When Will China Block Its Open-Weight Models

Foreign Policy & Geopolitics
Published

August 9, 2026

Foreign Policy, Defence & Geopolitics technology-geopolitics AI

The playable strategy is now public. The genie is out of the bottle — this approach cannot be reversed through policy changes alone.

Origin

Developed by Pranay Kotasthane and Anupam Manur in Anticipating the Unintended #356 (August 2026), analysing why China permits open-weight AI model exports and when restrictions might emerge.

What it says

China exports open-weight AI models for five reasons:

  1. Lower costs — Chinese firms spend dramatically less on training. DeepSeek’s R1 cost only $294,000 through distillation and mixture-of-experts architecture.
  2. Geopolitical prestige — open-weight models demonstrate frontier AI competitiveness, raising China’s international influence.
  3. Undermining competitor business models — commoditising rivals’ core products compresses American labs’ pricing power.
  4. Capital absorption capacity — state-led investment mechanisms and financial repression provide artificially cheap capital; over $150 billion invested in semiconductors alone since 2014.
  5. Infrastructure play — free models drive global AI adoption, creating demand for energy, compute, and hardware in sectors China dominates. Alibaba’s cloud revenue grew 34% year-on-year with triple-digit AI-related growth from developers using free Qwen models on paid cloud infrastructure.

Three thresholds must be crossed before graduated restrictions emerge:

  1. Domestic consolidation — the “Hundred Model War” must consolidate to an oligopoly of 5–8 full-stack giants. Underway.
  2. Ecosystem lock-in — global developers must be deeply integrated with Chinese cloud infrastructure, creating high switching costs. Requires 12–18 additional months.
  3. Commoditisation saturation — further open releases must provide minimal strategic value, with non-Chinese labs independently sustaining commoditisation. Not yet reached.

Timeline: graduated restrictions most likely beginning late 2028 — delayed releases, capability-based licensing, channel-conditioned releases, and tiered access tied to diplomatic alignment.

Applied

  • When Indian companies adopt Chinese LLMs to reduce costs — the strategic imperative is to avoid lock-in to any single source while building domestic capacity in applications, industrial data, and domain-specific fine-tuning.
  • When designing an AI strategy that exploits the current open window while preparing for its closure.
  • When distinguishing between Chinese hardware risks (genuine security concern) and Chinese model risks (manageable through fine-tuning and inspection).

When it falls short

The three-threshold model assumes rational state behaviour; a geopolitical crisis or domestic political shift could accelerate restrictions before the thresholds are met. The framework also assumes China’s open-weight models remain competitive — a rapid American breakthrough could change the calculus.