On Thursday, 12 November, OpenAI announced local processing and inference for its API in India. Eligible businesses and developers can have supported requests processed, and the model computation that generates the responses performed, inside the country. The offering initially supports GPT-6 Luna. It extends a data-residency option introduced in India in May 2025, which allowed eligible customers to store data here but did not cover where the model actually does its work.
That distinction is real. It is also being asked to carry more political weight than it can bear.
What was actually announced
According to The Economic Times and The Hindu, the new capability is limited in ways the applause has already begun to ignore.
It currently supports GPT-6 Luna through OpenAI’s Responses and Chat Completions APIs, for direct API usage. Eligible customers get it by creating a new API project, selecting India for both storage and processing, and using the India regional endpoint. The company has particularly pointed to financial-services customers with data-localisation requirements.
Nitin Bawankule, OpenAI’s Head of Enterprise for India, put the commercial case plainly. “Some Indian businesses have not fully deployed AI for important work because they need the information involved to be processed within the country,” he said. “Local inference gives them more control over where supported workloads run and helps address a practical barrier to adoption.”
Fair enough. A bank that was blocked from using a model because prompts were leaving the country now has a product to point to. That is a genuine convenience. It is not a transfer of control over the model, the weights, or the company that trains them.
India, Australia and Japan, OpenAI said, are among the latest markets to receive API inference residency. This is a regional product rollout, not a special compact with New Delhi. OpenAI models were also earlier made available for in-country inference through Amazon Web Services’ Bedrock platform, announced in August 2026. Foreign platforms are localising the plumbing because Indian buyers have started to ask where the data goes. That is welcome. It is still their plumbing.
A regional endpoint is not a national stack
The temptation, in a country that has spent a decade talking about Atmanirbhar technology, is to treat every localisation announcement as a sovereignty event. The two are not the same. Data residency answers a storage question: where the files sit at rest. Local inference answers a processing question: where the computation runs when a prompt becomes a reply. Both matter to companies that must show a regulator that sensitive information did not leave India during a workflow.
Neither answers the questions that actually decide who holds power. Who trained the model? Who updates it? Who sets the rules for what it will and will not say? Who can suspend the endpoint? Whose law can demand logs or cooperation in a dispute? On every one of those counts, GPT-6 Luna running in India remains an OpenAI model, offered by an American company, on terms that company writes. A subpoena, a policy change or an outage does not become an Indian event merely because the inference happened in an Indian region. The buyer has more control over the geography of a workload. The seller retains control over the stack.
The same-week policy signal New Delhi should not ignore
The timing makes the confusion harder to excuse.
Days before OpenAI’s announcement, senior government officials told Business Standard that the Centre is likely to push indigenous AI systems, large language models and agents in government-owned critical sectors, including defence and aerospace. One official said this could include asking critical departments, at the Centre and in the states, to completely avoid frontier models developed by non-Indian companies. Some sectors, including defence, are already working on air-gapped models that can run on the edge, and have been asked to review how IndiaAI Mission models might be deployed.
That is a serious policy direction, even if it is still being telegraphed through unnamed officials rather than a notified rule. An air-gapped defence model and a Chat Completions call to an India endpoint are not two versions of the same idea. They are opposite answers to the same question: who is allowed into the Indian state’s most sensitive computational rooms.
Under the Mission, Business Standard reported, the government has approved 20 models by 12 startups, including general-purpose large language models and sector-specific small ones. Sarvam and BharatGen have released several models; so have Varya, Soket’s EKA-1, and ZenteiQ’s Dhir 2B and Dhir 5B. Officials said companies including Gnani, Gan AI, Fractal, Tech Mahindra and Shodh AI are working to release models over the next 12 to 15 months.
One official added that some domestic models had achieved “up to 80 per cent of the capacity of leading frontier AI models.” That is a striking claim. It is also, so far, an unsourced official assertion, not a published, independently audited benchmark. If the models are close, show the evaluations.
Follow the money, not the slogan
The IndiaAI Mission, launched in March 2024, has a five-year outlay of ₹10,371.92 crore. That figure is not in dispute. It appears in the government’s own Rajya Sabha reply to Unstarred Question No. 1668, answered on 13 February 2026 by the Minister of State for Electronics and Information Technology, Jitin Prasada.
The same reply is less convenient when one leaves the outlay and looks at the releases. Against a revised estimate of ₹173 crore for 2024–25, IndiaAI received ₹21.79 crore. Against a revised estimate of ₹800 crore for 2025–26, it had received ₹379.15 crore as of 9 February 2026. For 2026–27, the budget estimate was ₹1,000 crore. The release column in the same table was blank. Add the two published release figures, and the total is about ₹401 crore: less than 4 per cent of the five-year outlay, as of the last official snapshot on the parliamentary record.
The reply is not a confession of failure. It also lists more than 38,000 GPUs onboarded, 14 service providers empanelled, 12 teams shortlisted for indigenous foundational models, and 27 India Data and AI Labs established, with 543 more identified. Those claims deserve to be tested. They do not erase the disbursement record.
The same was reported in April and again on 12 November. Nine months later, the November coverage was still leaning on the same official numbers. If later releases have accelerated, the government has not, in the material on record, replaced the February snapshot. A sovereignty project that will not update its own spending figures is asking the public to take a leap of faith.
The internal allocation is equally revealing. Of the ₹10,371.92 crore, Compute Capacity accounts for ₹4,563.36 crore, Foundation Models ₹1,971.37 crore, and Startup Financing ₹1,942.5 crore. Safe & Trusted AI is allocated ₹20.46 crore. Overheads and contingency, at 1 per cent, get ₹102.69 crore: five times the safety-and-trust pillar. That may be defensible in a spreadsheet. It is a strange way to advertise seriousness about models now being discussed for defence and aerospace.
A Tata partnership is not an Indian model
OpenAI’s India push did not begin this week. Earlier this year it launched an “OpenAI for India” initiative, which, the company has said, included a partnership with the Tata Group to develop domestic AI data-centre capacity.
Infrastructure on Indian soil, built with an Indian industrial partner, is useful. It is still a hosting arrangement around a foreign model family. A data centre that serves an American model does not become an Indian foundation model because the cooling towers are in an Indian state. India can rent access. It cannot outsource the stack it wants behind a defence air gap and pretend the rental is possession.
The missing doctrine, and the footage from Indian homes
The same week illustrated how confused that conversation remains.
On 12 November, MeitY Secretary S. Krishnan chaired a meeting with representatives of home-services platforms including Pronto, Urban Company and Snabbit, along with robotics companies, on first-person video and sensor data collected from gig workers to train physical AI models. The Economic Times, which first reported the consultation, said officials were concerned about workers’ rights, consumer privacy and data sovereignty, and asked whether footage was being sold or transferred to developers outside India.
Pronto has begun in-home recording through its Verified service. Urban Company and Snabbit have decided against adopting the practice. Some executives at the meeting, MediaNama reported, citing ET, criticised the idea of collecting data in India and selling it to overseas developers, saying it “did not put the country in a favourable light.”
This is not a footnote. If Indian households and workers are becoming a raw-material supply chain for someone else’s robots, the country is living the opposite of the story it tells itself about AI power. Local inference for enterprise chatbots, and the possible export of footage from Indian kitchens, can sit in the same news cycle only because New Delhi has not written a clear doctrine: what must be indigenous, what may be foreign but processed in India, and what should never leave an air gap, a home, or the country.
MediaNama has also noted a legal lag that makes that doctrine overdue. The Digital Personal Data Protection Act allows companies to send personal data abroad unless the government restricts a country, and it has not restricted any so far. Additional conditions under the DPDP Rules, the publication reported, take effect only in May 2027.
The test is not the press release
OpenAI is free to sell local inference in India. Enterprises are free to buy it. The government is free to court those investments, and to work with Tata or anyone else on data-centre capacity.
None of that relieves New Delhi of the harder job. If critical sectors are to be steered away from foreign frontier models, the IndiaAI Mission has to be funded, measured and independently evaluated, not merely announced. If “up to 80 per cent” is a real figure, it should be demonstrated. If the February spending snapshot is stale, it should be replaced. If footage from Indian homes is being considered as training fuel for physical AI, the state should decide, in public, what is permitted and what is not.
Local inference is a useful feature, not sovereignty. Sovereignty is what remains when the vendor’s endpoint is switched off. The test is whether India is building that residue, or merely renting a more flattering address for the same dependence.
