Why in News

• Indian start-ups are increasingly using Chinese open-weight AI models like DeepSeek, Qwen and Kimi because they are cheaper than many US models. However, China may restrict the global sharing of AI model weights and training data, raising concerns for India about AI dependence, costs and digital sovereignty if access becomes limited.

Why Indian Firms Are Switching

• Chinese open-weight models perform almost as well as American frontier models, lagging by roughly six months, but are free to download and cheap to run — removing the recurring API bill that dominates an Indian start-up’s cost sheet.
• DeepSeek reportedly trained its R1 model for about $294,000 — a tiny fraction of what leading American labs spend — enabled by distillation and architectural efficiency gains.
• India is largely an AI adopter and integrator, not a frontier model builder — so cheaper foundations directly translate into faster diffusion.

Dimension 1: Why China Gives Its Models Away — The Five Logics

Dimension 2: What Would Make Beijing Close the Gates

• Consolidation: the AI industry becomes concentrated among a few major companies, making it easier for government to control them.
• Lock-in: global users become highly dependent on Chinese AI models, cloud services and tools, making it harder to switch to alternatives.
• Saturation: Chinese AI models become widely adopted, while other open-weight models from the US and elsewhere can no longer sustain competition.

Dimension 3A: Opportunities for India (Short Term)

• Cheap models let Indian firms deploy AI in agriculture, health, education, logistics and citizen services at affordable price points, lowering entry barriers for start-ups beyond metro hubs.
• Open weights can be fine-tuned on Indian languages — essential for a multilingual population and last-mile governance delivery, strengthening India’s “adoption-first” strategy.

Dimension 3B: Risks and Vulnerabilities (Medium to Long Term)

• Technological dependence: building critical services on a foreign stack creates switching costs that grow silently over time.
• Strategic dependence on a competitor: India’s dependence is on a country with an unresolved boundary dispute and a large trade deficit — as pharmaceutical APIs and rare earths have already shown, supply chokepoints can become leverage.
• Data and security concerns: training-data bias, censorship embedded in model behaviour, and the possibility of backdoors or covert data transfer in hosted versions.
• Governance gaps: who is liable when an open-weight foreign model produces harmful, discriminatory or false output in a public service?
• Innovation trap: if the cheapest path is always to integrate someone else’s model, incentives to build domestic foundational capability weaken.
• Sudden-stop risk: if restriction arrives around 2028-29, the cost shock will hit exactly when AI has become embedded in critical workflows.

• Openness as a strategy: open-weight AI is not simply about promoting free technology — China can use it to expand global adoption, extend influence and pressure competing AI companies.
• Unintended effect of US restrictions: US restrictions on advanced chips pushed Chinese firms toward efficient architectures and model distillation — restrictions can produce unexpected technological responses.
• Atmashakti, not complete self-reliance: India doesn’t need total AI self-sufficiency — it should build strength in compute, data, chips, foundational models and Indian-language AI, while using global technologies where useful.
• Focus on AI diffusion: priority should be affordable, widespread AI adoption in agriculture, health, education and governance, rather than frontier AI for prestige alone.

Challenges Before India