Goldman Sachs: Chinese AI Labs Are About to Start Charging for Their Open-Weight Models

What the 'Paid Weights' Model Looks Like

Goldman's thesis describes a shift from today's dominant approach to a new revenue-generating one :

• Current model: Open-weight models released under permissive licenses like MIT or Modified MIT. Anyone — including large cloud providers like AWS, Google Cloud, and Alibaba Cloud — can download, host, and serve the weights at no cost to themselves .
• Projected model: A "paid weights" approach where cloud platforms and high-revenue "Model as a Service" (MaaS) businesses must pay commercial licensing fees to the model developer. The weights remain downloadable, but large-scale commercial use triggers a separate paid agreement .

Goldman Sachs analyst Ronald Keung framed the shift as turning surging global adoption into direct revenue for the first time . The bank forecasts China's AI model API and subscription revenue will grow from approximately 35 billion RMB in 2026 to 879 billion RMB by 2030, driven by a 25-fold surge in daily token consumption — and paid licensing would let developers capture a share of that growth .

Why Now: Performance Parity and Zero Revenue

Two key factors explain why this shift is happening now :

1. Frontier-level performance. Chinese open-weight models like Moonshot AI's Kimi K3 and Zhipu AI's GLM-5.2 now rival top US proprietary models in benchmark performance . Goldman's report characterized the trajectory as "from last year's DeepSeek cost-efficiency moment to this year's Zhipu GLM intelligence moment" . The earlier performance gap that made free distribution a necessary adoption strategy has effectively closed .

2. Massive global adoption with zero developer revenue. Because models ship under open-source terms, cloud platforms can host and resell them without paying the original developer. Rising usage has not translated into model-builder revenue . As Goldman put it, the open-weight format means "anyone can download and host the weights for free. The labs earn almost nothing from that use" .

Kimi K3: The Real-World Proof

Moonshot AI's Kimi K3, released on July 27, 2026, is the clearest real-world example of the licensing shift Goldman describes . Kimi K3 is a 2.8-trillion-parameter open-weight model — the largest ever released . It ships under a custom Kimi K3 License, not MIT or Apache 2.0, and carries explicit commercial triggers :

• For most users: Free to use, modify, fine-tune, and redistribute the model .
• For large MaaS businesses: If the licensee or any affiliate operates a "Model as a Service" business with aggregate revenue exceeding $20 million USD over any consecutive 12 months, they must enter into a separate commercial agreement with Moonshot before any commercial use .

The New York Times noted that Moonshot "is also looking to profit from its model's appeal by requiring large-scale users to enter into commercial licensing agreements" . One analysis put it bluntly: the license is "open source, but ultimately paid" .

GLM-5.2: The Counter-Example

In contrast, Zhipu AI's GLM-5.2 — released June 13, 2026 as a 744-billion-parameter Mixture-of-Experts model — uses a standard MIT license with no usage restrictions . The weights are fully open: free to download, self-host, fine-tune, and use commercially without any revenue thresholds . Zhipu monetizes through its hosted API and GLM Coding Plan, not through weight licensing .

Goldman's thesis is that other developers may follow Moonshot's lead rather than Zhipu's, moving toward tiered commercial terms as their models reach parity with closed-source rivals .

What This Means for Developers and Businesses

The shift has immediate practical implications :

• For cloud providers: Hosting Chinese open-weight models may soon carry licensing costs, altering current cloud economics .
• For MaaS businesses: Companies like Together AI, Fireworks AI, or any platform reselling inference as a service will need to watch license terms carefully. Kimi K3's $20M revenue trigger is a concrete threshold to track .
• For end users and enterprises: Internal use and direct API access through the model developer's own service are likely to remain unaffected. The new licensing targets third-party commercial hosting .

Goldman itself notes this is a description of an incentive, not a signed plan — no developer has committed to a broad paid-weights model yet . But with Kimi K3 already in the market under revenue-tiered terms, the direction is clear. As one analyst summarized, "the era of unrestricted, free-for-all open weights from DeepSeek, Qwen, and GLM may be shorter than developers assumed" .