国产 AI 算力与芯片自主 2026

Executive Summary · English

(Data as of 2026-08-01) This report tracks China's AI compute and chip sovereignty. Domestic AI chips took 52.3% of the China market in Q1 2026, crossing 50% for the first time. On July 9 the Greater Bay Area's first all-domestic Ascend 10,000-card cluster went live in Shaoguan; Huawei debuted the Ascend 950 SuperPoD at WAIC 2026 (July 17-20). On export controls, a U.S. Commerce official testified on July 14 that actual H200 deliveries to China remain "trivial" after Beijing suspended purchases in January. Kimi K3 (2.8T parameters, the largest open-source model ever) topped the frontend Code Arena and fully released its weights on July 27; Moonshot AI closed a $3.5B+ Series F on July 29 at a $35B post-money valuation; DeepSeek V4 GA launched July 20 with day-0 domestic-chip adaptation. Musk's reported $1B+ acquisition of APR Energy confirms "power is the new compute." Anthropic is proceeding toward an October Nasdaq IPO while OpenAI reportedly weighs delaying to 2027. Based on public information only; not investment advice.

• 01Domestic AI chips took 52.3% of China's market in Q1 2026 (first time above 50%), with Ascend at ~37%. On July 9 the Greater Bay Area's first all-domestic Ascend 10,000-card cluster went live in Shaoguan (30 supernodes / 11,520 cards / 9,000P / all Ascend 910C / ~RMB 5.5B investment); at WAIC 2026 (July 17-20) Huawei debuted the Ascend 950 SuperPoD (UnifiedBus interconnect, up to 1,024 cards); 950PR entered mass production in April and 950DT is reportedly pulled forward to August.
• 02H200 standoff: the U.S. approved exports on Jan 13 with conditions (25% revenue share, full non-refundable prepayment); China's MOFCOM suspended purchases within 24 hours. On July 14 Commerce official Kessler testified that actual deliveries to China remain "trivial" despite ~$10B in approved licenses; Chinese firms reportedly ordered 2M+ chips for 2026 vs. ~700K Nvidia inventory, and Nvidia guides zero China H200 revenue.
• 03Moonshot AI released Kimi K3 on July 16: 2.8T parameters, the largest open-source model ever (1M context, KDA + LatentMoE activating 16 of 896 experts). It topped Arena.ai's Frontend Code Arena at 1,679 points, surpassing Claude Fable 5 (1,631); full weights hit Hugging Face on July 27. On July 29 Moonshot closed a $3.5B+ Series F at $35B post-money; a pre-IPO round reportedly starts in August targeting $50B pre-money, with a Hong Kong listing possible within six months.
• 04DeepSeek V4 GA launched July 20: a dual-model matrix (V4-Pro 1.6T params / 49B active; V4-Flash 284B / 13B active), 1M context, MIT license, and novel peak-valley pricing, with Cambricon Siyuan 690 adapted day-0. Zhipu launched its "Touch High" AGI program on July 11 (cumulative H-share fundraising ~RMB 31.5B) and relaunched the GLM Coding Plan on July 31 with transparent credits from RMB 118/month.
• 05"Power is the new compute" confirmed globally: APR Energy (1GW+ mobile gas-turbine fleet) was acquired for a reported $1B+, surfacing via an FTC early-termination notice (No. 20261350, dated 2026-05-14); multiple outlets report Musk bought it personally, with capacity mainly powering xAI's Grok data centers. Speed of power procurement has replaced chips as the top constraint on AI expansion.
• 06Capital markets diverge: Anthropic is proceeding toward an October Nasdaq IPO at a ~$965B Series H-1 valuation, with investor meetings from July 15 and bankers viewing $1T+ as the base case; OpenAI reportedly leans toward delaying its IPO to 2027. Cambricon's market cap first crossed RMB 1 trillion on July 1; DeepSeek's first round settled at ~RMB 351B valuation, with a second round reportedly discussed at $70-74B.