Open weights with paid layers
Placement: a judgment drawn from the evidence, open to correction (dispute it).
The model
- Definition
- A company releases its AI model weights publicly (usable, finetuneable, self-hostable) at no license fee, and instead earns revenue from hosting, API access, enterprise subscriptions, or cloud infrastructure built around those models.
- In one line
- Model weights are released openly and revenue comes from hosting, API access, enterprise deployment or cloud.
- What compounds
- Open release drives adoption, which drives cloud and API revenue, which funds the next open model.
- Who accelerates
- Every company and country can run capable models itself; margins are lower than closed rivals.
- What it displaces
- Closed models as the only route to frontier capability.
- Stops working when
- If closed models pull far enough ahead that buyers pay the premium regardless.
Evidence
Companies 3
Evidence items 3
Grades 1A 2B
Latest dated 2026-09-11
Counter-evidence 1
Evidence items 3
Grades 1A 2B
Latest dated 2026-09-11
Counter-evidence 1
The three labels
2. Rides · loop gain positive. Each step of the curve makes its own loop stronger.
Owning the choke points · money lands in where machine choices are made and paid.
Network loop · more participants (people or machines), more value per participant.
Same rung and family
Named companies 3 companies · grade per item
| Company | Claim | Date | Gr | Source |
|---|---|---|---|---|
| Alibaba Cloud (Qwen) China · public · best A |
Cloud Intelligence Group external revenue grew 40% YoY to RMB41.6B (~US$6B) in the March quarter of fiscal 2026, with AI-related product revenue posting triple-digit YoY growth for the 11th consecutive quarter and reaching an annualized run rate of about US$5.2B (30% of Cloud's external revenue); Model Studio (Qwen-based Model-as-a-Service) saw eightfold customer growth YoY | 2026-05-13 | A | Alibaba Group |
| DeepSeek China · startup · best B |
Annualized revenue of $400-500M as of July 2026, per The Information, generated via paid API access to its open-weight models; charges $2.19 per million output tokens versus roughly $60 for a comparable OpenAI model, while maintaining 70-80% gross margins on its V4 flagship | 2026-07-15 | B | PYMNTS (citing The Information) |
| Moonshot AI (Kimi) China · startup · best B |
Revenue run rate hit roughly $1B annualized by August 2026, with a year-end 2026 target of $2B; open-weight K3 model generates 'far lower margins' than closed-weight rivals, with OpenRouter data showing about 300 billion tokens/day generated by K3 | 2026-09-11 | B | TechCrunch |
Counter-evidence Reversals, failures and disputes
| Moonshot AI's own reporting concedes its open-weight strategy runs at 'far lower margins' than closed-weight competitors like OpenAI ($40B revenue) and Anthropic ($65B revenue) cited in the same article, meaning open-weighting is a volume/adoption play rather than a margin advantage | 2026-09-11 | TechCrunch |
Sources 3
Alibaba Group alibabagroup.com · 2026-05-13
PYMNTS (citing The Information) pymnts.com · 2026-07-15
TechCrunch techcrunch.com · 2026-09-11
PYMNTS (citing The Information) pymnts.com · 2026-07-15
TechCrunch techcrunch.com · 2026-09-11