Version 9

Chinese AI providers are advancing on model capability, open-weight distribution, capital, and physical infrastructure while facing a structural compute spending gap versus the US. GLM-5.2 — a 753B-parameter MIT-licensed model released June 16 by Z.ai — tops the Artificial Analysis Intelligence Index v4.1 [8], costs approximately $4.40 per million output tokens [9], and has drawn sharply divided assessments: Nathan Lambert calls it the first open-weight model to work credibly as a general agent in coding harnesses [10], while Zvi Mowshowitz argues it is heavily distilled from Claude Opus, trails frontier closed models substantially, and sits in an awkward commercial position [12]. DeepSeek's $7.4B raise at a $50B valuation [14] and disclosed MW-to-GW data center plans [15] represent the infrastructure dimension. US hyperscalers are projected to spend approximately 8.3x more than Chinese hyperscalers on AI infrastructure by 2027 [18]. The US-ASML dispute over whether a banned EUV tool reached China remains unresolved: ASML says it actively tracks all ~314–340 EUV units worldwide and none have ever shipped to China [20][21].

The GLM-5.2 debate is a proxy for a larger question: whether Chinese open-weight labs have closed enough of the capability gap to put meaningful commercial pressure on closed-model incumbents. The 8.3x projected compute spending gap suggests US labs retain a structural advantage in scaling, but if distilled open-weight models can deliver near-frontier performance at a fraction of the price, compute scale alone may not maintain commercial separation.

Through 2025, US models led weekly token consumption on OpenRouter. From early 2026, Chinese models became the primary growth driver, reaching 61% of the platform's weekly token market share by mid-2026 [1], with China holding 6 of the top 9 spots in a spring 2026 global usage ranking [2]. OpenAI's June 2024 cutoff of API access for Chinese developers created an opening that Baidu, Alibaba, and Moonshot AI actively exploited [3][4]. Observers disagree on causes: Rohan Paul treats the shift as a structural competitive gain [5]; developer-community commentators argue developers route to whatever delivers the best cost-performance ratio at a given moment, making the gains real but price-conditional [6].

The open-weight model pipeline in mid-June 2026 produced two significant releases. MiniMax M3 (428B total parameters, 23B activated, block sparse attention 9x faster than its predecessor) shipped June 13 with day-0 MXFP8 weights [7]. GLM-5.2, released June 16 by Z.ai under an MIT license, is a 753B-parameter mixture-of-experts model that tops the Artificial Analysis Intelligence Index v4.1 with a score of 51, expands context to 1 million tokens, and costs approximately $4.40 per million output tokens [8][9]. Its reception divides sharply. Nathan Lambert describes it as the first open-weight model to work credibly as a general agent in coding harnesses like Claude Code, placing it alongside OpenAI and Anthropic's latest closed models on Arena's agent leaderboard, and argues GLM-5.2 puts economic pressure on Anthropic's Claude Code revenue by offering an open alternative [10]. The Vercel CEO described its coding ability as 'almost shocking' [11]. Zvi Mowshowitz credits the achievement — calling GLM-5.2 the strongest currently available open-weight model, scoring around Opus 4.7 on traditional benchmarks — but argues it is very likely heavily distilled from Claude Opus, citing its behavioral voice and reliance on a Claude harness; he notes distilled models tend to overperform on benchmarks and underperform on less common tasks, and that GLM-5.2 sits in an awkward commercial position — not cheap enough for bulk tasks nor strong enough for the hardest tasks compared to closed-model alternatives [12]. Milk Road AI argues open-source models will not win the AI race and the market narrative around DeepSeek's early 2025 launch misread the episode as validation of open-weight approaches [13].

At the infrastructure and financial layer, DeepSeek completed a $7.4B financing round at a $50B valuation on June 16, becoming China's most valuable AI startup; founder Liang Wenfeng held approximately 90% pre-round and personally contributed around $3B [14]. Job postings for MW-to-GW scale data center design signal a move from software-first to vertical integration of physical compute [15]. China established a government-chartered Space Computing Industry Innovation Center led by Beijing University of Posts and Telecommunications before Elon Musk announced the AI1 satellite [16]. A hardware teardown of SMIC's N+3 process found it reaches TSMC N6-class logic density via DUV multi-patterning but at higher cost, with the resulting Kirin 9030 Pro trailing current flagship SoCs [17]. Against this, Rohan Paul notes that US hyperscalers are projected to spend approximately 8.3x more than Chinese hyperscalers on AI infrastructure by 2027, arguing compute advantage remains firmly on the US side [18].

US Commerce Secretary Howard Lutnick personally raised with ASML leadership the concern that a banned EUV lithography machine may have reached China [19]. ASML denied the claim publicly, saying it has never shipped an EUV machine to China, that it actively tracks all ~314–340 EUV units worldwide, and that the reports are 'inaccurate and damaging to our reputation' [20][21][22]. The dispute is unresolved. US export controls apply to EUV machines in part because they incorporate US intellectual property [23]; ASML separately sells older DUV machines to China, a distinction central to the current disagreement.