China's open-weight AI strategy is neither charity nor a workaround — critically examine the geopolitical logic behind this approach and its implications for India's technology sovereignty.

Open-weight models release parameters, not training data or code. China's DeepSeek, Qwen and Kimi releases are therefore best read as calculated statecraft — buying global reliance on Chinese AI foundations — offering India large but revocable gains.

The geopolitical logic

• Commoditising the complement: free weights erode the pricing power of American frontier labs while keeping the world's application layer running on Chinese architectures [1].
• Not merely chip-control evasion: cost compression came from distillation and architectural efficiency — DeepSeek R1 used roughly 2,000 GPUs against about 25,000 for comparable US training — showing capability, not compulsion [1].
• Deliberately time-limited: Chinese models trail the US frontier by about six months, and access is projected to be "graduated" around late 2028 [1].

The critical counter-view

• The 2028 date is an analytical projection, not declared policy; and the releasing firms are competitors, not a cartel — Alibaba's Qwen markets itself against DeepSeek, so coordinated closure is hard to sustain [1].
• Yet the risk is asymmetric: access is unilaterally revocable by a state with which India has unresolved border and trust disputes, and the Economic Survey warns precisely against such "fragile dependencies" [2].

Implications for India's technology sovereignty

• The saving is at the inference layer — fine-tuning and serving — and builds no domestic pre-training capability that would survive a cutoff.
• Weights without corpora cannot be audited or reproduced; alignment behaviour travels with the parameters, so local hosting relocates data, not embedded value judgements.
• The substitute lags: IndiaAI's common compute crossed 34,000 GPUs [3], while the ₹10,300 crore outlay funds compute, datasets, skilling and startups together — an ecosystem, not a frontier lab [4].

Dependence is thus manageable only if treated as a dated lease, not a windfall. India should mandate portability and open weights for IndiaAI-funded models, back the Survey's decentralised, application-driven path [2], and use subsidised compute [3] for efficiency research — converting a borrowed advantage into genuine technological self-reliance.