Version 4

Chinese AI models held 61% of weekly token consumption on OpenRouter by mid-2026 [1], with China occupying 6 of the top 9 positions in a global model usage ranking from spring 2026 [3]. MiniMax M3 — a 428B-parameter mixture-of-experts model — released June 13 with day-0 open weights on HuggingFace, CUDA/ROCm support, and 9x faster prefill than its predecessor, illustrating the pace of Chinese model releases fueling that shift [7][8]. DeepSeek disclosed hiring for MW-to-GW scale data center infrastructure, interpreted as a move toward vertical integration of physical compute [9]. SemiAnalysis's teardown of SMIC N+3 chips finds China reached TSMC N6-class logic density via DUV multi-patterning, but at higher cost and complexity than EUV-based production [10].

Chinese AI providers are now competing across the full stack — model capability, open-weight availability, and physical infrastructure — not just on API pricing. If DeepSeek and peers own compute at gigawatt scale, the cost-performance dynamics currently favoring Chinese models on platforms like OpenRouter may persist even if U.S. labs narrow the pricing gap. The semiconductor picture complicates the story: export controls have not stopped China's chip progress, but have forced a more costly and complex manufacturing path that may constrain the pace of future gains.

Through most of 2025, U.S. models led weekly token consumption on OpenRouter, the API routing platform widely used by developers and AI startups. From early 2026, Chinese models became the primary growth driver, reaching 61% of the platform's weekly token market share by mid-2026 [1]. Social media commentary describes the trajectory as going from roughly 1% to 60% of developer API traffic within twelve months [2], though that figure circulates as secondary amplification rather than primary analysis. A separate global AI model usage ranking from spring 2026 shows China holding 6 of the top 9 positions [3]. Observers differ on causes: Rohan Paul treats the OpenRouter gains as a structural competitive shift [4]; Arnaud Bertrand attributes the turn to a specific Chinese model release in 2026 [5]; developer-community commentators argue that developers route to whatever model delivers the best cost-performance ratio at a given moment, making the gains real but conditional on price [6].

MiniMax M3, released June 13, 2026, is a concrete example of the Chinese model releases driving this dynamic. The model has approximately 428B total parameters with 23B activated — consistent with a mixture-of-experts architecture — and was available on HuggingFace on day 0 with open MXFP8 weights [7][8]. It supports both CUDA and ROCm inference stacks at launch, and its block sparse attention achieves 9x faster prefill than its predecessor M2.7 [8]. Day-0 open weights with cross-platform hardware support reduce the friction between model release and developer adoption in routing decisions.

At the infrastructure layer, DeepSeek disclosed job openings for roles explicitly scoped to the design and delivery of MW-to-GW scale data center infrastructure, building on data center operations hires in Ulanqab, Inner Mongolia from April 2026 [9]. SemiAnalysis interprets the combined hiring pattern as a deliberate shift from a software-first posture to vertical integration of physical compute at gigawatt scale. The semiconductor context is more mixed: SemiAnalysis's teardown of SMIC N+3 finds that China has reached TSMC N6-class logic density through aggressive DUV multi-patterning and design-technology co-optimization, but pays for that in cost, complexity, and process risk [10]. The Kirin 9030 Pro, fabricated on N+3, performs comparably to three-year-old Android flagships and trails current top-tier SoCs from Apple, Qualcomm, MediaTek, and Samsung [10]. Export controls have not stopped SMIC and Huawei from shipping advanced silicon, but have forced a different and more complex manufacturing path in the absence of EUV lithography [10].

On the product side, Tencent's WorkBuddy became China's #1 PC productivity AI agent by daily active users in early June 2026, with integrations across GitHub, Jira, Notion, Gmail, Google Drive, and Slack [11][12]. Tencent subsequently released an enterprise edition targeting team-level workflows and global markets [13]. Analyst Jeffrey Towson frames this as a deliberate ecosystem push toward a full enterprise suite [14]. Together, the model-layer gains on OpenRouter, new open-weight releases like MiniMax M3, infrastructure investment by DeepSeek, and WorkBuddy's product expansion describe Chinese AI providers competing across multiple layers simultaneously.