Reliance and Adani at one hundred billion US dollar quanta, Microsoft and Google at fifteen to seventeen billion, and the IndiaAI Mission at one and a quarter billion: 2025 reframed AI infrastructure as a flagship Indian sectoral opportunity at quanta few sectors have ever attracted. The country head evaluating entry encounters not one regulatory architecture but six, and an entrant typology that prices capital differently from a standard MNC. What architectures does the entrant actually navigate, and which is the binding constraint?
The visible architecture is the IndiaAI Mission. Approved by Cabinet in March 2024 with an outlay of approximately one and a quarter billion US dollars over five years, the Mission has empanelled thirty-eight thousand GPUs across ten domestic providers including Yotta, Jio, Tata Communications, CtrlS, NxtGen, and E2E Networks. Subsidised access at sixty-five rupees per GPU-hour runs through the IndiaAI Compute portal. A third tranche announced in February 2026 added twenty thousand sovereign GPUs, with a one-hundred-thousand-GPU target by late 2026.
The build cadence is faster than the Mission's own architecture. The NVIDIA-Yotta twenty-thousand-GPU Blackwell cluster activates in August 2026. Larsen and Toubro is building gigawatt-scale NVIDIA AI factory infrastructure across Chennai and Mumbai. Sarvam AI received approval in April 2025 to build India's Sovereign LLM Ecosystem.
The headline announcements sit at a different scale entirely. Reliance Industries has announced approximately one hundred and ten billion US dollars in AI and data-centre commitments since late 2025, anchored by the Jamnagar gigawatt campus. Adani Group has announced one hundred billion. Microsoft and Google have committed seventeen and a half and fifteen billion respectively, with Google's commitment substantially flowing through the Adani partnership for the Visakhapatnam campus. Tata Group through TCS has committed six and a half billion. Yotta two billion.
What an entrant building AI infrastructure in India actually navigates is six regulatory architectures operating in parallel, none of which was designed for AI infrastructure as an integrated category.
The IndiaAI Mission is one of six architectures the entrant navigates: the Mission, the Telecommunications Act, the DPDP Act, state data-centre policies, the power and grid architecture for hyperscale loads the Indian grid was not built to carry, and US export controls on advanced GPUs.
The Telecommunications Act 2023 brings data centres into its scope where the centre interconnects with public networks or carries certain categories of traffic. The Act's implementing rules and the Department of Telecommunications' authorisation framework are still being constructed. The Digital Personal Data Protection Act 2023 introduces data localisation requirements for specific categories of personal data; the implementing rules are at draft stage as of April 2026. The operative scope, which categories of data are storable in cloud infrastructure outside India, which require Indian residence, which require Indian-controlled processing, is the architecture the rules will define.
The rules under the DPDP Act are not downstream compliance; they are upstream of the build, because the data classes a centre's customers will store determine what compute the centre must be designed to host.
The state-level data-centre policy stack adds the fiscal layer. Maharashtra, Tamil Nadu, Telangana, Uttar Pradesh, and West Bengal have each issued state-level policies with differentiated incentive packages: capital subsidy, electricity duty exemption, stamp duty refund, water charge concessions, dedicated industrial estate allocation. The state's selection is now as material a regulatory decision as central scheme eligibility.
The power architecture is the constraint the announcements most consistently understate. Andhra Pradesh has committed approximately five gigawatts of data-centre capacity in Visakhapatnam alone to Google, Meta, Reliance Industries, Tata Consultancy Services, and others; India's existing data-centre capacity totals approximately 1.3 gigawatts spread across Mumbai, Hyderabad, Chennai, and Bengaluru, built over three decades. The Visakhapatnam commitments alone are roughly three times India's installed load, in a city that had near-zero data-centre infrastructure six months prior. Google's one-gigawatt Visakhapatnam campus, anchored by a fifteen-billion-dollar investment over five years, has been cleared by the Andhra Pradesh state cabinet for a dedicated private power-distribution licence, with the grant proceeding through the state regulator; the first hyperscale instance of a route a data-centre developer in Uttar Pradesh first took in 2023. The investors have committed to build the captive generation themselves; the surrounding agricultural and household consumers in the same regions are already on rationed grid supply, and the precedent of state-level commitments to large captive generation has been uneven on delivery.
The export-control overlay is the architecture entrants underestimate next.
US Bureau of Industry and Security export controls determine which advanced GPUs reach India under what conditions; the Indian regulatory architecture has no instrument that controls this, and the centre's compute scale is conditioned on a regime outside Indian sovereignty.
The capital structure of AI infrastructure entry adds a second binding variable beneath the regulatory stack, and the state is not the primary capital pool. The Union Budget 2026-27, presented on 1 February 2026, allocated one thousand crore rupees to the IndiaAI Mission for the year, half the two thousand crore allocated in 2025-26 against actual utilisation of approximately eight hundred crore. The Parliamentary Standing Committee on Communications and Information Technology flagged the cut as a fifty percent pruning of the flagship programme; mid-year discussions of expanding the Mission to twenty thousand crore rupees gave way to contraction. The state's annual disbursement to the Mission now sits two orders of magnitude below the private quanta announced in the same year, and the architecture's flagship instrument has been tacitly relegated to a regulator-and-subsidy role rather than a primary capital pool.
The capital entering Indian AI infrastructure is concentrated in two entity classes that can fund both the data-centre build and the captive power alongside it: Indian conglomerates with diversified balance sheets and US hyperscalers with global cost-of-capital advantages; the mid-tier MNC enters a market priced by capital pools whose cost, horizon, and parallel-power capability are not comparable.
Reliance and Adani at one hundred billion US dollar quanta operate against an internal cost of capital and a return horizon a standalone hyperscale entrant cannot replicate. Microsoft and Google at fifteen to seventeen billion operate against a global parent-balance-sheet advantage. The mid-tier MNC enters a market structure where the binding capital decision has already been priced by these two entity classes, with the Indian state's flagship Mission funding a fraction of either's annual deployment cadence.
For a country head evaluating AI infrastructure entry, the operative reading is that the IndiaAI Mission's announcement is the visible architecture, the build determinant sits in the intersection of six regimes the Mission does not control, and the competitive structure sits in two entity classes whose capital pools are not comparable. The investor who treats IndiaAI as the architectural entry point has read the announcement; the architecture is the six-stack the announcement does not articulate, and the market structure is the bimodal entrant typology the announcement does not name.