China's AI capital expenditure forecast to reach 600–700 billion yuan ($84–98B) in 2025, with government investment accounting for up to ¥400B. Core AI industry scale exceeds 1 trillion yuan. ByteDance alone plans ¥160B in AI capex. Alibaba unveiled a ¥380B three-year compute investment plan.
Shenzhen's Longgang district releases 'Ten AI Lobster Measures' — the world's first local government subsidy policy for OpenClaw (OPC) agent platform development. Up to RMB 2M for critical code contributions and embodied device integration. RMB 10M equity investment for seed-stage OPC projects. 30% hardware subsidies for AI NAS devices. Free deployment zones and open public data access in low-altitude airspace, transport, healthcare, and urban governance.
DeepSeek announces V4 model will run inference on Huawei 950 PR chips via CANN Next, marking the first frontier-class AI model purpose-built for Chinese semiconductor infrastructure. Withholds pre-release access from Nvidia and AMD, granting Huawei multi-week head start. Chinese hyperscalers (Alibaba, ByteDance, Tencent) place bulk orders for hundreds of thousands of Huawei units. Training still uses Nvidia GPUs, but inference sovereignty achieved.
Following DeepSeek V4's successful deployment on Huawei Ascend silicon (April 12), demand surged for the Ascend 950 chip. ByteDance, Alibaba, and other major Chinese tech firms are scrambling to secure orders. Huawei's Ascend supernode architecture has been confirmed to support DeepSeek V4 inference and training workloads.
China released draft rules governing interactive and human-like AI services (April 2026), alongside outbound personal data certification requirements and safety standards (effective Jan/Mar 2026). Combined with existing algorithm registration and content labeling requirements, China maintains the strictest generative AI governance globally. Focus areas: application deployment, security, labeling, and preventing outbound data leakage.
DeepSeek launched V4 preview: 1 trillion parameters MoE architecture, rewritten to run natively on Huawei Ascend and Cambricon chips. Alibaba Cloud and Tencent Cloud deployed same day. Huawei targeting 750K Ascend 950PR units in 2026, expects AI chip revenue to climb 60% to ~$12B. V4 demonstrates inference sovereignty — full stack without NVIDIA dependency.
US Commerce/BIS guidance clarifies that H100/H200/Blackwell licence requirements apply to China-headquartered companies operating outside China (Malaysia, Singapore, third-country jurisdictions). Closes 2025 grey-channel route (industry estimate: hundreds of thousands of restricted chips reached Chinese-linked entities via Malaysia during AI Diffusion Rule unenforcement gap).
DeepSeek (Hangzhou) closed first external funding round at ~50B yuan / $7.4B, valuation above $50B (China's most valuable AI startup). Novel capital structure: only China's National Artificial Intelligence Industry Investment Fund (NAIIF) received equity, voting rights, and freedom from lock-up. Every other investor — including Tencent and battery maker CATL — accepted 5-year lock-ups and surrendered all voting rights. Founder Liang Wenfeng contributed ~$3B (largest single stake). Capital channeled into limited partnership managed by Liang rather than into DeepSeek directly.
Within 24-72 hours of the Anthropic Fable 5 / Mythos 5 global disable (June 12), routing-aggregator OpenRouter's top-4 most-used model positions were occupied by DeepSeek, MiniMax, Tencent, and Xiaomi Chinese open-weight models. DeepSeek V4 Pro lists at $3.48 per million output tokens vs Fable 5's $50 — 14x pricing gap. Budget-constrained organisations (academic labs, regional governments, SMEs, developing-market deployments) switched within days. First measurable market-substitution event directly attributable to a US frontier-AI export-control firing.