Open models as a new frontier for soft power in the US-China AI competition
B · Useful, limited A short RAND Expert Insights perspective that compares the US and Chinese AI action plans and model licences; credible institution and specific licence detail, but policy argument without original data.
Summary based on the full text·Original: about 25 minutes to read
In short
"Open Models, Soft Power, and the Spectrum of U.S.-China Artificial Intelligence Competition" is a RAND Expert Insights paper (March 2026) by Owen J. Daniels and Hanna Dohmen of RAND's Center for the Geopolitics of AGI. It argues that the US-China AI contest is moving from compute and frontier performance toward influence over the global ecosystem, where open-weight models act as soft power. Chinese developers release strong models under more permissive licences, which could lock researchers and businesses, especially in the developing world, into Chinese technology. The authors recommend stronger US support for open models, recalibrated export controls and incentives for permissive licensing.
Key numbers
- nearly 1,000%growth in DeepSeek model downloads on model-sharing platforms since R1's release in January 2025 (US CAISI)
- 63%developers in a 41-country 2025 survey who prefer open-source AI because of lower costs
- 700 millionmonthly active users above which Meta's Llama licence restricts distribution
- 70%+ vs halfemerging-market respondents vs US respondents who expect positive societal benefits from AI (Google/Ipsos)
As reported by the source.
English·Web page·opens rand.org
What it reports
- The US "Winning the Race: America's AI Action Plan" (2025) treats open models as having "geostrategic value"; China's "Global AI Governance Action Plan" frames open source as global sharing aimed at the Global South.
- China has called for open models since its 2017 Next Generation AI Development Plan; DeepSeek R1 and Moonshot AI's Kimi K2 (2025) competed with leading US closed and open models.
- US Center for AI Standards and Innovation: DeepSeek model downloads rose nearly 1,000% since R1's January 2025 release.
- Licences compared: DeepSeek R1 under MIT; Kimi K2 under a modified MIT requiring attribution for products above 100 million monthly active users or $20 million monthly revenue; Meta Llama restricts businesses above 700 million monthly active users; Google Gemma uses a proprietary licence; OpenAI's gpt-oss (August 2025, first open release since 2019) uses Apache 2.0.
- Google/Ipsos polling: more than 70% of respondents in emerging markets (Brazil, Mexico, South Africa, UAE) expect positive societal benefits from AI vs about half in the US.
- 2025 McKinsey, Mozilla Foundation and Patrick J. McGovern Foundation survey across 41 countries: 63% of developers prefer open-source AI for lower costs.
- Recommendations: encourage US firms to release open models with permissive licences, codify and fund the National AI Research Resource, partner with cloud providers on compute credits, and keep export controls while avoiding limits on inference compute that push global capital toward Chinese chips.
How they got it
A qualitative policy analysis of the two national action plans, published licence terms, and secondary data from government, polling and industry surveys. The authors acknowledge that adoption data are limited and leaderboard rankings change fast.
Limits
- No original measurement of who uses Chinese vs US open models, or why.
- Framed for US policymakers; security risks of open models are noted but not weighed in detail.
- Funded from RAND operations and philanthropic gifts listed in the paper, including donors active in AI safety.
Why it matters for companies
Model choice is now also a supply-chain and geopolitical decision. Many of the strongest open-weight models come from Chinese labs under licences more permissive than Llama or Gemma, which matters for firms that want to self-host or fine-tune. Licence terms, export controls and government pressure can change the options available during the life of a system. Procurement and legal teams should track model provenance and licence conditions as they would for any critical supplier.
Summary based on the full text. Read the original for the full argument and data.Read the original on rand.org