RAND Corporation (兰德) report: open-source models are now the soft‑power battleground in US–China AI rivalry

Key finding: open-source is geopolitics

A major report from RAND Corporation (兰德) frames the next phase of US–China AI competition not as a race for raw capability but as a contest to shape the global AI ecosystem — and it argues open‑source models are the prime tool. Who sets default model architectures, licensing norms and developer ecosystems will influence future technical standards and the information environment in which societies operate. It has been reported that RAND calls this the “geostrategic value” of open source: once a model becomes the global default, its design choices and embedded priorities travel with it.

Evidence on the ground: Chinese momentum, developer gravity

China’s recent wave of open releases — notably DeepSeek (深寻)’s R1 and Moonshot AI (月航AI)’s Kimi K2 — reportedly matched top‑tier Western models on performance while adopting permissive licenses. U.S. National Institute of Standards and Technology data showed R1 downloads on sharing platforms surged nearly 1,000% after release. A 2025 developer survey across 41 countries found 63% of developers prefer open models for cost and visibility of weights; combined with permissive licences like MIT and Apache 2.0, RAND says this creates a powerful diffusion engine. Why do developing markets matter? RAND highlights surveys showing strong optimism about AI in Brazil, Mexico, South Africa and the UAE — fertile ground for systems that are easy to adopt and adapt.

Policy prescriptions and geopolitical tradeoffs

RAND recommends a two‑pronged U.S. response: build attractive alternatives and tighten chokepoints. On the supply side it urges expansion of the National AI Research Resource (NAIRR), subsidised cloud partnerships with AWS, Azure and Google Cloud, and a federal “trust” rating for open‑source models tied to transparency and license openness. On the restriction side, the report argues the U.S. should recalibrate export controls — shifting attention from training compute (already partially circumvented by distillation and algorithmic efficiency) toward limiting large‑scale inference capacity that fuels global deployment. RAND cautions, however, that heavy‑handed controls could accelerate Chinese chip self‑reliance; export policy must be precise, not wholesale decoupling.

Bigger picture: narrative, not just hardware

RAND’s core judgment is stark: China’s open‑source push is less about capturing market share than about embedding technical standards and narrative framing into global AI infrastructure, much as Beijing did with 5G and data‑center investments. It has been reported that Beijing’s “open and shared” diplomatic messaging is a strategic tool to lower adoption barriers and create path dependencies in the Global South. Ultimately, the report argues, technology alone won’t win soft power — the U.S. must couple competitive, permissive alternatives with political and economic narratives that resonate abroad, or risk ceding the default architecture of tomorrow’s AI stack.