India's AI Infrastructure Problem isn't Capacity. Its Location
As terrestrial AI expansion collides with power, land, and deployment delays, Indian companies are beginning to test whether some compute workloads belong beyond Earth.
News
- The Vatican Enters the AI Governance Debate
- Microsoft PullBack, Uber Budget Blowout Put AI Coding Costs in Focus
- Shastra VC Launches $100 Million Deeptech Fund
- Persistent, IIM Ahmedabad Launch AI Value Compass for Enterprise AI
- India Launches National AI Doctors Mission to Train Doctors in AI
- HCLTech Says 43% of Enterprise AI Projects May Fail
Image Credit- Chetan Jha/ MIT Sloan Management Review India
Key Takeaways
01
India’s terrestrial AI buildout is expanding, but production AI demand is beginning to move faster than the land, power, cooling, and deployment cycles needed to support it.
02
Recent partnerships show that Indian AI infrastructure planning is shifting from simple capacity accumulation toward workload-specific compute placement, including early orbital pilots.
03
Orbital inferencing remains experimental, but it is now being evaluated as one response to terrestrial latency, concentration, and infrastructure timing limits.
The first bottleneck in artificial intelligence was access to chips. The second is proving harder to solve. The world can now finance more accelerators than it can quickly house, power, cool, and connect. The International Energy Agency’s 2025 Energy and AI report estimates that electricity demand from global data centers will rise from roughly 485 terawatt-hours in 2025 to nearly 950 terawatt-hours by 2030, with AI-intensive facilities accounting for a substantial part of that increase.
A March 2026 academic paper, Concentrated Siting of AI Data Centers Drives Regional Power-System Stress Under Rising Global Compute Demand, finds that dense AI compute concentration is placing measurable stress on regional power systems as hyperscale clustering intensifies. This is no longer simply a utility issue. It is becoming a deployment issue.
Adding chips is only one part of AI expansion. Operators also need land near fiber corridors, utility upgrades, high-density cooling architecture, and permitting cycles that do not slow deployment for months or years. In several markets, those supporting systems are proving less elastic than AI demand itself.
India is entering that same phase. In late April, Google began work on its AI-focused data-center hub in Visakhapatnam with AdaniConneX and Bharti Airtel Nxtra as infrastructure partners. Yotta Data Services is continuing its sovereign GPU-backed campus expansion. Reliance Jio is building domestic compute infrastructure around larger AI workloads. These terrestrial investments point to a common market condition: AI-ready demand is broadening faster than conventional build cycles can comfortably absorb.
That helps explain why two orbital compute announcements in one quarter no longer look like isolated startup experiments. In February, NeevCloud signed a deal with Agnikul Cosmos to build an AI-focused orbital data-center platform in low Earth orbit, with a proof-of-concept launch targeted before the end of this year and commercial operations expected in 2027. In the first week of May, Pixxel and Sarvam AI announced Pathfinder, another orbital data-center satellite initiative expected to reach orbit by the end of 2026.
Two separate orbital compute pilots in less than a quarter suggest that part of the Indian AI infrastructure market is beginning to test a harder question: can a purely terrestrial hosting model efficiently serve every production AI workload now coming online?
Production Inference Is Arriving Faster Than the Ground Can Absorb It
Enterprise AI infrastructure has historically scaled in familiar ways. More workloads led to more racks, larger campuses, and additional cloud commitments. That model remains viable when the objective is broad compute accumulation. It becomes less reliable once AI systems move from pilots to live operational dependency.
Narendra Sen, Founder and Chief Executive of RackBank and NeevCloud, argues that infrastructure execution has now overtaken model development as the more immediate bottleneck. Building a terrestrial AI-ready facility requires land acquisition, utility approvals, high-density cooling systems, fiber connectivity, and regulatory clearance. In practice, he says, that process can consume as much as 24 months before production workloads begin.
AI deployment is not moving on a 24-month schedule.
The mismatch is sharpest in inference. Training remains periodic, centralized, and capital-intensive. Inference is continuous. AI systems embedded in industrial automation, border monitoring, healthcare diagnostics, logistics routing, autonomous machines, and maritime surveillance run continuously. A border sensor does not pause. A diagnostic system does not wait for the next build cycle. That infrastructure requirement is categorically different from anything conventional enterprise compute were designed to serve.
“Training is a one-time or infrequent intensive event. Inference is the continuous operational reality of every deployed AI system, every second of every day,” Sen says.
That distinction exposes a timing gap that terrestrial planning does not solve. Ground-based hyperscale facilities are built during construction time. Production inference demand expands in deployment time. The two are no longer moving together. India compounds this globally. Its AI deployment is accelerating in sectors where infrastructure gaps are most acute: rural healthcare, port logistics, border surveillance, and precision agriculture. These are not use cases that can wait for a 24-month data-center build cycle. They are use cases that will route around terrestrial constraints if terrestrial build speed cannot keep pace with deployment speed. That is the condition orbital compute is being designed to exploit.
Orbital Compute Is Being Tested Where Terrestrial Inference Starts to Fail
The immediate use case for data centers in space is narrower than much of the surrounding rhetoric suggests. These systems are not being designed to replace hyperscale AI campuses. They are being tested under workloads where terrestrial inference introduces too much distance, too much dependence on connectivity, or too much centralization.
NeevCloud’s planned orbital platform sits squarely in that category. Sen says the targeted applications include defense surveillance, autonomous drones, offshore monitoring, industrial telemetry, disaster response, and remote healthcare systems. These are workloads in which inference requests originate far from terrestrial compute centers and in which response quality degrades if every request must be routed back to a central facility.
“We are not just building a data center in space. We are building an entirely new layer of orbital inferencing infrastructure,” Sen says.
The platform is expected to place AI inference, storage, and compute modules in low Earth orbit at an altitude of roughly 350 to 500 kilometers (about 220 to 310 miles), powered by solar energy and managed through NeevCloud’s orchestration layer. The first configuration is designed within a 500-kilogram payload envelope, aimed at lightweight, continuous inference rather than dense model training.
The technical model is equally central to economics. Saraniya P., Vehicle Director of Agnibaan SOrTeD at Agnikul Cosmos, says the company is not launching a standalone satellite platform solely for hosting compute hardware. Instead, the upper stage of the launch vehicle is being repurposed into the orbital hosting layer once the primary payload is deployed.
“The launch vehicle itself becomes the infrastructure. That integration is what makes the economics fundamentally different from a conventional satellite deployment.” Saraniya P., Vehicle Director, Agnibaan SOrTeD, Agnikul Cosmos |
This changes the cost comparison. Conventional orbital systems duplicate hardware by requiring a separate satellite bus after launch. Agnikul’s approach attempts to reduce that duplication by converting existing launch architecture into operational infrastructure. Sen estimates that early deployments may still cost 10 to 20% more than terrestrial alternatives. That premium narrows as launch frequency increases, and the per-kilogram cost of reaching low-Earth orbit continues to fall. For workloads where the alternative is building a new terrestrial facility in a remote or underserved region, the orbital cost premium may already be competitive. The relevant benchmark is not the standard cloud hosting cost. It is the cost of serving workloads where terrestrial latency or terrestrial absence begins to interfere with execution.
Compute Location Is Becoming a Strategic Variable, Not Just a Cost Line
The NeevCloud-Agnikul pilot becomes more meaningful when placed beside the rest of India’s recent AI infrastructure activity. Google’s Visakhapatnam project with AdaniConneX and Bharti Airtel Nxtra represents long-horizon, hyperscale terrestrial capacity. Yotta’s sovereign GPU-backed campus expansion represents concentrated domestic compute ownership. Reliance’s AI compute investments point to larger local model-serving infrastructure. NeevCloud-Agnikul and Pixxel-Sarvam extend the same conversation into distributed orbital inferencing.
The commercial logic is straightforward. The strategic implications are less so. India has spent the past two years building sovereign compute capacity specifically to reduce its dependence on foreign cloud services for sensitive workloads. Orbital infrastructure introduces a new version of that question. An orbital platform operated by an Indian company with Indian launch infrastructure is a categorically different proposition from hyperscale capacity owned by a US hyperscaler and co-located domestically. For defense, border management, and critical national infrastructure workloads, compute sovereignty does not end at the atmosphere. Where inference physically runs and who controls the hardware running it are likely to become governance questions as much as procurement ones.
This does not yet amount to a broad shift away from terrestrial data centers. It does indicate that parts of the Indian market are beginning to divide AI workloads by compute location rather than by compute volume alone. India’s next AI buildout is therefore becoming as much a placement problem as a capacity problem.
Research Context
Primary interviews with senior leaders at NeevCloud, RackBank, and Agnikul Cosmos, conducted in April 2026. Quantitative sources: IEA Energy and AI report (2025); academic study Concentrated Siting of AI Data Centers Drives Regional Power-System Stress Under Rising Global Compute Demand (March 2026). Market evidence drawn from Indian AI infrastructure announcements, April–May 2026. |
Role | Required Action |
C-Suite: Budget for Workload Placement | By the next infrastructure budgeting cycle, organizations deploying AI into field operations, surveillance, industrial automation, logistics, or remote diagnostics need a workload map that separates centralized training demand from latency-sensitive inference demand. Without that distinction, terrestrial capital expenditure is likely to remain misaligned with operational bottlenecks. |
Technology Leaders: Classify Before You Commit | Within the next twelve months, CIO and infrastructure teams need to identify which production AI requests are tolerant of centralized hosting and which begin to degrade under terrestrial latency, fiber dependence, or remote connectivity gaps. That classification determines whether future spending belongs in conventional cloud contracts, sovereign GPU hosting, or distributed inference layers. |
Boards and Governance: Review Compute Concentration Risk | As AI systems move deeper into critical operations, infrastructure resilience reviews need to account for compute concentration risk, terrestrial build delays, and foreign cloud dependency. Model oversight without compute-location oversight leaves a strategic blind spot. |
India’s first AI infrastructure push followed a familiar formula: secure more chips and build more terrestrial capacity. Its next phase is likely to be shaped by a different question. Where should different AI decisions physically run if land, grid, and fiber cannot expand at the same speed as deployment demand? India’s orbital AI pilots do not settle that question, but they make one point harder to dismiss. The future AI infrastructure race may be defined less by how much compute companies can buy and more by where they can intelligently place it.
