AIKosh and IndiaAI: A CTO's Guide to Building Enterprise AI in India

India has the builders. GitHub counts 21.9 million Indian developers, more than five million of them added in 2025.⁴⁵ Indian enterprises have the intent, too: in EY's survey of more than 200 enterprise leaders, 47% had several generative AI use cases live in production. Only 10% had taken them enterprise-wide, and 78% named integration and data readiness as the main barrier.⁴⁴ The IndiaAI Mission has spent two years building national parts for that work. This guide covers what each part is, how good it is in practice, and what Indian law asks of the system you build from them.

• 15,999+datasets on AIKosh on October 2, 2026, alongside 355 models from 671 organizations¹
• 38,000+GPUs empanelled on the IndiaAI Compute Portal, with a subsidy of up to 40% for eligible users¹⁵
• May 132027, the day the DPDP Rules' notice, security, breach and logging duties take effect¹⁴

What AIKosh is, and what it offers a builder

AIKosh, launched as AIKosha and often searched as AI Kosh, is the IndiaAI Mission's catalogue of datasets, models, use cases and toolkits. It opened on March 6, 2025 with more than 300 datasets and 80 models.² MeitY's latest release counted more than 15,000 datasets, 300 models and 30 toolkits in August 2026,³ and on October 2 the live portal showed 15,999+ datasets, 355 models, 22 sectors and 671 contributing organizations.¹ The IndiaAI Division under MeitY runs it, on ₹199.55 crore of the mission's ₹10,371.92 crore outlay, about 1.9%.⁴ Use has grown more slowly than supply; the latest official figure is 17,500 registered users in February 2026.¹¹

Access is easy. Anyone can browse, and downloading open items needs a free account made with a phone one-time password. Organizations register with a GST, PAN, CIN, UDYAM or TAN document.⁹ The terms are where a CTO should slow down. Contributors keep their IP and choose the license for each item, AIKosh disclaims responsibility for accuracy and license compliance, it may share a downloader's name and email with the contributor, and it forbids scraping.⁵ Quality varies widely. Much of the recent growth is bulk uploads from a few news accounts, one of them with more than 1,400 datasets, and critics noted at launch that MeitY never said how quality would be verified or how residual personal data would be removed.¹⁰¹²

How to use AIKosh data in a commercial system

1. 1Register the company Use an organization account, not an employee's personal one, and assume contributors can see who downloaded their data.
2. 2Record provenance For each item, keep the access level, the declared license, the upstream source, the version, the download date and a file hash.
3. 3Read the license Flag share-alike and non-commercial terms before anything touches a client's IP. A share-alike license passes its terms to whatever you derive from the data.
4. 4Assume personal data is present Scan health, identity and call-transcript data with an anonymizer such as Presidio or ARX before it enters a pipeline, whatever the dataset card says.
5. 5Sample before you train Check content against the metadata through the preview API, and pull data through the API rather than by scraping, which the terms forbid.
6. 6Ask early for restricted data Restricted items need the contributor's approval, and there is no published turnaround.

Personal data deserves the most care. The DPDP Act excludes only two kinds of public personal data from its scope: data a person made public about themselves, and data someone was legally required to publish.¹³ Data that others posted about a person is not clearly covered, and the Rules' research exemption, which applies from May 2027, does not open personal data to general AI training.¹⁴

IndiaAI compute: who gets subsidized GPUs

The IndiaAI Compute Portal lists capacity that approved providers offer through the mission. It passed 38,000 GPUs from 14 providers by March 2026, in data centers in Mumbai, Hyderabad, Bengaluru, Noida and Jamnagar, and a further 20,000 announced in February were still under process.¹⁵ The menu runs from inference cards to the newest training accelerators, on demand or reserved.¹⁶ The subsidy is a ceiling of 40%, subject to approval, and PIB put the subsidized rate at under ₹100 an hour, against global rates above ₹200.¹⁶¹⁷

Eligibility is the part most guides get wrong. The portal names DPIIT-recognized startups and MSMEs with AI experience and ₹50 lakh of annual revenue or ₹1 crore of funding, researchers who meet publication thresholds, students and fellows, early-stage startups with an official letter, and government bodies. It has no category for an ordinary large private company. Requests under 5,000 GPU-hours are approved automatically; larger ones go to a committee that takes applications from the 1st to the 25th of each month and publishes decisions on the 10th of the next. An approval lapses unless use starts within 30 days.¹⁸

Delivery is real and still small next to the headline. Hours sanctioned are not hours used, and no official utilization figure exists. PRS reports that the mission spent ₹19 crore in 2024-25 against a budget of ₹552 crore, and that utilization has stayed below half since launch.²¹ For an eligible startup, the portal is worth using when a job fits under the 5,000-hour line and can start within the month. A large company, a services firm or a foreign company should plan to buy from the same Indian providers, or its usual cloud, at commercial rates.

India's own models: read the license before the launch note

The mission backs 20 foundation-model proposals, 12 large models and eight small ones, and the builders keep the IP.²⁰ That clause matters more than the word sovereign. Sovereign describes who funded a model; its license decides whether you can ship it. Sarvam's 30B and 105B models, released on March 3, 2026, are mixture-of-experts models covering 22 Indian languages under Apache 2.0, and the 30B, with 2.4 billion active parameters, is built for practical deployment, including voice agents.²²²³ BharatGen's 2026 models are a different case. Their license allows non-commercial research only and requires written permission from BharatGen even to serve a model through a free hosted API.²⁴

Treat launch claims and independent tests separately. Sarvam reports strong results on its own Indic benchmark. The independent tester Artificial Analysis gave Sarvam-105B an Intelligence Index of 18, against 33 for an open-weight model of similar size from a global lab.²⁸ An Indian model can still be the right choice when data residency, cost or Indian-language speech matter, which is why the choice should rest on your own evaluation set and not on either party's leaderboard.

How AI performs in Indian languages

The gap to English is measurable. On MILU, a benchmark built from Indian state and regional exam questions, the leading model in late 2024 scored 81.75% in English and about 70% in Tamil and Odia.²⁹ On BhashaBench, which asks the same questions in English and Hindi across agriculture, finance, law and Ayurveda, the same model's Hindi accuracy trailed its English accuracy by 7 to 10 points in every domain.³⁰ Romanized, code-mixed input is harder again. In a June 2026 study of 19 models, heavy code-mixing cut accuracy by up to 17.5 points, and one widely used open model replied in English to 99.7% of code-mixed prompts.³¹ An English reply to a customer who wrote in Hinglish is a product failure even when the facts are right.

Speech is where Indian conditions are hardest. Voice of India, a 2026 benchmark from IIT Madras and Josh Talks, tested 14 speech systems on 536 hours of real 8 kHz phone calls from 675 districts. Most systems exceeded a 20% word error rate, the authors' usability line, and none stayed under it in every language. Error rates by district ranged from about 4% to 44%, and models that scored well on clean public test sets often did substantially worse on real calls.³³

Safety does not carry across languages either. IndicSafe sent 6,000 prompts in 12 Indian languages to 10 models and found that the same prompt received a different safety judgement in nearly 90% of cases when the language changed. One model refused 61.6% of harmless Punjabi prompts.³⁴ A guardrail tuned on English cannot be assumed to work on Hindi, Tamil or Hinglish traffic. Production use at scale exists, mostly where the domain is narrow and the content is grounded: Bhashini, the government's language platform, reported more than 4 billion language inferences by February 2026.³⁵

What Indian law asks of an AI system

India has no AI Act. MeitY's AI Governance Guidelines of November 2025 conclude that a separate AI law is not needed for now, govern AI through existing law and sector regulators, and point toward graded liability that rewards documented due diligence.³⁶ Several rules already bind, and the largest set of duties has a date.

The DPDP Rules were notified on November 13, 2025. Consent Manager registration opens on November 13, 2026, and the operating duties start on May 13, 2027.¹⁴ In January 2026 MeitY proposed pulling those duties forward to November 2026; no amendment had been notified by early October.⁴² For an AI system, the duties read like a specification. Notices must itemize the data and the purpose, and withdrawing consent must be as easy as giving it. Personal data and the logs of its processing must be kept for at least a year, which needs a deliberate policy for prompt and inference logs. A breach needs a detailed report to the Data Protection Board within 72 hours. Anyone under 18 is a child, and processing a child's data needs verifiable parental consent. A Significant Data Fiduciary must run a yearly impact assessment and audit, and check that its algorithmic software is not likely to put people's rights at risk.¹⁴

Penalties reach ₹250 crore for failing to protect personal data, and the Act covers processing outside India when it relates to offering goods or services to people in India, so a global company serving Indian users is in scope.¹³ The Data Protection Board had no appointed members as of August 1, 2026.⁴³ A system launched today will still be running when the duties start, so the cheaper path is to build them in now.

Foreign model APIs are allowed. The DPDP Act lets the government restrict transfers to named countries, and no such list exists.¹³ The hard limits are sectoral. A payments business must keep payment data in India and delete anything processed abroad within 24 hours, and SEBI leaves liability for an AI tool's output with the regulated firm, never with the model vendor.⁴¹³⁹ In practice, the regulator of your customer, more than any benchmark, sets the hosting envelope.

A build sequence that holds through May 2027

Data residency and evaluation decide more of an Indian AI system's success than the choice of model. The way an Indian-language system is tested is covered in LLM and AI agent evaluation, checking whether a dataset is fit for a use case in AI data readiness, and the EU's comparable rulebook in EU AI Act compliance.

We build these systems as part of our AI development work: Indian-language evaluation sets scored by native speakers, models chosen per language on measured results, hosting matched to the data class, and the DPDP and CERT-In records designed in from the first release.