Calls for greater restraint around powerful AI systems are getting louder as concerns grow over autonomous agents, cyber risks, deepfakes and increasingly capable models. India, however, is entering this debate while still expanding compute access, building indigenous models and trying to turn AI adoption into domestic capability. Slowing frontier AI and slowing an entire AI ecosystem are not necessarily the same thing.
Artificial intelligence spent the past few years being discussed largely in terms of speed. Faster models. Faster coding. Faster content. Faster research. Faster automation.
In 2026, another question has moved closer to the centre of the conversation: should some parts of AI development be moving this quickly at all?
The concern is not without evidence. The International AI Safety Report 2026, developed with guidance from more than 100 experts and an advisory panel drawing nominees from over 30 countries and international organisations, says capabilities in general-purpose AI are improving rapidly while the frameworks used to assess and manage their risks remain immature. It also points to an “evaluation gap”, where performance in controlled testing does not always predict how systems will behave in the real world.
The debate, however, becomes more complicated when it reaches countries that are still building their AI foundations.
India is not starting from the same position as the United States or China. It is expanding access to GPUs, backing indigenous foundation models, training talent and pushing AI deeper into enterprises and public services. The IndiaAI Mission alone carries an approved outlay of ₹10,371.92 crore, while the government said in March that more than 38,000 GPUs had been onboarded for its common compute facility.
For India, therefore, the question may be less about whether AI needs guardrails and more about where those guardrails should be placed.
A pause on highly capable frontier systems is one proposition. Slowing access to compute, skills, local models and practical AI deployment is another.
Here are four reasons why that distinction matters.
1. Not every call to slow AI is a call to stop using it
The phrase “slow down AI” makes a complicated debate sound deceptively simple.
There are researchers concerned about hypothetical superintelligence. Others are focused on current frontier models acquiring stronger cyber, biological or autonomous capabilities. Regulators are examining deepfakes, discrimination, privacy and accountability. Enterprises, meanwhile, are trying to determine whether AI can summarise documents, analyse customer data or make employees more productive.
These are not equivalent risks.
A frontier model capable of independently performing sophisticated cyber tasks presents a different policy problem from an Indian startup building speech recognition for regional languages. An autonomous agent receiving access to critical systems requires different safeguards from an enterprise using AI to classify customer queries.
That distinction is increasingly important because the global safety conversation itself is becoming more specific.
The International AI Safety Report focuses on rapidly developing general-purpose systems, including language, vision and agentic models. It does not argue that every AI application should stop. Instead, it examines mechanisms including model evaluations, dangerous-capability thresholds, monitoring and conditional safety commitments.
The report also acknowledges the difficulty policymakers face. Regulating too early can entrench ineffective interventions, while waiting for definitive evidence can leave societies exposed to serious harms.
India is part of that conversation rather than outside it.
In his foreword to the 2026 report, Electronics and Information Technology Minister Ashwini Vaishnaw wrote that global AI risk-management frameworks remain immature and that evidence gaps around advanced systems need attention.
“These gaps must be addressed alongside innovation,” he said.
The word alongside matters.
India’s AI strategy so far has largely concentrated on expanding the infrastructure beneath the technology rather than simply trying to produce the world’s largest model.
The IndiaAI Mission covers compute infrastructure, indigenous models, datasets, application development, startup financing, skills and a Safe and Trusted AI programme.
By March 2026, more than 38,000 GPUs had been brought into the common compute facility for use by startups, researchers and academic institutions at subsidised rates. That expansion matters because access to computing power has become one of the biggest barriers to serious AI development.
A university research group developing an Indian-language model needs compute. So does a startup training an agriculture model on local crop data or a healthcare company experimenting with medical-document processing.
Restricting risky deployment does not necessarily require restricting the capacity to build and test those systems.
Prime Minister Narendra Modi made a similar distinction at the India AI Impact Summit in February.
“Technology is powerful, but direction must always be set by humans,” he said.
That frames the Indian problem more accurately than a binary choice between accelerating AI and stopping it.
The technology can keep developing while the rules governing its use become stricter.
2. India is still building what established AI powers already have
There is another problem with treating an AI slowdown as though every country would be starting from the same line.
They are not.
The global AI economy is already highly concentrated around countries and companies with enormous advantages in capital, computing infrastructure, semiconductor access and research talent.
That means any slowdown would begin after those advantages have already accumulated.
For India, affordable compute is therefore not simply an infrastructure programme. It is part of reducing the cost of entering the AI market.
The IndiaAI Mission was originally approved with an outlay of ₹10,371.92 crore over five years. Its remit stretches from computing capacity and foundation models to datasets, startup funding, skills and responsible AI.
The scale of the compute programme has also expanded substantially from its initial ambition of more than 10,000 GPUs. The government said in March that the common facility had crossed 38,000 GPUs.
That does not suddenly place India on equal footing with the world’s biggest AI infrastructure markets.
It does, however, make experimentation possible for organisations that would struggle to finance large computing requirements independently.
The same logic applies to indigenous models.
India’s linguistic and economic diversity creates problems that cannot always be solved simply by importing a general-purpose model trained primarily on global or English-heavy datasets.
A customer speaking a regional language to a banking assistant, a farmer asking a crop question, a citizen navigating a government service and a business analysing an Indian regulatory document all require systems that understand local language and context.
This is why domestic model development is not only about technological nationalism or competing on benchmark rankings. Part of the case is much more practical: localisation.
If Indian organisations do not build that capability, AI will not disappear from the country. Businesses and consumers may instead become more dependent on models, infrastructure and interfaces developed elsewhere.
That creates a second form of concentration.
The risk is not only that India lacks a frontier model. It is that increasingly important layers of its digital economy could depend on technology whose training choices, commercial terms and product priorities are decided outside the country.
In that context, slowing domestic capability building could have consequences long after the current AI cycle.
3. India has AI adoption. It still needs deeper AI capability
The third reason is visible inside Indian companies.
India is no longer merely experimenting with AI.
Deloitte’s State of AI in the Enterprise 2026 found that 40% of Indian respondents reported significant or full AI usage, compared with roughly 28% globally.
At-scale deployment was particularly high in product development at 62%, strategy and operations at 56%, and marketing and sales at 55%.
Investment intentions were even stronger. Ninety-four per cent of Indian organisations surveyed expected their AI spending to increase over the following year.
Those numbers suggest that companies are moving beyond the phase in which generative AI is an isolated experiment run by a small innovation team.
But the same research exposes an important weakness.
Depending on the AI capability measured, only 0% to 4% of Indian organisations reported the highest level of expertise.
India may therefore be adopting AI faster than it is building deep expertise around it.
That gap matters.
Buying access to an AI assistant is relatively easy. Building the institutional capability to evaluate its output, integrate it with proprietary data, secure it, measure returns, redesign workflows and understand when it should not be trusted is considerably harder.
“Ambition is translating into enterprise-wide execution,” S Anjani Kumar, Partner at Deloitte India, said while discussing the findings.
But execution alone will not determine who captures the long-term value.
If companies rely predominantly on externally built tools without developing internal expertise, India could become a large market for AI without becoming equally important in the creation of AI technology.
That distinction also extends to employment.
AI will automate parts of existing jobs, including some entry-level and repetitive work. The transition is unlikely to be painless. Companies may need fewer people for certain tasks while demanding more AI, data and judgement skills from the workers they retain.
Slowing AI adoption does not necessarily prevent that transition. Global software and business models can still change around Indian workers.
Building domestic expertise may instead determine whether India participates primarily as a consumer of that transformation or as one of its developers.
This is why skills, research capacity and access to experimentation matter alongside enterprise adoption.
The competitive question is not simply how many Indians use AI.
It is how many can build it, evaluate it, adapt it and decide where it should be deployed.
4. Keeping the AI push going cannot mean AI at any cost
The argument for continued investment becomes weaker if it is interpreted as an argument for unrestricted deployment.
There are already enough warning signs to make that position difficult to defend.
General-purpose models can hallucinate. AI-generated impersonation is complicating fraud and misinformation. Automated systems can reproduce bias in their underlying data. Agentic systems introduce new questions when software is allowed not merely to recommend an action, but to execute one.
There are also less visible concerns around privacy, cybersecurity, energy consumption and the concentration of computational power.
The International AI Safety Report 2026 is particularly cautious about how much is still unknown.
Researchers cannot always predict which behaviours will emerge during model development. Existing evaluations do not consistently predict real-world performance. Developers also hold substantial proprietary information about training data, internal testing and deployment that governments and independent researchers may not see.
The report therefore describes what amounts to a policy dilemma: acting with incomplete evidence can produce poor regulation, but waiting for perfect evidence can leave society exposed.
That problem becomes more important for India precisely because adoption is accelerating.
An inaccurate chatbot producing a poor restaurant recommendation is one category of failure. An inaccurate system influencing a loan, medical decision, educational assessment or government service is another.
Scale can magnify benefits, but it can magnify errors too.
India’s AI programme has started acknowledging this through its Safe and Trusted AI pillar, which includes work around areas such as bias mitigation, privacy, explainability, model auditing and other responsible-AI mechanisms.
That may ultimately be as important as the GPU numbers.
The more AI enters high-impact sectors, the less useful a simple “move fast” philosophy becomes.
Healthcare systems may need clinical validation and human oversight. Financial models may need bias and explainability checks. Government applications may require stronger accountability and appeal mechanisms. Advanced agents may need restrictions on what systems and data they can access without human authorisation.
Frontier models demonstrating dangerous capabilities may justify stricter evaluation or delayed deployment without requiring an Indian university to abandon an Indic-language research project.
That is where the global slowdown debate could prove useful for India.
It forces the country to think about safety before its AI ecosystem becomes much larger.
But safety and capability do not have to be built sequentially.
India does not need to first build AI and then decide how to govern it. Nor does it necessarily need to stop building until every governance problem has been solved.
The two systems can develop together.
That means expanding compute while expanding independent testing. Supporting domestic models while developing evaluation standards. Encouraging enterprise adoption while requiring stronger controls in high-risk applications. Training engineers while also training auditors, policymakers and workers who will have to supervise automated systems.
The global argument around slowing AI is therefore more nuanced for India than the headline suggests.
A blanket pause would not return the global industry to a level playing field. Existing leaders would retain their infrastructure, models, capital, talent and data advantages.
At the same time, ignoring safety in the name of catching up could create a different disadvantage by weakening trust in the technology India is spending heavily to build.
India’s choice is not simply between pressing the accelerator and hitting the brakes.
The more consequential question is where each is needed.
Frontier systems with poorly understood capabilities may require more caution. High-risk applications may require stronger testing and human accountability. Deepfakes, privacy breaches and discriminatory algorithms require safeguards.
But compute access, Indian-language models, AI research, skills, startups and responsible enterprise experimentation address another problem: ensuring India develops enough capability to participate meaningfully in an AI economy that is already taking shape.
For a country still closing gaps in infrastructure and specialised expertise, moving too quickly carries risks.
So does standing still.
Disclaimer: All data points and statistics are attributed to published research studies and verified market research. All quotes are either sourced directly or attributed to public statements.