India is building its AI infrastructure. The next challenge is making it pay

On the morning of February 19, the exhibition halls of Bharat Mandapam in New Delhi filled with the hum of an industry convinced it was watching history. Prime Minister Narendra Modi stood for photographs between Sam Altman and Dario Amodei, two men whose companies did not exist 10 years ago but now help shape how much of the world thinks. By the end of the week, India’s two largest conglomerates had pledged roughly $210 billion between them to build AI and data infrastructure.

The photographs will fade, as photographs from summits do; the arithmetic underneath them will not. India is expected to need between 19 and 23 gigawatts of data-centre capacity by 2030, up from roughly 1.6 GW today, requiring an estimated $350-435 billion in investment. The summit’s pledges were not the finish line. They were, at best, a substantial downpayment.

CRISIL’s analysis, published in June, names commercial viability as one of three tests by which India’s AI ecosystem will ultimately be judged and notes that nearly half of all AI-related capital deployed in the country between 2022 and June 2026 has already gone into infrastructure. The early evidence on what that infrastructure is producing is encouraging, but uneven. Dun & Bradstreet’s 2026 survey of Indian businesses found that 73% report some measurable AI return, while only 23% report broad or strong returns across multiple projects. AI value, in other words, remains concentrated in pockets: proof of concept, not yet proof of scale.

The question India now has to answer is not how much computing capacity it can build. It is how much of that capacity it can convert into value deliberately, workload by workload, rather than trusting that scale alone will do the converting.

India’s AI Infrastructure Boom Needs a New Measure of Success

Most of what is driving India’s data-centre expansion shares one characteristic: it is hungrier than what came before. CRISIL’s finding that nearly half of tracked AI capital has gone into infrastructure means utilisation, not capacity alone, now belongs in the same sentence as gigawatts and GPUs.

The temptation, in any capital cycle this large, is to keep score in the units that are easiest to announce: megawatts switched on, campuses broken ground, chips imported. India would do better to keep score in the units that are harder to announce and far more informative: workloads actually supported, productivity actually gained, revenue actually generated, costs actually reduced.

Not Every AI Workload Deserves the Same Investment

Part of the discipline this next phase requires is distinguishing between two kinds of AI demand. Training a large model is a sprint: enormous, concentrated compute, run occasionally, at a scale only a handful of organisations will ever need. Inference, the everyday business of a model actually answering a question, approving a claim, or flagging a transaction, is a different animal entirely: distributed, constant, and unglamorous. Of the 7-9 GW of AI infrastructure India may need by 2030, projections suggest 5-6 GW will be inference, not training.

This is the quieter decision Indian enterprises now face: evaluating each workload on its own terms and being honest about which applications justify premium infrastructure and which can run perfectly well on smaller, cheaper, more specialised models.

Sovereignty Should Become a Workload-level Investment Decision

Sovereign AI raises questions that have nothing to do with raw performance or cost: where a workload is processed, where the underlying data physically resides, which jurisdiction ultimately governs the infrastructure, and how much operational control an organisation is willing to give up in exchange for convenience. For sectors such as BFSI, healthcare, government, and critical infrastructure, the case for domestic or sovereign deployment is genuine. For a great deal of ordinary enterprise AI, it does not.

The mature response is not a single sovereignty policy applied uniformly across the economy, but a segmentation exercise repeated workload by workload until sovereignty becomes an insurance policy priced case by case.

The Real Cost of AI Is Bigger Than the Cost of Compute

The costs that are harder to see compound quietly beneath them: moving data between systems, storing it, carrying it across networks, the energy consumed along the way, operating the models themselves, and integrating everything.

As AI moves out of experimentation and into daily operation, the organisations that will actually see a return are the ones with real visibility into what each workload costs end to end, visibility that lets them find, and stop funding, the projects quietly delivering too little for what they consume. Governance, seen this way, is no longer only a risk-and-compliance function. It becomes an economic discipline in its own right: the mechanism by which a business finally learns where its infrastructure spend is going, what it costs, and whether the value coming back justifies it.

From AI Capacity to AI Productivity

There are signs Indian businesses are starting to close this loop. SAP’s 2026 India research found that AI already supports a third of business tasks at the average company, and that 74% of organisations report satisfaction with their AI returns so far.

The next phase has to go further: from isolated pockets of productivity to an enterprise-wide accounting of economic impact, with AI investment judged against revenue growth, operating-cost reduction, employee productivity, customer experience, and the speed of decisions actually made. Done well, this becomes a more disciplined investment model altogether: one in which infrastructure capacity follows demonstrated business value, rather than being poured in ahead of any clearly defined outcome and left to justify itself later.

The Advantage That Actually Compounds

India’s AI infrastructure cycle remains a genuine opportunity for productivity, for competitiveness, for the kind of long-term growth that changes what a country’s economy is capable of. As capacity becomes abundant, the ability to identify, prioritise, and operate the workloads that generate enough value to justify their cost will matter more than the ability to build the next data centre.

India has navigated a version of this choice before: choosing an interoperable payments rail over letting every bank dig its own well, the smartphone over the copper wire it never fully laid. It does not need to

relearn that instinct now. It needs to apply it again, at the scale of a $435-billion bet. India’s real AI advantage will come not from building more infrastructure than anyone else, but from extracting the highest economic value out of every unit of capacity it builds.

(The author is VP & MD, Qlik India. Views are personal.)