India's market regulator wants a kill switch, and the state wants its own GPUs

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

Speaking at a brokers' convention in Mumbai on 24 October, SEBI chairman Tuhin Kanta Pandey said the regulator is exploring an AI sandbox and finalising responsible-AI guidelines, Business Standard reports. The framework is tiered and covers accountability, data controls, kill switches and human oversight, in line with the IOSCO toolkit. His warning to market intermediaries was blunt: "Algorithms now drive a growing share of trades. Artificial Intelligence is reshaping how people invest... Test your systems. Plan for failure."

Separately, on 7 October C-DAC issued a request for empanelment of AI service providers under the IndiaAI Mission, with a closing date of 29 October. As Medianama reports, it seeks up to 10,000 AI compute units for five years, deployable in clusters of 4,000, 1,000 or 512 units, and requires data, logs, backups and disaster recovery to stay in India. Minister Ashwini Vaishnaw is quoted: "Our demand estimate is for 10,000 GPUs, and the first tender that we are coming out with is for 5,000."

Why it matters

Taken together, these show India building both the rules and the infrastructure for AI in regulated markets. A kill-switch requirement turns a vague principle, human oversight, into something that can be inspected: who can stop the system, how fast, and whether it has been tested. Brokers running algorithmic or AI-assisted strategies should expect to show the switch works, not just that it exists.

A kill switch is only as good as the person allowed to pull it, and the drill that proves it works.

The compute tender matters for the same institutions. Data residency is often the first objection a compliance team raises about AI. Sovereign capacity with in-country logs and disaster recovery removes one excuse, though price and access terms will decide whether regulated firms actually use it. My view is that a sandbox plus local compute could let Indian firms test agentic tools under supervision rather than in the dark, which is better than the alternative many US banks chose this week: blocking them outright.

What I'd watch

• The text of SEBI's draft guidelines, especially how "kill switch" and human oversight are defined for each tier.
• Who is eligible for the AI sandbox, and whether agentic trading tools are in scope.
• Who is empanelled under the C-DAC process, at what price, and whether regulated entities can use the capacity.