The India AI Governance Guidelines, released by the Ministry of Electronics and Information Technology (MeitY) on 5 November 2025, do not impose any mandatory requirements on foundation model developers. There is no registration requirement, no mandatory assessment of a model before deployment, and no requirement to disclose training compute or model architecture. The Guidelines are based on seven principles, with innovation over excessive regulation placed at the forefront. This appears to be a deliberate policy choice rather than a temporary gap before a new law is introduced. It reflects the current position of the drafting committee and sets India apart from the European Union and the United States, both of which have attempted to use computing power as a way to identify and regulate the most powerful AI models.

The Number Problem

The European Union has set a threshold of 10^25 FLOPs. Under Article 51(2) of Regulation 2024/1689, a general-purpose AI model trained using more than this amount of computing power is presumed to pose systemic risk. This can trigger notification to the European Commission under Article 52, along with additional requirements relating to model evaluation, incident reporting and cybersecurity under Article 55.

The United States took a different approach. Executive Order 14110, issued in October 2023, introduced a reporting threshold of 10^26 FLOPs. However, the Executive Order was revoked about fifteen months later by Executive Order 14179. The main lesson from the US approach is therefore not the number itself, but the fact that computing power can be used as a practical way to draw a regulatory line around the most powerful AI systems. At the same time, such a line can disappear quickly when government policy changes.

Both approaches also face the same basic problem. The amount of computing power used to train frontier AI models has increased by roughly four to five times every year for the past decade. A fixed threshold can therefore become outdated very quickly. A threshold written into law needs a mechanism for regular review, otherwise the number may lose its relevance within months. The EU AI Act recognises this problem and provides for review under Article 51(3). No comparable review mechanism currently exists in India's AI regulatory framework.

Where India Is Actually Heading

MeitY's stated position is that existing laws should apply to AI wherever those laws already cover the relevant activity. However, there are signs that a more prescriptive approach may be developing underneath this position. Reports indicate that a government committee is considering stronger AI rules. There has also been discussion within the Ministry about whether large GPU clusters should be subject to registration or end-use disclosure requirements.

This distinction is important because a model-based threshold would focus on the AI model itself, requiring developers to identify and report when their models cross a particular level of computing power. A hardware-based threshold, on the other hand, would focus on the infrastructure used to train the model and could allow the government to monitor large computing clusters and training activity more directly. This could give the state greater visibility into AI development because GPUs are imported, licensed and, in many cases, supplied by companies outside India. The hardware layer may therefore be easier for the government to monitor than the technical details of a privately developed model.

India's AI rules may consequently remain largely voluntary for some time, while the infrastructure used to build increasingly powerful AI systems could become subject to regulation much sooner. This would represent a different regulatory path from simply imposing obligations on developers after a model has crossed a particular capability or compute threshold.

The Open-Source Question

One important question is how India would treat open-weight models. The EU's approach provides an exemption for certain open-source or open-weight models, but that exemption does not apply once a model crosses the systemic-risk threshold. If India eventually introduces its own threshold, it will need to decide whether to follow the same approach or create a different one. That decision could have a major impact on Indian AI companies and research labs that are building on open-weight models.

The same is true of a possible hardware-based threshold. If India chooses to regulate large GPU clusters rather than simply regulating models after they are developed, the compliance burden could look very different from the EU model. These two questions where India sets the threshold and whether that threshold is based on models or hardware may ultimately determine the shape of India's foundation model regulation.