Nvidia lock.

China’s challenge: Lacking top-end chips, Beijing is forced into second-best solutions — distributed networks of less-powerful processors. This introduces efficiency losses, higher costs, and maintenance headaches.

Beijing’s hedge: Subsidies + central planning can stretch scarce resources further. Long-term, if Huawei/Cambricon can achieve commercially viable AI chips, China could carve out a sustainable domestic ecosystem — though still years behind.

Huawei – positioned uniquely as both a telecom & AI hardware player. UB-Mesh could become the backbone of China’s distributed AI infra. Rating: 8/10.

China Telecom / Unicom / Mobile – beneficiaries of the state-mandated push to interconnect clusters nationwide. Rating: 7.5/10.

Cambricon (688256 CH) – Beijing-backed AI chip designer. Investors already betting it will be “China’s Nvidia.” Rating: 7/10.

Nvidia ($NVDA) – short-term demand hit (can’t sell advanced GPUs to China) but black-market demand underscores irreplaceability. Rating: 9/10.

TSMC ($TSM) / Samsung ($005930 KS) – restricted from supplying advanced nodes to China, which keeps their highest-margin production tied to U.S./allied customers. Rating: 8.5/10.

Western hyperscalers ($MSFT, $GOOG, $META, $ORCL) – continue to enjoy uncontested access to top compute. Reinforces U.S. lead. Rating: 9/10.

China AI startups – forced to compete with second-best hardware, black-market supply, and fragmented infra. Efficiency disadvantage = margin compression. Rating: 5/10.

European utilities/infra providers – excluded from China’s domestically protected market, while U.S. players capture the bulk of hyperscale spend. Rating: 5/10.

Gas/coal-heavy power providers in China – will be leaned on to power these clusters but without subsidy upside. ESG headwinds mount. Rating: 4/10.

U.S. structural lead widens: Export controls + Nvidia’s dominance + hyperscaler capex ensure the U.S. stays ahead for at least the rest of this decade.

China’s mitigation strategy: Central planning + Huawei UB-Mesh + subsidies keep them in the game, but at higher cost and lower efficiency.

Tradeable edges: