AI in India: Deep Dive

Facts in this report are as of 12 October 2026. It draws on gazette texts, regulator and company pages, law-firm commentary and news coverage gathered on 11 October 2026, and its key facts were re-checked against primary sources on 12 October 2026. Several Indian government portals (PIB, MeitY, indiaai.gov.in) refused access or did not load; the IndiaAI compute portal (compute.indiaai.gov.in) was read on 12 Oct 2026. The report says where a gap remains. Rupee and dollar conversions use ₹96 per US dollar, the X-Rates monthly average for October 2026 (X-Rates monthly average, 2026); 1 lakh is 100,000 and 1 crore is 10 million. Source labels: primary means the official text was read; secondary means news, law-firm or vendor coverage; single source, Wikipedia only and company claim mean what they say. This is research, not legal advice. Related documents in this set: Regulation and security, Costs, Hardware, Companies, Countries overview, Asia-Pacific and Indian languages (models, speech and OCR).

Key takeaways

• The hard deadlines are in data-protection law, not AI law. India has no AI statute. DPDP duties phase in on 13 Nov 2026 (consent managers only) and 13 May 2027 (notice, consent, security, 72-hour breach reports, rights). Security failures carry penalties of up to ₹250 crore (about $26 million).
• Platform, media and security rules bite sooner. Businesses that Meta bills with India as the "Sold-To" country must move every WhatsApp Business account to rupee billing by 31 Dec 2026, or Meta stops delivering their messages from 1 Jan 2027. Synthetic audio and video have needed labels since 20 Feb 2026. CERT-In's binding 6-hour incident reporting (since 2022) covers attacks on AI and ML systems.
• Subsidised compute is real but narrow. The IndiaAI compute portal lists 45,081 empanelled GPUs. An H100 SXM lists at ₹153 per GPU-hour (GST extra), or ₹91.80 (about $0.96) after the maximum 40% subsidy, against ₹217–362 for commercial on-demand H100 hours; the minister's "about ₹65" cannot be reproduced from the portal. Eligibility is by named category (researchers, DPIIT-recognised start-ups and MSMEs, students, government entities); large enterprises and services companies are not one.
• The February 2026 AI Impact Summit was declaratory. Its declaration, endorsed by 88 countries and organisations, is non-binding. "Over $200 billion" of investment is an official expectation, and company pledges are pipeline, not capital.
• Sarvam has the broadest verifiable 2026 output: open-weight 30B and 105B models (Apache-2.0, March 2026), a $1.5 billion valuation, and rupee-priced speech and LLM APIs. Gnani's 30B model (Apache-2.0, 10 Aug 2026) and BharatGen's Param2-17B (Feb 2026) are the other verifiable 2026 open releases. Most other "sovereign" claims are early.
• In-country processing exists only for named models on named clouds. No vendor's own API processes requests in India. AWS Bedrock offers India-only routing for the previous Claude generation (Opus 5, Sonnet 5, Haiku 4.5), Kimi K3 and GPT-5.6 Terra and Luna, and Google processes Gemini 3.5 Flash in Mumbai; the newest models (Claude 5.5, GPT-6, Gemini 3.6 and later) have no documented India processing on any cloud.
• Usage is huge in total and average per person. India has 100 million+ weekly ChatGPT users. But 18.5% of working-age Indians use generative AI (world: 18.8%), and Anthropic's Feb 2026 brief ranked India 101st of 116 countries per person.
• IT services cut headcount in FY26 as AI revenue grew. The top five firms shed a net 7,389 staff in FY26. In Q2 FY27 (to 30 Sep 2026) TCS reported annualised AI revenue above $3 billion, more than 10% of revenue, and headcount up 4,258 to 598,056 (secondary, several outlets agree).
• For businesses in India, the near-term work is DPDP readiness, WhatsApp and Indian-language automation, and governance of staff AI use. Indian AI/ML base pay averages about ₹10 lakh ($10,700) a year, against US medians above $120,000, and is projected to grow 15–20% a year. The main new risk is model access: a June 2026 US directive took two frontier models offline for 19 days.

India bets on cheap compute, home-grown models and light-touch law

India's AI strategy in October 2026 rests on three choices. First, the state subsidises access to GPUs supplied by empanelled private providers through the IndiaAI Mission's compute portal. Second, it funds domestic foundation models: Sarvam, BharatGen, Gnani.ai, Soket and Gan.ai were selected under the same mission. Third, it regulates through existing law instead of a new AI statute. The November 2025 AI Governance Guidelines say so explicitly, which in practice leaves binding obligations to the DPDP Act, the IT Rules and sector regulators such as the RBI and SEBI. The electronics and IT minister, Ashwini Vaishnaw, framed the goal at Davos in January 2026. He said India is "clearly in the first group" of AI powers and will become the largest supplier of services at the application layer of the AI stack (Business Today, Davos, 21 Jan 2026). That is the layer at which most business automation is built and bought, so national policy explicitly backs it.

Private capital moves faster than policy. Indian conglomerates and US hyperscalers have announced gigawatts of data-centre capacity, OpenAI and Anthropic both call India their second-largest market, and India's largest IT firms are becoming channel partners for US model vendors. The national averages are less striking: per person, Indian generative-AI use sits at the world average, and the IMF scores India's AI preparedness below most large emerging-market peers. A very large, young, software-heavy user base sits on top of a much larger population and SME economy that has barely started.

For Indian businesses, and for foreign companies that work with Indian teams, three things follow, and the rest of this report sets out the evidence. Adoption is broad but shallow. The dated obligations concern personal data and synthetic media rather than "AI" as such. And unit costs are increasingly priced in rupees: GPU-hours, speech minutes, WhatsApp messages and consumer AI plans.

The IndiaAI compute portal lists 45,081 empanelled GPUs, allocated by application

The IndiaAI Mission's outlay is ₹10,371.92 crore (about $1.08 billion) according to Wikipedia (Wikipedia, Artificial intelligence in India), and a report built on a MeitY statement rounds it to ₹10,372 crore (TelecomTalk, 25 Mar 2026). Its most visible output is a shared GPU pool. A MeitY statement dated 25 Mar 2026 said more than 38,000 GPUs had been onboarded through the AI Compute Portal and were being provided "to Indian start-ups and academia at an affordable rate" (TelecomTalk, 25 Mar 2026; re-read 11 Oct 2026). The same report lists 190 approved projects, the largest groups being 78 with government entities and 46 with start-ups and MSMEs. It is a single report; the PIB release could not be read. The portal's own empanelled-GPU list now totals 45,081 GPUs across 14 bidder rows (undated PDF; column sums check): H100 16,496; B200 8,192; H200 7,916; TPU v6e (Trillium) 5,548; L40S 1,973; L4 1,608; others smaller. Yotta is the largest bidder with 17,408 (IndiaAI, empanelled GPUs) (primary). The 18,693 (an early launch report) and 34,381 (a later PIB factsheet; Communications Today, search summary) and 38,000+ figures are earlier announcements. On 20 Feb 2026, Jayant Chaudhary, Minister of State for Skill Development and Entrepreneurship, said compute access in India "currently costs around Rs 65 per hour" and that the mission aims to scale to more than one lakh (100,000) GPUs through a public-private model (Storyboard18, 20 Feb 2026) (secondary).

Capacity on the portal is not the same as use. A Lok Sabha disclosure reported in August 2026 put 93.18 lakh GPU-hours sanctioned across 237 approved projects, with 15 compute service providers empanelled (Communications Today) (secondary; the parliamentary record was not read). By our arithmetic, 9.318 million GPU-hours equals about 1,060 GPUs running for a whole year, or about 2.4% of the 45,081 empanelled GPUs; the period over which those hours were sanctioned is not stated, so the comparison only shows direction. Even so, the "38,000" or "45,081 GPUs" headline describes what is listed, not what is busy, and the gap between the two is the most common way this programme is misquoted.

The portal's own rate card now sets the price. The IndiaAI compute portal publishes list rates before subsidy (calculator footer "Last Updated 08-10-2026"; Ready Reckoner rates as of 13 Mar 2026; GST extra): H100 SXM ₹153.0 per GPU-hour on demand (₹134.1, ₹125.1 and ₹117.0 on 1-, 6- and 12-month reservations); H200 SXM ₹140.0; B200 SXM ₹290.7; B300 SXM ₹351.0; A100 80GB ₹135.9; RTX PRO 6000 ₹114.6; L40S ₹67.5; L4 ₹44.86 (IndiaAI price calculator; IndiaAI Ready Reckoner) (primary). The subsidy is "Max upto 40% subject to approval from IndiaAI", so an H100 SXM hour at the maximum subsidy is ₹91.80 (about $0.96). Vendor guides' "₹92" (DEV Community; Tensordata) is therefore the post-maximum-subsidy level, not a pre-subsidy price. The minister's "around Rs 65" and the guides' ~₹67 would need about 57% off the H100 list and cannot be reproduced from the portal; this report does not use them as a rate. Business Standard reported access at about ₹65 per GPU-hour around 29 Jun 2026 (Business Standard), but the page returned an access error and was not read. Vaishnaw has said AI compute in India costs less than half the global market rate (Outlook Business).

Eligibility is by named category (IndiaAI eligibility criteria) (primary): researchers and academia (h-index ≥5, or publication in an AI/ML journal with impact factor above 2.5, or 150+ citations); start-ups and MSMEs with DPIIT recognition and either ₹50 lakh annual revenue or ₹100 lakh funding; India AI fellows; students; early-stage start-ups (with a letter from the IndiaAI CEO or a JS-rank officer); government entities; and early-stage researchers. Large enterprises and services companies are not a named category, so on our reading they qualify only as DPIIT-recognised start-ups or MSMEs. Requests under 5,000 GPU-hours are auto-approved and larger ones go to the PMEC committee; requests are submitted on the 1st–25th of a month, the approved list appears on the 10th of the next, approvals are valid 30 days, and users pay the provider beyond the subsidy. Login is via DigiLocker, e-Pramaan or Jan Parichay. A business planning GPU-heavy work should budget ₹153 per H100 SXM hour before subsidy and at best ₹91.80 after the maximum 40%, and should first confirm that it falls in a named category, for most businesses as a DPIIT-recognised start-up or MSME.

Twelve model-building teams, three launched at the summit

Sarvam AI was the first foundation-model grantee. It was selected in April 2025 from 67 proposals and given government-supported GPU access (Business Today, 16 Mar 2026; Wikipedia, Sarvam AI). A second round on 30 May 2025 added three teams: Soket AI (a planned open-source 120-billion-parameter model), Gnani.ai (a 14B voice model) and Gan.ai (a 70B multilingual model focused on text-to-speech). The mission had received 506 applications by 30 Apr 2025 (The Week / PTI, 30 May 2025; Inc42). Pre-announcement press reports, citing unnamed DST sources, list twelve companies across two phases, adding Avaatar, BharatGen at IIT Bombay, Zenteiq, Gen Loop, Intellihealth, Shodh AI, Fractal Analytics and Tech Mahindra Maker's Lab (secondary; no official list was read). Separately, the minister told Parliament in July 2026 that 20 sovereign-model proposals (12 LLMs and 8 SLMs) had been identified for support (Indian Infrastructure, 27 Jul 2026) (secondary). Public support for Sarvam was reported at about ₹246.72 crore (about $25.7 million) in a summary citing PIB (secondary), and three sovereign models, from Sarvam, Gnani.ai and BharatGen, were unveiled at the February summit (Free Press Journal). The models section below shows what each has actually shipped.

New leadership, and data platforms with no public usage figures

The mission changed leaders in 2026. Abhishek Singh, the IndiaAI Mission CEO, became Director General of the National Testing Agency from 1 April 2026 (Entrepreneur India, 2 Apr 2026; Storyboard18). Saurabh Vijay, CEO of UIDAI, was given additional charge of the IndiaAI Mission after Singh's transfer (order dated 31 Mar 2026) (Elets eGov, June 2026) (secondary; indiaai.gov.in blocked), so directories that still list Singh as mission CEO are out of date. Vaishnaw is still shown as Cabinet Minister for MeitY (Wikipedia, MeitY) (Wikipedia only). The mission's data and skills platforms could not be measured: the AIKosh dataset portal exposes no counts, and the FutureSkills pages were unreadable. Bhashini, the government translation platform launched in July 2022, has MoUs with Snapdeal, the Information and Broadcasting ministry and Federal Bank and underpins ONDC's "Saarthi" app (Wikipedia, Bhashini), but no usage figure was found; Indian languages covers its translation role. Profiles of the officials and founders are in People.

The AI Impact Summit produced a non-binding declaration and "over $200 billion" of expected investment

India hosted the AI Impact Summit at Bharat Mandapam, New Delhi, from 16 to 21 February 2026 (Wikipedia's infobox says 16 to 20 February). Prime Minister Modi opened it on 19 February, with more than 20 heads of state and attendees including Sundar Pichai, Sam Altman, Demis Hassabis, Dario Amodei and Mukesh Ambani (Wikipedia, AI Impact Summit) (Wikipedia only). The formal output is the New Delhi Declaration on AI Impact, a non-binding text adopted by 88 countries and international organisations, with both the US and China endorsing (Fortune, 23 Feb 2026; IANS, 21 Feb 2026). A later analysis by the Observer Research Foundation gives 91, and an 89 figure also circulates; this report uses the 88 reported on the day. ORF lists outputs from seven thematic working groups, several endorsed by only 22 to 24 countries (ORF). Thirteen frontier and domestic AI companies signed the voluntary "New Delhi Frontier AI Impact Commitments" (ORF). They agreed to publish anonymised, aggregated insights into real-world AI use and to strengthen testing in under-represented languages (Fortune).

The headline money figure is an expectation, not a commitment. Vaishnaw said "over $200 billion in AI and deep-tech investment is expected in the country over the next two years" (Fortune). That is about ₹19.2 lakh crore. The "$250 billion" figure found in search summaries is not supported by any page read and should not be used. Named pledges were large. Reliance-Jio pledged $110 billion over seven years. Adani pledged $100 billion for renewable-powered AI data centres by 2035. L&T announced a venture with NVIDIA, AMD widened its partnership with TCS to up to 200 MW, and Blackstone led a $600 million investment in Neysa. OpenAI agreed to be the first customer of TCS's data-centre unit. Anthropic opened a Bengaluru office. Google announced two $30 million funds, and Microsoft restated its Global South and India commitments (see the infrastructure table) (Fortune). Each of these is single-source; the company releases were not read.

The criticism matters as much as the pledges. Fortune noted that the declaration is non-binding and that "operational details are thin", with compute, data and frontier know-how still concentrated elsewhere. Wikipedia records criticism that the event was trade-fair focused and that some non-Indian products were labelled as Indian (Wikipedia, AI Impact Summit). Vaishnaw called for AI "for, by, and of the humans" and acknowledged disruption in IT services alongside reskilling (Business Today, 19 Feb 2026). For businesses, the summit's lasting effect is commercial rather than regulatory. It brought US model vendors' India offices and Indian data-centre deals forward, and it made "sovereign" and "multilingual" features that Indian enterprise buyers now ask vendors about. The wider context is in Timeline and Landscape.

Governance stays principles-first, with hard dates only for personal data and synthetic media

The AI Governance Guidelines favour existing law and voluntary compliance

MeitY released the India AI Governance Guidelines on 5 Nov 2025 under the IndiaAI Mission. They set out seven "sutras" (principles), recommendations across six pillars (infrastructure, capacity building, policy and regulation, risk mitigation, accountability, and institutions), an action plan and implementation guidance. They prefer existing laws to a new AI statute, and voluntary compliance first (Indian Infrastructure, 7 Nov 2025; S.S. Rana). Two points carry practical weight. Using personal data to train AI must comply with the DPDP Act, and the guidelines call for grievance redress on bias, discrimination, misuse and deepfakes. This is secondary coverage; the MeitY PDF was not read.

DPDP: three commencement dates and penalties of up to ₹250 crore

The DPDP Act and Rules are the only AI-relevant law in India with dated duties confirmed from the primary text. The commencement notification, G.S.R. 843(E), and the DPDP Rules, G.S.R. 846(E), are both dated 13 Nov 2025 and were uploaded to the e-gazette on 14 Nov (Gazette, G.S.R. 843(E); Gazette, G.S.R. 846(E); MeitY, DPDP Rules G.S.R. 846(E)). The MeitY copy of the notification is identical to the Gazette file (MeitY copy). The text settles two errors that still circulate. The phase dates are 13 Nov 2026 and 13 May 2027, not 14 Nov. And the Act's Schedule sets penalty maximums ("may extend to"), so Wikipedia's statement that the "minimum penalty for breach is 50 crore" is wrong (Wikipedia, DPDP Act).

Sources: G.S.R. 843(E) and G.S.R. 846(E) (primary; both re-read 12 Oct 2026).

The Rules make the May 2027 duties concrete. Under Rule 7, a Data Fiduciary that becomes aware of a personal-data breach must inform each affected Data Principal and the Board without delay. It must then give the Board a fuller report within 72 hours, or a longer period the Board allows on written request, covering the facts, causes, mitigation, findings about who caused the breach, remedial steps and the notices sent to individuals. Rule 8(3) requires personal data, traffic data and processing logs to be kept for at least one year, and Rule 6(1)(e) requires logs and personal data needed to detect and investigate unauthorised access to be kept for one year. Rule 8 and the Third Schedule set erasure periods by class, with "three years" for several classes. Rule 10 requires verifiable parental consent for children's data. Rule 13 requires Significant Data Fiduciaries to carry out an annual DPIA and audit reported to the Board, to check that algorithmic software does not risk Data Principals' rights (r.13(3)), and to localise personal and traffic data in categories specified on a committee's recommendation (r.13(4)); the gazette does not say which entities are SDFs. On cross-border transfers, Rule 15 makes transfers abroad "subject to" requirements the Central Government may set "by general or special order" about making data available to a foreign State or its agencies, while the power to restrict transfers to notified countries is Act s.16(1). Both apply only from 13 May 2027, and no s.16 notification or Rule 15 order was found (MeitY's index was unreadable, so absence is not confirmed). Commentary treats the May 2027 deadline as firm, with no grace period expected (CyberPeace; Hogan Lovells).

Source: DPDP Act 2023, Gazette of India (primary).

Three practical readings follow. First, the 13 Nov 2026 tranche is narrow. It switches on the consent-manager provisions, not the general duties, so the common belief that "DPDP starts in November" overstates what is due this year. Second, the ₹250 crore ceiling attaches to security-safeguard failures, not to every breach. On our reading, a fiduciary that is breached and fails to notify is exposed under two separate heads of the Schedule, with maximums of ₹250 crore and ₹200 crore; confirm with counsel. Third, the Act has no "legitimate interests" or "contractual necessity" ground, and it puts no direct duties on processors; fiduciaries pass them on through processing contracts (Hogan Lovells). Reusing customer chats or records to tune or evaluate models beyond the purpose stated in the notice will therefore likely need fresh consent (our inference; confirm with counsel).

The IT Rules now require labels on synthetic audio and video

The Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026 were notified on 10 Feb 2026 and came into force on 20 Feb 2026. They define "synthetically generated information" as audio, visual or audio-visual content made or altered by a computer resource so that it appears real. Lawful synthetic content must be prominently labelled, visibly or with an audible prefix. Sources differ on how firm the metadata or provenance duty is: one says markers are required "where technically feasible", another that they are "permanent" and cannot be removed (secondary). Significant social media intermediaries (more than 50 lakh registered users in India) must also ask users to declare synthetic content and verify those declarations. Prohibited synthetic content includes child sexual abuse material, non-consensual intimate imagery, false documents and deceptive depictions of real people or events, and intermediaries that offer generation tools must use technical measures to prevent it. The deadline for acting on government or court takedown orders falls from 36 hours to 3 hours (amended Rule 3(1)(d)). Complaints about non-consensual intimate imagery, including morphed or deepfake nudity, must be acted on within 2 hours (amended Rule 3(2)(b); two secondary sources agree) (iPleaders; Gala Law; Business Today, 10 Feb 2026; Hogan Lovells, 13 Feb 2026; both re-read 11 Oct 2026) (secondary; Gazette G.S.R. 120(E) not read). A draft dated 21 Apr 2026 would require labels to stay visible for the full length of the content. Comments were due on 7 May 2026, and the draft has not been confirmed as final (S.S. Rana). The Internet Freedom Foundation argues that the rules undercut their own purpose (Internet Freedom Foundation).

Enforcement so far has been aimed at platforms. On 2 Jan 2026, MeitY's Cyber Laws Division wrote to X's India Chief Compliance Officer over obscene Grok-generated content. It demanded an action-taken report within 72 hours and warned that X could lose its Section 79 safe-harbour protection (Business Today, 2 Jan 2026; exchange4media; The Logical Indian). Around 11 Jan, "government sources" said about 3,500 items had been blocked and about 600 accounts deleted (Storyboard18); no action-taken report or enforcement decision has been published (secondary, several outlets). That month Indonesia, Malaysia and the Philippines blocked Grok temporarily (Wikipedia, 2026 in artificial intelligence); Grok has the details. For any business that produces or commissions voice bots, avatars or marketing video for Indian audiences, the safe default is to label every synthetic output. Whether a business API customer carries intermediary duties under the IT Rules was not checked and is a question for counsel.

Sector regulators: RBI has a framework, SEBI a consultation, CERT-In a binding 6-hour incident rule

RBI's FREE-AI framework is the most concrete sector document. The committee was set up on 26 Dec 2024, chaired by Dr Pushpak Bhattacharyya of IIT Bombay. RBI's press release confirms that its report, released on 13 Aug 2025, contains seven "Sutras" and 26 recommendations under six pillars: Infrastructure, Policy, Capacity, Governance, Protection and Assurance (RBI press release, primary; Storyboard18, 14 Aug 2025; Outlook Money). It is a set of recommendations, not a direction, and no RBI AI direction appears on the first page of RBI's notifications list (12 Oct 2026). A law-firm reading says several provisions are meant for Master Directions. It covers banks, NBFCs, payment system operators and RBI-regulated fintechs, with obligations expected to "flow down" to technology vendors by contract. Its touchpoints are a board-approved AI policy, stricter consent and retention rules, a documented model lifecycle, AI checks in product approval, telling customers when they deal with AI (with a way to challenge decisions), and prompt reporting of AI failures (Bar and Bench, 18 Sep 2025). These details remain secondary, because the report PDF (FREE-AI report PDF) timed out.

SEBI has consulted but not finalised. Its consultation paper of 20 Jun 2025 proposes model governance with human oversight, fallback plans and vendor agreements, plus plain-language disclosure of AI use (Lakshmisri; Taxmann). No final guideline was found.

CERT-In has one binding rule and several guidance documents. The binding rule is its Directions under s.70B(6) of 28 Apr 2022: service providers, intermediaries, data centres, body corporates and government bodies must report listed cyber incidents within 6 hours of noticing them. Annexure I item (xx) lists attacks on "systems/servers/software/applications related to Artificial Intelligence and Machine Learning", and logs must be kept for 180 days within India (CERT-In Directions, 28 Apr 2022) (primary). Its bill-of-materials guidelines, version 2.0 of 9 Jul 2025 (CERT-In, SBOM/QBOM/CBOM/AIBOM/HBOM guidelines v2.0) (primary), add an AI bill of materials (AIBOM) covering training-data sources, dataset licensing and privacy compliance, intended use, model licensing, security requirements and known vulnerabilities (Medianama, Jul 2025; AZB & Partners, 29 Jul 2025; Khaitan). Their "recommended" and "encouraged" language makes them non-binding guidance that shapes procurement, not a s.70B direction (Mondaq). Newer items are guidance or of unknown status: a late-April 2026 high-severity advisory on faster AI-enabled attacks, aimed at organisations including MSMEs (OECD.AI incident monitor); a "Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation", v1.0, dated 25 May 2026 (CERT-In Blueprint) (primary PDF; binding status not stated); and the 10 Jun 2026 "Guidelines regarding AI-Accelerated Vulnerability Protection and Response Requirements for OEMs and Technology Providers" (CISG-2026-03), whose title and date CERT-In's listing confirms (CERT-In listing). Their reported requirements (a secure development lifecycle, "both traditional and AI-assisted security testing", updated bills of materials, faster patching by severity, vulnerability disclosure) rest on one secondary summary (TeamLease RegTech); the body could not be downloaded and the binding status is unknown. ICMR's 2023 "Ethical Guidelines for Application of Artificial Intelligence in Biomedical Research and Healthcare" are guidance ("may be followed"), not law; the text's line about "legal status" refers to CDSCO's clinical-trial rules, and the guidelines page lists no newer edition (ICMR guidelines) (primary). Its site also lists a 20 Feb 2026 event on a regulatory pathway for AI-based medical devices. On DeepSeek, no Indian government notice was verified. A Finance Ministry advisory reported in February 2025 could not be confirmed; see DeepSeek.

The pattern across RBI, SEBI and CERT-In is consistent. Each asks for documentation that most SME-grade AI projects do not produce today: a written AI policy, a model and data inventory, human oversight, customer disclosure and incident reporting. A regulated business can ask its AI vendors to deliver these as standard artefacts now, before the AI-specific rules become binding. One duty already binds: CERT-In's 6-hour incident reporting, which covers attacks on AI and ML systems.

US and EU rules follow work done for those markets

Automation built in India for US users brings US and sometimes EU rules with it, whether an Indian business sells it abroad or a foreign company has it built by an Indian team. Colorado repealed and re-enacted its AI Act as SB 26-189, an automated decision-making technology (ADMT) law signed in May 2026. From 1 Jan 2027, the developer of an ADMT used in consequential decisions such as hiring, lending, housing or insurance must give the deployer technical documentation: intended uses, categories of training data, known limitations, and instructions for appropriate use and human review. Both sides must keep compliance records for three years (Colorado General Assembly, SB26-189; re-read 11 Oct 2026). The attorney general's draft ADMT and chatbot rules are open for comment until 26 Oct 2026, and a separate Chatbot Safety Act (HB 26-1263) also takes effect on 1 Jan 2027 (Colorado Attorney General). Texas TRAIGA took effect on 1 Jan 2026 and applies to developers and deployers doing business in Texas (Wikipedia, TRAIGA). California's CPPA regulations on automated decision-making technology (ADMT) took effect on 1 Jan 2026. A business already using ADMT for a significant decision before 1 Jan 2027 must comply by 1 Jan 2027 (later users from first use), and risk assessments conducted in 2026–27 are submitted by 1 Apr 2028 (CPPA approved text, primary; National Law Review). A federal executive order of 11 Dec 2025 (EO 14365) directed the Attorney General to set up a Justice Department task force to challenge state AI laws (Axios, 11 Dec 2025), but state laws stay in force while those challenges proceed. Colorado is the exception to watch: in X.AI LLC v. Weiser (D. Colo.), after the United States joined as a plaintiff on 24 Apr 2026, a minute order of 27 Apr 2026 bars the attorney general from beginning enforcement of SB 24-205, or of legislation replacing or amending it enacted that session, until 14 days after a ruling on xAI's preliminary-injunction motion, due within 28 days after final implementing rules; whether that reaches SB 26-189 is arguable (CourtListener docket, primary). In the EU, the AI Omnibus entered into force on 27 Jul 2026; the Commission's AI Act page links its final text as Regulation (EU) 2026/1744 (primary; OJ publication on 24 Jul 2026 rests on law firms). Transparency rules under Art. 50 apply from 2 Aug 2026, and Annex III high-risk duties move to 2 Dec 2027 (European Commission). Art. 50(1) requires that people be told when they are dealing with an AI system (AI Act Art. 50). A grace period to 2 Dec 2026 applies only to machine-readable marking duties for generative systems already on the market before 2 Aug 2026, and two new prohibitions, on systems built to generate realistic intimate imagery or child sexual abuse material, apply from the same date (AI Act Service Desk timeline and Commission page, primary; NatLawReview (K&L Gates), 31 Jul 2026). Whether an India-based vendor counts as a "developer" or a "deployer" under each US law depends on the contract and the facts, so ask counsel. Regulation and security and Americas and Europe cover these rules in depth.

Compliance calendar, October 2026 to 2028

Sources: DPDP Gazette notifications above; Meta WhatsApp Business Platform pricing; OpenAI, Agent Builder; Business Today on ChatGPT Go free year, 4 Nov 2025; ChatMaxima, WhatsApp API pricing India; Gallabox, WhatsApp pricing overview; CERT-In Directions, 28 Apr 2022; TRAI, comments on the draft TCCCPR Third Amendment; the US and EU sources cited in the previous subsection.

Sarvam leads India's model push; most other sovereign claims are still early

The sovereign-model programme has moved from grants to releases, but unevenly. Sarvam's 2026 output can be checked on its blog and on Hugging Face. BharatGen has a 2026 release with modest uptake; Gnani's August 2026 model has the most downloads of the Indian repos listed. Soket and CoRover rest mainly on company claims. Hugging Face download counts below were read on 11 Oct 2026 unless dated. They show developer interest, not production use or revenue.

Sources: Sarvam blog; Sarvam, 30B and 105B, 6 Mar 2026; Hugging Face, Gnani 30B card; Google, Gemini Apps languages; Claude docs, multilingual support; Claude help, interface languages; Hugging Face, sarvamai; Wikipedia, Sarvam AI; MediaNama, June 2026; BharatGen; Hugging Face, BharatGen; Gnani; Hugging Face, Gnani; Soket AI; Hugging Face, Soket; CoRover; Hugging Face, CoRover; Hugging Face, Krutrim; HN listing, Krutrim; Anthropic, Bengaluru office.

Sarvam matters most because it is the only domestic option with both open weights and a commercial API priced in rupees. Sarvam's blog (6 Mar 2026) and model cards describe both as mixture-of-experts models trained in India on IndiaAI compute and released under Apache-2.0 (vendor; primary for what Sarvam states). Its 2025 Sarvam-M model drew criticism for fine-tuning a foreign base model. A later report says NVIDIA, Activate, Glade Brook and Gaja added $74 million to complete a $300 million round (Entrepreneur India) (single source). Company-stated figures of ₹45.1 crore revenue in FY2025-26 (about $4.7 million), more than 2 million interactions a day and 10 million API calls a day are unverified (ARC coverage). So are reported MoUs with Tamil Nadu and Odisha, including a 50 MW compute facility in Odisha; Sarvam's blog does list a post on sovereign AI partnerships with Indian states (9 Feb 2026). Its price list is the practical anchor. Text-to-speech costs ₹3 per 1,000 characters, and speech-to-text ₹30 per audio hour (₹45 with diarisation) (Sarvam API pricing; re-read 12 Oct 2026). Sarvam's own 105B LLM is listed at ₹29.28 / ₹10.98 / ₹73.20 per million input / cached / output tokens (about $0.31 / $0.11 / $0.76). Document digitisation costs ₹0.50 a page, extraction ₹1.00 and translation ₹0.005 a character pay-as-you-go, on prepaid credits with no expiry (page undated; GST not stated). Rupee pricing removes currency risk on the speech leg. No independent 2026 benchmark compares current flagships (Claude 5.x, GPT-6, Gemini 3.x) with Indian models. One preprint on romanised code-mixed prompts found Claude Opus 4.6, Sarvam-30B and Gemini 3.5 Flash all losing about 7–8 points of accuracy at 75% code-mixing (arXiv 2606.30790), and another finds Indian scripts take 2.1 to 8.0 times the tokens of English on average, depending on the tokenizer (arXiv 2607.24276) (both preprints). Indian languages has the detail on models, speech, OCR and token costs. Other models and Top models compare global options.

Infrastructure: 1.6 GW built, several times that promised

India's operating data-centre base is about 1.6 GW (JLL, June 2026), and forecasters disagree on how fast it will grow. Approvals and pledges far outrun what is being absorbed. Andhra Pradesh alone has approved more than 6.5 GW of AI data-centre projects in Visakhapatnam (Wikipedia, Opposition to AI data centers) (Wikipedia only), more than JLL's 6 GW national forecast for 2029.

Sources: Fortune, 23 Feb 2026; Wikipedia, AI in India; Times of Oman; Investment Monitor; Verdict; Wikipedia, Visakhapatnam; Economic Times, Apr 2026; TechCrunch, 18 Feb 2026; Wikipedia, Amazon Web Services; Yotta.

Three cautions apply. Pledges are not additive. The $200 billion national expectation, Reliance's $110 billion and Adani's $100 billion overlap and run over different periods. Treat them as pipeline, not capital. Power and local politics are real constraints. Citizen groups have raised concerns about the Visakhapatnam projects, and a "People's Convention on AI Data Centers" on 6 Sep 2026 demanded a moratorium (Wikipedia, Opposition to AI data centers) (Wikipedia only). The Vizag campus is a 2026–2030 build. It does nothing for API latency or prices for Indian SMEs this year. Hardware covers the global accelerator picture.

Where your data is actually processed

New Indian capacity does not yet mean in-country inference from the frontier model APIs: no vendor's own API processes requests in India, and OpenAI offers India storage-at-rest residency on approval. OpenAI's data-controls guide lists an India region (in.api.openai.com) with regional storage but no regional processing. In-region processing exists only for the US, the EU and the UAE. Non-US regions need approved abuse-monitoring controls and a Modified Retention contract amendment, and where a region lacks regional processing, OpenAI "may also process and temporarily store Customer Content outside of the Region" (OpenAI, Data controls; re-read 12 Oct 2026). OpenAI also charges a 10% uplift on regional-processing endpoints for models released on or after 5 Mar 2026 (OpenAI API pricing). Anthropic's first-party API setting inference_geo accepts only "global" and "us", and US-only inference costs 1.1 times the normal price (Claude data residency docs). Its regional compliance page said residency is available in Asia-Pacific regions "including Japan, South Korea, Singapore, India, and Australia" through cloud platforms (Claude regional compliance; read 11 Oct 2026, not re-readable on 12 Oct 2026); on the documented evidence, India processing for Claude exists only through Bedrock's India profiles for the previous generation. xAI documents a US regional endpoint at 1.1 times the price for two Grok models, and no India endpoint (Grok). TechCrunch's report that HyperVault will help OpenAI meet Indian "data residency" requirements describes intent, not a product.

On the clouds, India-only processing exists for some models, not the newest. The cloud pages below are primary, read 12 Oct 2026.

Two AWS caveats matter. Its regional-availability page also lists Claude Opus 5 and Sonnet 5 as In-Region in Mumbai on the bedrock-mantle endpoint, but the model cards omit it, so confirm in the console. And AWS says not to rely on the prefix alone (call GetInferenceProfile) and that prompts and outputs "may be stored in the opt-in Regions for abuse detection purposes" (Hyderabad is opt-in, so still in India). Bedrock's Mumbai prices were not captured. In short: no vendor's own API processes requests in India, and the newest models (Claude 5.5, GPT-6, Gemini 3.6 and later) have no documented India processing on any cloud read.

DeepSeek's privacy policy says personal data is collected, processed and stored "in People's Republic of China" (DeepSeek privacy policy). Indian clouds also sell managed endpoints (vendor-reported). E2E Networks TIR offers an "OpenAI-style endpoint served from Indian data centers" for DeepSeek R1/V3, Llama 4, GPT-OSS 120B, Qwen3 32B and Sarvam (E2E Networks, inference endpoints), and Yotta Shakti Cloud's shared endpoints offer OpenAI-compatible APIs for Llama 3.1/4, DeepSeek V3/R1 (INT4), Qwen3 and Mistral Large 2 from Navi Mumbai and Greater Noida (Yotta Shakti Cloud, AI endpoints). No managed API was found for Sify, Tata Communications or Jio. If your data must be processed in India, accept an in-country promise only for a named model and region that has been verified: a Bedrock in. profile (previous-generation Claude, Kimi K3, GPT-5.6 Terra and Luna), a Bedrock In-Region open-weight model in Mumbai, or an open-weight model on an Indian cloud. For the newest frontier models, India processing is not documented anywhere. Gateways and hosting choices are covered in Tool stack, and Indian-language hosting options in Indian languages.

At the maximum subsidy, an IndiaAI H100 hour costs a quarter to two-fifths of a commercial one

India's GPU prices fall into three tiers: the IndiaAI portal rate for eligible users, domestic commercial clouds, and global specialist and hyperscaler clouds. Several commercial figures come from undated vendor pages.

At the commercial rates above, eight H100s running all month cost ₹12.5–20.9 lakh a month excluding GST (₹14.7–24.6 lakh with it), or about ₹1.5–2.5 crore a year. That calculation comes from the Hardware research. At the maximum 40% subsidy, the portal's H100 SXM rate of ₹91.80 (about $0.96) is about a quarter to two-fifths of the ₹217–362 commercial on-demand range and below the $1.70–2.35 contract index (₹163–226). Even the ₹153 list rate (about $1.59) is below that index. That supports the "less than half the global rate" claim for those who qualify. Two commercial options come close: E2E's spot price of ₹70 an hour for interruptible capacity, and its 3+ month commitment rate "as low as ₹155.90" per GPU-hour, about the IndiaAI list rate. The less obvious finding is that domestic commercial on-demand H100 rates are not cheap by world standards. In dollars, ₹217–362 is $2.26–3.77, and E2E's ₹255.55 is $2.66. That is in line with or below the global on-demand median of about $3.37 but above the $1.70–2.35 contract index, so in-country hosting costs about the same as abroad on Indian clouds, and more on hyperscalers (AWS Mumbai is 1.2 times US prices). Renting versus calling an API points the same way: in the Costs analysis, a $3.99-an-hour H100 beats a hosted open model at $0.17 per million output tokens only above roughly 6,500 tokens per second of sustained output. At SME volumes the API is cheaper than the staff time needed to run a GPU, so self-hosting in India is a data-residency decision, not a saving.

The unit prices that set automation costs in India

Plivo's India page gives ₹0.38 and ₹0.25 a minute and ₹200 a month for a number. Exotel's own blog is reported (not read directly) to give about ₹0.80–1.00 a minute outbound and ₹0.30–0.50 inbound on Indian numbers (Exotel blog); Twilio India and Exotel rate cards were not read, and a reseller's ₹4–6 a minute for Exotel AI agents conflicts and is not Exotel's price. On rules, TRAI's Direction of 4 May 2024 moved promotional, transactional and service voice calls onto DLT-based 140-series numbering and set a plan for 160-series numbers for service and transactional calls (TRAI Direction, 4 May 2024) (primary). Comments hosted on TRAI's site (April 2026) show that the draft TCCCPR Third Amendment Regulations 2026 define A2P calls, add pre-declaration and termination charges for them, lower the complaint threshold for suspending a sender and add operator "AI-flagging" (TRAI, TCCCPR comments); a fully automated voice agent appears to fall within A2P as drafted (inference). The final rules were reported as notified in the Gazette on 18 Sep 2026 (CareerIndia; Policy Edge) (secondary); the official text was not read, and no explicit AI-voice rule is confirmed. Any voice-agent quote therefore needs a telephony line item checked with the carrier; Indian languages builds a rupee voice-agent minute leg by leg. WhatsApp costing should be confirmed against Meta's own rupee rate-card files, because the two vendor rate tables differ by a factor of 1.15 and Meta's page, as read on 11 Oct 2026, did not confirm the service-message charge. Voice and workflow platforms are compared in Tool stack and Frameworks.

Adoption is huge in total and only average per person

Consumers: 100 million weekly ChatGPT users and free premium tiers

Sam Altman said in February 2026 that India had about 100 million weekly active ChatGPT users, the second-largest base after the US, and more student users than any other country (TechCrunch, 15 Feb 2026; TechRepublic). OpenAI's Christopher Lehane said at Davos that Indian usage had grown 2.5 times in a year (Business Today, 19 Jan 2026). Pricing has been aggressive. ChatGPT Go launched in India on 18 Aug 2025 at under $5 a month (TechCrunch, 18 Aug 2025). From about 4 Nov 2025 it was free for a year, with billing at ₹399 a month afterwards unless cancelled. OpenAI says Go helped double its paid subscriber base in India (Stocktwits) (secondary). Google has run an 18-month free AI Pro offer for eligible Jio users and a free-year offer for undergraduates that the snippets say was valid to 15 Sep 2025 (Outlook Business) (snippets). The upshot for employers (our inference) is that a large share of SME staff already use capable AI on personal accounts, outside any company data policy. OpenAI and Gemini cover the vendor details.

Work and enterprise: software-heavy and concentrated in four states

Anthropic's India Country Brief (16 Feb 2026; data from 13 to 20 Nov 2025) is the most detailed primary usage source for India. India accounted for 5.8% of global Claude.ai consumer use, second only to the US, but ranked 101st of 116 countries per person. Four states produced more than half of Indian use: Maharashtra 15.5%, Tamil Nadu 13.2%, Karnataka 12.7% and Delhi 10.5%. India ranked first in the world for the share of software-related tasks (45.2%), ahead of Vietnam (42.1%) and Egypt (39.2%). Work-related use was 51.3% against 46% globally (Anthropic, India Country Brief). The data covers Claude.ai Free, Pro and Max users located by IP address. It is one vendor's consumer product, not the national workforce. The September 2025 edition had given India 7.2% of global use and a per-capita usage index of 0.27, against 3.62 for the US, and found coding was over half of Indian usage against about a third globally (Anthropic Economic Index, Sept 2025). Anthropic's June 2026 data release (April and May 2026; Claude.ai chat and Cowork on Free, Pro and Max plans) gives India a per-capita index of 0.31 and 0.30, with 6.92% and 7.12% of all usage, against 4.25 and 3.87 for the US (Anthropic Economic Index data) (primary). The brief's 5.8% and 101st-of-116 figures remain valid for their earlier release; no new rank is derived from the June file.

Enterprise deployments are mostly announced through vendors. Anthropic opened its Bengaluru office on 16 Feb 2026, led by managing director Irina Ghose, its second Asian office after Tokyo. It named partners including Air India (Claude Code), CRED (a vendor-reported "2x faster feature delivery"), Cognizant ("deploying Claude to 350,000 employees"), Razorpay, Pratham and Adalat AI, and a Ministry of Statistics (MoSPI) MCP server (Anthropic, Bengaluru office). TechCrunch's report on OpenAI's February 2026 India deal names eight Indian enterprise partners, among them PhonePe, Pine Labs and MakeMyTrip, and says OpenAI plans Mumbai and Bengaluru offices later in 2026 alongside its New Delhi presence (TechCrunch, 18 Feb 2026). Survey data says workplace use is high. Stanford's AI Index 2026 reports that more than 80% of employees in India use AI at work semi-regularly or regularly, against 58% globally, and that India had the sharpest rise in AI nervousness (+14 points) of any country surveyed (Stanford HAI, Public Opinion). EY's 2025 figure of 62% of Indian employees using generative AI regularly rests on a sample reported as 15,000 employees across 29 countries, so it is not an India-only base (Storyboard18 on EY). Measured company outcomes are rare. Meesho claims its generative-AI voice bot cut some customer-call costs by 75%, with no baseline given (TechCrunch, Nov 2024) (company claim). Claude and Use cases carry more case evidence.

MSMEs: high claimed use, low clarity

The best MSME figure comes from a retail-heavy sample. The PayNearby MSME Digital Index 2026 (published 23 Jun 2026; 10,000 retailers and MSMEs) found that 71% of retailers use AI tools in daily operations but 26% are unsure which AI features would help. It also found that UPI and Aadhaar-enabled banking make up 61% of transactions at semi-urban and rural MSME retail stores (Newswav on PayNearby, June 2026). A Vi Business study reports that 57% of MSMEs see AI as core to growth and 25% have integrated it, but the page could not be read (CXOToday) (unverified). No named survey gives an Indian SME AI budget. The gap between "using AI tools" and "unsure which features help" is the clearest finding for small businesses. On our reading, many have AI in their apps without having redesigned any workflow around it.

Government deployments: many platforms, few measured outcomes

Government AI is visible in platforms rather than results: Bhashini, the MoSPI MCP server, AWS Outposts in National Informatics Centre data centres under MeghRaj 2.0, and Adalat AI's court work. NPCI's new UPI limits of 1 Aug 2025 are a rule change, not a measured result (Wikipedia, UPI) (Wikipedia only). No outcome figures were found for Kisan e-Mitra, court tools, tax analytics or health deployments; PIB and ministry annual reports were blocked.

India in global indices: what each one measures

Sources: Microsoft AI Diffusion Report 2026 Q2; Microsoft AI Diffusion Report 2026 Q1; Anthropic India brief; Anthropic, Sept 2025; Anthropic Economic Index data, June 2026; IMF DataMapper, AI_PI; Stanford HAI; TOP500.

The indices seem to disagree because they measure different things, and taken together they make sense. On breadth (Microsoft), India tracks the world average almost exactly and is not pulling ahead (+0.9 points in Q2 2026, against +1.0 worldwide). On intensity per person (Anthropic), India is low because a huge population dilutes a large absolute user base. On workplace self-reports (Stanford), India is near the top, which probably reflects a sample of connected, urban workers (our inference; the survey's sampling was not read). On preparedness (IMF), India scores below most large emerging-market peers. Vaishnaw cited a Stanford ranking that placed India third on AI penetration, preparedness and talent; that ranking was not checked against Stanford's data, and the Microsoft and IMF figures do not support it. In short, India's AI users are a large, urban, work-oriented and software-heavy minority. The rest of the economy, including most SMEs outside Maharashtra, Tamil Nadu, Karnataka and Delhi, has barely started. Other countries' scores are in Countries overview, Asia-Pacific, Americas and Europe and Middle East and Africa.

IT services and GCCs: headcount fell in FY26 as AI approached or passed a tenth of revenue

Large IT firms report AI as a share of revenue while cutting staff

The five largest Indian IT services firms (TCS, Infosys, Wipro, HCLTech and Tech Mahindra) shed a net 7,389 employees in FY26, after adding 12,718 the year before. It was the first net decline in about two decades (Forkast; IANS, 25 Apr 2026) (secondary). In the same period the firms began reporting AI as a run-rate revenue line.

Sources: Outlook Business on TCS Q2 FY27; YourStory on TCS Q2 FY27; The Tribune on TCS Q1 FY27; DT Next; The Quint; PeopleMatters; Entrepreneur India; TechMarketView; Let's Data Science; DQ India on HCLTech; Wipro Q1 FY27 release; Whalesbook on LTIMindtree; Accenture 8-K exhibit; Motley Fool, Accenture Q4 FY26 call transcript; Business Standard on Infosys Topaz.

AI is not the only cause of the decline. TCS's CEO reportedly said the cuts were not solely driven by AI, and that redeployment had failed for roles such as manual testing. Its HR chief said it had doubled its intake of fresh graduates with advanced AI skills (DT Next; The Quint). Some of their clients reportedly want 20–30% price cuts on new contracts (Let's Data Science) (secondary). Infosys cut its guidance in the same quarter that it reported AI growth, which suggests AI revenue is not yet offsetting weakness in the rest of the business. Meanwhile the large firms are becoming channels for US model vendors. Tata Group will roll out ChatGPT Enterprise starting with "hundreds of thousands" of TCS employees. TCS will use Codex, and it is the first organisation outside the US in OpenAI's certification programme (TechCrunch, 18 Feb 2026). Anthropic's India announcement says Cognizant is deploying Claude to 350,000 employees. For scale, NASSCOM estimates technology-industry revenue of $315 billion in FY2026, up 6.1%, and IT-BPM is about 7.5% of GDP (Wikipedia, Information technology in India) (Wikipedia only; the article carries a machine-generated-text warning). More detail is in Companies.

GCCs have become transformation hubs

The NASSCOM–Zinnov landscape report counts more than 1,700 GCCs in FY2024, with estimated revenue of $64.6 billion (about ₹6.2 lakh crore) and more than 1.9 million employees. It says more than half of GCCs have become portfolio and transformation hubs, and that global roles grew at a 40% compound annual rate over five years. The same page also says "over 2,975 centres", which contradicts the 1,700 figure and was not resolved (NASSCOM). Wikipedia cites a projection of up to 2,200 GCCs by 2030 and lists Bengaluru, Pune, Hyderabad, Noida, Gurgaon, Chennai and Navi Mumbai as the main hubs (Wikipedia, Global capability centre). No figure for GCC AI headcount was found. For businesses hiring AI staff in these hubs, GCCs are, on our reading, rivals for the same talent. The sources do not size the effect.

Talent: a 6–6.5 lakh AI workforce, paid about a tenth of US levels

India's AI talent pool is estimated at 6 to 6.5 lakh people, projected to pass 1.25 million by 2027, according to a NASSCOM–Deloitte paper as reported. Deloitte South Asia says skills are growing about 13% a year against demand growth of 25–35% (Entrepreneur India). A Parliamentary report covered in March 2026 repeats the figures and says demand is outpacing supply (Storyboard18). An IIM Ahmedabad and BCG study says India's top 500 firms will need at least 25,000 to 30,000 advanced AI/ML practitioners within three to five years. It puts India's share of the world's AI professionals at 4.5% (Careers360). The developer pipeline is very large. India added more than 5 million developers to GitHub in 2025, over 14% of new accounts, and GitHub expects India to account for one in three new developers by 2030 (GitHub Octoverse 2025). NPTEL lists 3,200+ courses and 3.79 crore+ enrolments (NPTEL).

By our arithmetic, the mid-market Indian AI/ML base of about ₹10.2 lakh is roughly an eleventh to a thirteenth of US median wages for data scientists and software developers, though the measures differ (self-reported base pay against official medians). A like-for-like survey gives a similar ratio: in the 2025 Stack Overflow survey, AI/ML engineers reported $17,436 in India against $189,500 in the US (Stack Overflow Developer Survey 2025). Two features matter for anyone budgeting AI staff or contracts in India. First, projected 15–20% annual pay growth and the demand-supply gap will push up the cost of multi-year budgets and fixed-price contracts set on today's salaries. Second, on our reading, the IT-services headcount decline releases experienced engineers into the market, including from roles such as manual testing, where TCS said redeployment had failed. No salary data was found for forward-deployed engineers, LLMOps engineers, AI evaluators or AI governance leads. Hiring and skills guidance is in Playbook, and terms are defined in the Glossary.

What this means for your business

Your deadlines: DPDP on 13 May 2027, WhatsApp and labels sooner

If your business collects customer data in India through WhatsApp, web forms, CRMs or e-commerce, DPDP is the obligation with statutory dates and penalty maximums of up to ₹250 crore. From 13 May 2027 you must give notice and take consent, keep reasonable security safeguards, tell affected individuals and the Board about breaches without delay and send the Board a fuller report within 72 hours, keep processing logs for a year, erase data on the Rule 8 and Third Schedule timetables, and honour data-principal rights. Seven months is short without a data inventory. Start with a data map, then consent and notice flows in your existing channels, automated retention and erasure, audit logging, and a breach runbook with both clocks: CERT-In's 6-hour incident report (binding since 2022, covering attacks on AI/ML systems) and the DPDP 72-hour report. The Act puts no direct duties on processors, so your processing contracts must pass them to any vendor that handles your customer data. The 13 Nov 2026 tranche covers consent managers only.

Two obligations come sooner. If Meta bills you with India as the "Sold-To" country, move every WhatsApp Business account to INR billing by 31 Dec 2026, or Meta stops delivering its messages from 1 Jan 2027. And label synthetic audio and video, as the IT Rules have required since 20 Feb 2026.

Automations that fit Indian SMEs

Most Indian SMEs already have AI tools, bundled in apps and free consumer plans; what they usually lack is workflow redesign, Indian-language channels and compliance, so budget for those first. Globally, AI projects without workflow redesign or a human escalation path fail more often (Use cases).

Subsidised compute: check eligibility first

At SME volumes an API costs less than the staff time needed to run a GPU, so self-hosting in India is a data-residency decision, not a saving. IndiaAI portal GPUs go only to named categories, such as DPIIT-recognised start-ups and MSMEs with ₹50 lakh revenue or ₹100 lakh funding; large enterprises are not one. Budget at the ₹153 list rate per H100 SXM hour, and at ₹91.80 only if the maximum 40% subsidy is approved. Requests go in on the 1st–25th of a month, and those under 5,000 GPU-hours are auto-approved.

Indian-language choices

No independent 2026 benchmark compares current flagships with Indian models, so decide on your own test set and include romanised, code-mixed messages: in one preprint Claude Opus 4.6, Sarvam-30B and Gemini 3.5 Flash each lost about 7–8 points of accuracy at 75% code-mixing. Indian scripts also take 2.1 to 8.0 times the tokens of English. Domestic open-weight options include Sarvam's 30B and 105B and Gnani's 30B (all Apache-2.0), and Sarvam's rupee prices remove currency risk on the speech leg. Indian languages compares models, speech and OCR, with costs.

Working with Indian teams from abroad

If your automation is built or run by an Indian team (your own, a GCC or a vendor), the obligations follow the work, not the location. US governance is often thin: 52% of US employees use AI at work (Gallup), 37% of firms with 250+ employees use AI in operations (US Census), 25% of organisations have no AI policy (ISACA), and 76% of respondents in a Salesforce/YouGov survey said AI tools lack company context (Salesforce).

Pay and prices. Indian AI/ML base pay of about ₹10.2 lakh is roughly an eleventh to a thirteenth of US medians and is projected to grow 15–20% a year, so index multi-year budgets and contracts. As evidence of what implementation partners charge, US boutiques that publish prices use three steps: Main & Machine lists a fixed-fee readiness audit of $3,500–8,500 over 2–4 weeks (about ₹3.4–8.2 lakh), a build sprint of $18,000–60,000 over 4–12 weeks (₹17.3–57.6 lakh) and managed service from $1,500 a month (₹1.44 lakh) (Main & Machine pricing) (vendor page; another of its pages shows a lower sprint range). For per-resolution pricing, Fin charges $0.99 per outcome (about ₹95) with a 50-outcome monthly minimum, and Zendesk's price varies by regional page (£1.30 committed on the India and UK pages; US$2.00, about ₹192, pay-as-you-go) (Fin pricing; Zendesk pricing). Indian providers mostly publish no prices: Antino Labs uses "pay for outcomes" wording (Antino), and Infyom prices n8n builds on complexity (Infyom). No Indian SME rate card for AI automation was found, so test any quote through a paid, fixed-scope pilot. Costs has the full analysis.

Evidence, not location. Where significant decisions are automated, the US state laws above apply; otherwise US companies rely on security questionnaires. Ask your Indian team or vendor to map its controls to the OWASP Top 10 for LLM Applications 2025, especially LLM06 "Excessive Agency" (narrow tools, minimal permissions, human approval for high-impact actions, authorisation enforced outside the model) (OWASP LLM06:2025), the OWASP Top 10 for Agentic Applications (9 Dec 2025) and the voluntary NIST AI RMF Generative AI Profile (NIST GenAI Profile); larger buyers are expected to ask for ISO/IEC 42001 (current text not verified). Vendor indemnities are narrow: Anthropic's commercial terms indemnify paid use, including training data and authorised outputs, but exclude customer inputs, modifications and combinations with non-Anthropic technology, and cap liability at the previous 12 months' fees (Anthropic Commercial Terms). Your contract with an implementation partner must therefore cover its wrapper code and your data. US-only endpoints cost about 10% more, and DeepSeek's official API sends data through China (see the residency section and DeepSeek).

Model access is now a regulatory risk for any business outside the US. On 12 Jun 2026 a US export-control directive (attributed to the Commerce Department in press reports) suspended all foreign-national access to Fable 5 and Mythos 5; Anthropic switched both off for every customer until 1 July, about 19 days (Anthropic, 12 Jun 2026; Al Jazeera, 1 Jul 2026). The Washington Post reported that the US government would vet users of OpenAI's GPT-5.6 Sol (Washington Post, 26 Jun 2026) (headline only; see Timeline). Require that every system can switch models without a rewrite, put model-availability clauses in contracts, and keep a tested fallback on open weights hosted in India or on a Bedrock India-profile model.

Risks

Licence details: Tool stack and Other models.

A sequence to May 2027

The undated items in the calendar's last row (RBI, SEBI, MeitY and the IT Rules label draft) need watching throughout.

Questions to ask any vendor or implementation partner

• Which model, endpoint and region will process the data, and is India processing documented for that model?
• What is the tested fallback if a model is withdrawn?
• Which DPDP duties does the processing contract pass to the vendor, and do its runbooks meet the CERT-In 6-hour and DPDP 72-hour clocks?
• Are an AI-use policy, a model and data register, an AIBOM and an incident runbook delivered as standard?
• Is synthetic audio and video labelled, and do automated calls follow TRAI's A2P rules?
• Are WhatsApp and telephony fees passed through at cost?
• Which licences cover anything hosted on your behalf, and what does the model vendor's indemnity exclude?
• How was Indian-language quality tested, and on whose data?

Two less obvious conclusions follow. First, the rules that matter most for AI in India are not AI rules: a data-protection statute, a media-labelling rule and a messaging platform's billing change will shape projects more over the next eight months than the AI Governance Guidelines. Second, India's cost advantage is in people, not compute: subsidised GPU-hours go to named categories, commercial Indian GPU-hours cost about what they do abroad and in-country processing covers only some previous-generation models, while Indian AI salaries are about a tenth of US levels. Plan for documented compliance and model independence, not cheap infrastructure. Future outlook looks beyond 2027.

Open questions and unverified items

Sources

Dates in titles are publication dates where known. Pages were accessed on 11 October 2026, and key facts were re-checked on 12 October 2026; titles note pages that could not be read in full.

1. X-Rates: USD to INR monthly averages, 2026
2. Business Today, 21 Jan 2026: Vaishnaw at Davos
3. Wikipedia: Artificial intelligence in India
4. Free Press Journal, Feb 2026: three sovereign models unveiled at the summit
5. TelecomTalk, 25 Mar 2026: MeitY statement on 38,000 GPUs
6. Communications Today: IndiaAI compute and GPU-hours sanctioned
7. Wikipedia: AI Impact Summit
8. Storyboard18, 20 Feb 2026: Jayant Chaudhary on one lakh GPUs at Rs 65 per hour
9. IndiaAI compute portal: price calculator ("Last Updated 08-10-2026")
10. IndiaAI compute portal: Ready Reckoner (PDF, rates as of 13 Mar 2026)
11. IndiaAI compute portal: eligibility criteria
12. IndiaAI compute portal: empanelled GPUs (PDF, undated)
13. Business Standard, c. 29 Jun 2026: subsidised GPU capability gap (not readable)
14. DEV Community: GPU cloud pricing in India 2026 (superseded by the IndiaAI portal)
15. Tensordata: GPU cloud pricing in India 2026 (superseded by the IndiaAI portal)
16. Outlook Business: Vaishnaw on AI compute at less than half the global rate
17. Business Today, 16 Mar 2026: The Sarvam moment
18. Wikipedia: Sarvam AI
19. The Week / PTI, 30 May 2025: IndiaAI foundation-model selections
20. Inc42: IndiaAI Mission picks Soket AI, Gnani.ai and Gan.ai to develop indigenous models
21. Indian Infrastructure, 27 Jul 2026: 20 indigenous AI model proposals identified
22. Entrepreneur India, 2 Apr 2026: Abhishek Singh moves from IndiaAI Mission to lead NTA
23. Storyboard18: New NTA Director General Abhishek Singh on the IndiaAI Mission
24. Elets eGov, June 2026: Saurabh Vijay given additional charge of IndiaAI Mission
25. Wikipedia: Ministry of Electronics and Information Technology
26. Wikipedia: Bhashini
27. Fortune, 23 Feb 2026: AI Impact Summit closes with the New Delhi Declaration
28. IANS, 21 Feb 2026: 88 countries and organisations endorse New Delhi Declaration on AI
29. ORF: strategic outcomes of the India AI Impact Summit 2026
30. Business Today, 19 Feb 2026: Vaishnaw's vision at the summit
31. Indian Infrastructure, 7 Nov 2025: MeitY unveils India AI Governance Guidelines
32. S.S. Rana & Co.: MeitY unveils India AI Governance Guidelines
33. Gazette of India: G.S.R. 843(E), DPDP Act commencement notification, 13 Nov 2025
34. MeitY: Digital Personal Data Protection Rules 2025, G.S.R. 846(E)
35. Gazette of India: G.S.R. 846(E), DPDP Rules 2025, 13 Nov 2025
36. MeitY: copy of DPDP commencement notification G.S.R. 843(E)
37. Wikipedia: Digital Personal Data Protection Act, 2023
38. CyberPeace: The DPDP compliance clock is running
39. Hogan Lovells: India's Digital Personal Data Protection Act 2023 brought into force
40. Gazette of India: Digital Personal Data Protection Act 2023 (Act No. 22 of 2023)
41. Business Today, 10 Feb 2026: Govt tightens rules on AI-generated content
42. Hogan Lovells, 13 Feb 2026: India introduces mandatory labelling for AI and 3-hour takedown
43. iPleaders: IT Rules 2026 amendment explained
44. Gala Law: India's 2026 IT Rules amendment regulating AI-generated content
45. S.S. Rana & Co.: India tightens oversight on AI-generated content under IT Rules
46. Internet Freedom Foundation: IT Intermediary Amendment Rules 2026 contradict their purpose
47. exchange4media: Grok AI misuse, X gets MeitY notice
48. The Logical Indian: MeitY issues notice to X over Grok deepfakes
49. Business Today, 2 Jan 2026: Centre issues notice to X over Grok content
50. Storyboard18: X takes down Grok content, deletes 600 accounts
51. Wikipedia: 2026 in artificial intelligence
52. Storyboard18, 14 Aug 2025: RBI unveils FREE-AI framework
53. Outlook Money: RBI panel proposes ethical AI framework for financial sector
54. Reserve Bank of India: press release on the FREE-AI committee report
55. Bar and Bench, 18 Sep 2025: RBI's FREE-AI framework, navigating the legal landscape
56. Reserve Bank of India: FREE-AI committee report (PDF; timed out)
57. Lakshmikumaran & Sridharan: SEBI's consultation paper and the winds of AI governance
58. Taxmann: SEBI proposes AI/ML governance framework for securities markets
59. Medianama, Jul 2025: CERT-In AI bill of materials guidelines
60. AZB & Partners, 29 Jul 2025: CERT-In's new BOM guidelines
61. Khaitan & Co: CERT-In's expanded BOM guidelines
62. Mondaq: CERT-In's new BOM guidelines
63. CERT-In: SBOM, QBOM & CBOM, AIBOM and HBOM guidelines v2.0 (PDF, 9 Jul 2025)
64. CERT-In: section 70B(6) Directions, 28 Apr 2022 (PDF)
65. CERT-In: Blueprint on AI-assisted vulnerability exploitation, v1.0 (PDF, 25 May 2026)
66. CERT-In: guidelines listing, CISG-2026-03 (10 Jun 2026)
67. OECD.AI incident monitor, 27 Apr 2026: CERT-In high-severity advisory on AI-enabled attacks
68. TeamLease RegTech: CERT-In guidelines for OEMs and technology providers (10 Jun 2026)
69. ICMR: guidelines page
70. Colorado General Assembly: SB26-189, Automated Decision-Making Technology
71. Colorado Attorney General: ADMT and Chatbot Safety rulemaking
72. Wikipedia: Texas Responsible AI Governance Act (TRAIGA)
73. National Law Review: California privacy regulations on ADMT, cybersecurity audits and risk assessments
74. California Privacy Protection Agency: approved ADMT, risk-assessment and audit text (PDF)
75. CourtListener: X.AI LLC v. Weiser, D. Colo. docket
76. Axios, 11 Dec 2025: Executive order on state AI laws
77. European Commission: AI Act regulatory framework page
78. EU AI Act, Article 50 (artificialintelligenceact.eu)
79. European Commission, AI Act Service Desk: implementation timeline
80. NatLawReview (K&L Gates), 31 Jul 2026: EU AI Omnibus ("Digital Omnibus on AI") enters into force
81. Meta: WhatsApp Business Platform pricing
82. OpenAI developer docs: Agent Builder (deprecation notice)
83. Business Today, 4 Nov 2025: ChatGPT Go free for one year in India
84. ChatMaxima: WhatsApp API pricing in India
85. Gallabox docs: WhatsApp pricing overview (INR tab)
86. Sarvam AI: blog
87. Sarvam AI, 6 Mar 2026: Sarvam 30B and 105B
88. Hugging Face API: Sarvam AI models
89. MediaNama, June 2026: Sarvam raises $234 million
90. BharatGen: home page
91. Hugging Face API: BharatGen models
92. Gnani.ai: home page
93. Hugging Face API: Gnani search
94. Hugging Face: gnani-evon-v3.3-30B-A3B model card
95. Soket AI: home page
96. Hugging Face API: Soket search
97. CoRover: home page
98. Hugging Face API: CoRover search
99. Hugging Face API: Krutrim AI Labs models
100. Hacker News listing: Krutrim stories
101. Anthropic, 16 Feb 2026: Bengaluru office and partnerships across India
102. Google Help: languages supported in Gemini Apps
103. Claude platform docs: multilingual support
104. Claude Help Center: using Claude in your preferred language
105. Entrepreneur India: NVIDIA leads $74 million investment in Sarvam AI
106. ARC: Sarvam building India's sovereign AI stack
107. Sarvam AI: API pricing
108. arXiv 2606.30790 (preprint): Indi-RomCoM, romanised code-mixed instruction following
109. arXiv 2607.24276 (preprint): tokenizer cost of Indian languages
110. Wikipedia: Opposition to AI data centers
111. IANS, 2 Sep 2026: data-centre capacity to reach 6 GW by 2029 (JLL)
112. Energetica India: India data-centre capacity to reach 12 GW by 2030 (Wood Mackenzie)
113. DEV Community: Reconciling India's data-centre capacity and dollar numbers
114. tele.net.in: India DC capacity to hit 9 GW by 2032
115. Outlook Business: data-centre capacity likely to reach 3 GW by 2030
116. IANS, May 2026: data-centre capacity up to 8 GW by 2030
117. TOP500: June 2026 list
118. Times of Oman: Google to invest $15 billion in Visakhapatnam AI hub
119. Investment Monitor: Google to establish $15bn AI hub in Visakhapatnam
120. Verdict: Google AI hub in Visakhapatnam
121. Wikipedia: Visakhapatnam
122. Economic Times, Apr 2026: Vaishnaw after the Google Cloud India AI Hub ground-breaking
123. TechCrunch, 18 Feb 2026: OpenAI taps Tata for 100 MW in India
124. Wikipedia: Amazon Web Services
125. Yotta: home page
126. Anthropic: regional compliance page (read 11 Oct 2026; not re-readable 12 Oct)
127. OpenAI developer docs: Data controls in the OpenAI platform (data residency regions)
128. Claude platform docs: data residency
129. AWS Bedrock User Guide: model support by Region (regional availability)
130. AWS Bedrock User Guide: Claude Sonnet 5 model card
131. Google Cloud, Gemini Enterprise Agent Platform: data residency
132. Microsoft Learn: Foundry Models sold directly by Azure, region availability
133. Oracle Cloud: Generative AI model endpoint regions
134. DeepSeek: privacy policy
135. E2E Networks: inference endpoints
136. Yotta Shakti Cloud: AI endpoints
137. OpenAI developer docs: API pricing
138. ECorpIT: India GPU cloud rental pricing, H100 and B200, 2026
139. E2E Networks: NVIDIA H100 cloud GPU pricing
140. E2E Networks blog: NVIDIA H200 price in India (19 Aug 2026)
141. E2E Networks: GPU cloud pricing
142. IntuitionLabs: H100 rental prices compared
143. Lambda: pricing
144. Crypto Briefing: NVIDIA H100 GPU rental costs surge
145. ainauten, 13 Aug 2026: NVIDIA DGX Spark now costs $4,699
146. Tech Insider: NVIDIA RTX PRO 6000 Blackwell price 2026 (secondary)
147. GPU Poet: RTX PRO 6000 Blackwell price tracker, July 2026
148. AWS Price List: Amazon EC2, ap-south-1 (9 Oct 2026)
149. Azure Retail Prices API
150. Plivo: India voice pricing
151. Bolna: pricing
152. Exotel blog: Twilio to Exotel migration for voice AI (not read directly)
153. TRAI: Direction on DLT-based voice calls, 4 May 2024 (PDF)
154. TRAI: comments on the draft TCCCPR Third Amendment (PDF, Apr 2026)
155. CareerIndia: TRAI anti-spam rules 2026 for AI voice agents
156. Policy Edge: TRAI finalises new spam rules
157. Retell AI: pricing
158. Analytics Vidhya, Nov 2025: Jio users get free Google AI Pro plan
159. TechStory: Anthropic launches rupee billing for Claude in India
160. TechCrunch, 15 Feb 2026: India has 100M weekly active ChatGPT users, Sam Altman says
161. TechRepublic: India has 100 million weekly ChatGPT users
162. Business Today, 19 Jan 2026: OpenAI bets big on India, says Christopher Lehane
163. TechCrunch, 18 Aug 2025: OpenAI launches a sub-$5 ChatGPT plan in India
164. Stocktwits: OpenAI to offer ChatGPT Go free in India for a year
165. Outlook Business: Google offers free one-year Google AI Pro to Indian college students
166. Anthropic, 16 Feb 2026: India Country Brief, Anthropic Economic Index
167. Anthropic Economic Index, September 2025 report
168. Anthropic Economic Index dataset, June 2026 release
169. Stanford HAI: AI Index 2026, Public Opinion chapter
170. Storyboard18, 23 Dec 2025: 62% of Indian employees use GenAI regularly (EY)
171. TechCrunch, 26 Nov 2024: AI helps India's Meesho cut customer call costs by 75%
172. Newswav, Jun 2026: PayNearby MSME Digital Index 2026
173. CXOToday: Vi Business MSME AI study (not readable)
174. Wikipedia: Unified Payments Interface
175. Microsoft AI Diffusion Report 2026 Q2 (PDF, Sept 2026)
176. Microsoft AI Diffusion Report 2026 Q1 (PDF, May 2026)
177. IMF DataMapper API: AI Preparedness Index
178. Forkast: India's IT giants are cutting jobs for the first time in 20 years
179. IANS, 25 Apr 2026: Top IT firms' hiring turns negative in FY26
180. Outlook Business, 8 Oct 2026: TCS Q2 FY27 results
181. YourStory, Oct 2026: TCS Q2 FY27 results
182. The Tribune: TCS Q1 FY27 results, AI revenue up 13.6% QoQ
183. DT Next: TCS likely to lay off 12,000 employees in FY26
184. The Quint: TCS workforce shrinks in Q3 FY26
185. PeopleMatters: TCS headcount falls by 11,151 in Q3 FY26
186. Entrepreneur India: From TCS to TechM, headcount declines
187. TechMarketView, 24 Jul 2026: Infosys trims guidance as AI revenue climbs (headline only; body not readable)
188. Let's Data Science: Indian IT giants face AI-driven FY26 reset
189. DQ India: HCLTech Q1 FY27 management on AI growth
190. Wipro: Q1 FY27 results press release (PDF)
191. Whalesbook: LTIMindtree Q1 FY27 results
192. Accenture Q4 FY26 earnings, Form 8-K exhibit (SEC)
193. Motley Fool: Accenture Q4 FY26 earnings call transcript, 1 Oct 2026
194. Business Standard, 29 May 2025: Infosys Topaz agents (headline only)
195. Wikipedia: Information technology in India
196. NASSCOM: India GCC Landscape Report, the 5-year journey
197. Wikipedia: Global capability centre
198. Entrepreneur India: Indian AI talent pool to touch 1.25 million by 2027
199. Storyboard18: India AI talent gap, Parliamentary report
200. Careers360: IIM Ahmedabad study on AI/ML talent demand
201. GitHub Octoverse 2025
202. NPTEL: home page
203. PayScale: Machine Learning Engineer salary in India
204. PayScale: Data Scientist salary in India
205. Stack Overflow Developer Survey 2025: Work (pay by role and country)
206. Storyboard18, 7 Jan 2026: AI and ML engineer pay in India
207. Tredence: AI engineer salary in 2026
208. US Bureau of Labor Statistics: Data scientists
209. US Bureau of Labor Statistics: Software developers
210. Gallup: AI in the workplace indicator
211. US Census Bureau, May 2026: AI use by businesses
212. ISACA 2026: AI safeguards and ROI remain elusive
213. Salesforce: AI tools lack job context
214. Main & Machine: pricing
215. Fin: pricing
216. Zendesk: pricing
217. Antino: artificial intelligence services
218. Infyom: n8n workflow automation services
219. OWASP GenAI: LLM06:2025 Excessive Agency
220. NIST: AI RMF Generative AI Profile
221. Anthropic: Commercial Terms of Service
222. Anthropic, 12 Jun 2026: statement on Fable 5 and Mythos 5 access
223. Al Jazeera, 1 Jul 2026: US lifts restrictions on Fable and Mythos models
224. Washington Post, 26 Jun 2026: US to vet users of OpenAI's latest model (headline only)