India’s artificial intelligence narrative is currently dominated by massive numbers. The IndiaAI Mission has successfully onboarded over 38,000 GPUs to democratize compute access, driving down costs to as low as INR65 an hour, while global hyperscalers have poured a staggering $80 billion into local data centers.
But according to Manoj Dhanda Founder and CEO of Utho Cloud, an Indian sovereign cloud and AI infrastructure platform, raw hardware is only a third of what production-grade enterprise AI actually requires.
“India has solved compute access faster than compute usability. To build a truly self-sustaining, sovereign digital ecosystem, the nation must look past flashy hardware announcements and address the hidden technical, economic, and geopolitical realities of the AI stack.”
“India has solved compute access faster than compute usability. To build a truly self-sustaining, sovereign digital ecosystem, the nation must look past flashy hardware announcements and address the hidden technical, economic, and geopolitical realities of the AI stack,” he told The Tech Panda.
He adds that the next phase of the mission should be measured not by GPUs onboarded, but by how many of them are running real production workloads a year from now.
The Missing Pieces
While the IndiaAI Mission has successfully onboarded over 38,000 GPUs to democratize compute access for startups, hardware is only one piece of the puzzle. Missing infrastructure layers, such as specialized data pipelines, proprietary networking fabrics, or localized security frameworks, remain. Indian startups must now build or acquire to scale these systems to enterprise grade.
“Getting access to a GPU is not the same as running a production AI system. That gap is bigger than most people realise. The IndiaAI Mission’s 38,231 GPUs, priced as low as Rs 65 an hour, solved a real problem: Indian startups could not afford this kind of compute before. But raw hardware is maybe a third of what enterprise-grade AI actually needs,” Dhanda says.
“Getting access to a GPU is not the same as running a production AI system. That gap is bigger than most people realise.”
Three layers are still missing, he says. First, high-speed networking between GPUs, the kind InfiniBand or NVLink-class fabric provides.
“Without it, a cluster trains only as fast as its slowest data link, no matter how many chips sit inside it.”
Second, data pipeline infrastructure.
“Most Indian enterprises have data scattered across legacy systems and paper trails, not sitting clean and labelled in one place, and building that pipeline is often a longer job than training the model itself.”
Read more: Beyond the OTP: “Aadhaar eSign has genuinely changed onboarding economics”
Third, localised security frameworks: ISO 27001, SOC 2, and increasingly MeitY empanelment, for anyone whose workload touches banking, healthcare, or government data.
“Not only this, orchestration software to schedule GPU jobs efficiently is its own specialised discipline, and very few teams in India have built it yet.”
Ownership Over Residency: Managing the Hyperscaler Surge
Global hyperscalers like Microsoft are investing tens of billions of dollars to build massive local data center hubs, capturing early enterprise workloads. Can India effectively balance its reliance on foreign big-tech capital with its strategic mandate for digital sovereignty and indigenous technology stack development?
Before responding, Dhanda talks about scale, to put the question in perspective. He points out that Microsoft, AWS, and Google have together committed more than US$80 billion to Indian data centres in roughly the last year. The IndiaAI Mission, India’s own flagship push for sovereign AI compute, has a total outlay of about US$1.25 billion.
“What India needs is precision about which layers must stay sovereign regardless of who is funding the data centre next door. Data residency, meaning the servers sit in India, is necessary but not sufficient. What matters more is who holds the encryption keys, and who can be forced to cut access by a foreign court order or a foreign sanctions list”
“That is not a small gap; it is a different order of magnitude,” he says.
“I don’t think the answer is to turn this capital away,” he continues. “It brings real jobs and real skilling.”
Microsoft’s own commitment includes training for over 20 million people.
“What India needs is precision about which layers must stay sovereign regardless of who is funding the data centre next door. Data residency, meaning the servers sit in India, is necessary but not sufficient. What matters more is who holds the encryption keys, and who can be forced to cut access by a foreign court order or a foreign sanctions list,” he explains.
“We already know what that dependency can cost,” he further points out. “An Indian company had its email and cloud-hosted data cut off overnight last year, not from a technical fault, but from sanctions on its ownership structure, enforced by a platform it did not control,” he recalls.
“Welcome the capital, but build the sovereign layer in parallel, not later. Every rupee of hyperscaler investment in India should be matched by a rupee of investment in Indian-owned infrastructure for the workloads that cannot afford to go dark,” he advises.
Earned, Not Bought: India’s Honest Fab Timeline
Under the Design Linked Incentive (DLI) Scheme, India is seeing the fabrication of homegrown chips and the rapid development of reusable IP cores. However, how close is India from breaking its reliance on foreign-manufactured chip components, and what remains the biggest bottleneck in the local supply chain?
Dhanda promises a real answer, not a diplomatic one. The news is good when it comes to chip design.
“On chip design, India is moving faster than people expect. The DLI scheme has already produced 16 tape-outs and 6 fabricated chips, including an indigenous AI chip built on an advanced 12-nanometer process. That part of the story is genuinely ahead of schedule,” he states.
“The real bottleneck is yield, the years of process engineering it takes a fab to reliably produce chips at scale without defects. That cannot be bought. It has to be earned one wafer at a time. Ten years is a realistic number. Anyone promising five is selling a headline, not a fab”.
But when it comes to manufacturing the story differs. He points out that India’s most-watched fab, the Tata-PSMC plant in Dholera, was originally promised to open at the 28-nanometre node by the end of this year. It is now expected to reach commercial production only in 2028, starting at the older 90-nanometer node instead.
“That is not a failure, fabs everywhere in the world slip on their first run. But it tells you the honest timeline, not the announcement timeline,” he says. “The government’s own target points the same way. It wants India ranked among the top four semiconductor manufacturing nations by 2035, not 2030. I’d treat that as the real marker for self-reliance, not the tape-out headlines we celebrate today.”
India has earmarked over INR76,000 crore in incentives already, with INR65,000 crore already committed to approved projects. So, the bottleneck, he says, is not money.
He says it straight. “The real bottleneck is yield, the years of process engineering it takes a fab to reliably produce chips at scale without defects. That cannot be bought. It has to be earned one wafer at a time. Ten years is a realistic number. Anyone promising five is selling a headline, not a fab”.
Mitigating Concentration in Critical Sectors
As critical sectors like banking, healthcare, and government systems embed AI deeply into their daily operations, they are increasingly relying on foreign cloud architectures. Leaving India’s strategic digital infrastructure on foreign commercial platforms creates permanent economic drains and severe national security vulnerabilities.
Dhanda points out that there are two separate risks here, and people often mix them up.
“One is economic: Indian banks and hospitals pay foreign hyperscalers indefinitely, in dollars, for infrastructure they will never own. That is a permanent outflow with no equity ever built. The other is a control risk, and it worries me more,” he says.
He further adds that the RBI’s own outsourcing guidance treats dependence above roughly 30 to 40 percent on a single vendor as a concentration risk serious enough to need board-level attention.
“Regulators do not write rules like that for hypothetical problems,” he says.
He recalls last year’s Nayara incident where the Indian company lost access to its email, collaboration tools, and cloud-hosted data overnight last year, not because of a technical failure, but because of sanctions on its ownership structure, enforced by a foreign platform it did not control.
“For certain workloads, control matters more than convenience. Government and banking data should live somewhere a decision made in another country cannot switch it off.”
“If that can happen to a corporate account, the same exposure applies to government or banking workloads sitting on comparable infrastructure,” he warns.
As an outcome, the Reserve Bank of India is launching the Indian Financial Services (IFS) Cloud through its subsidiary, Indian Financial Technology and Allied Services (IFTAS), to strengthen national data security and regulatory control.
“That is not a symbolic move. It’s an admission that for certain workloads, control matters more than convenience. Government and banking data should live somewhere a decision made in another country cannot switch it off,” he comments.
Building AI That Fits India’s Unique Economic Fabric
Moving past the immediate hype of hardware procurement and semiconductor manufacturing timelines reveals the true benchmark of success, a mature, self-sustaining Indian AI ecosystem by 2030, anchored by localized disruption in critical socioeconomic sectors.
“By 2030, we don’t think the winning story will be about how many GPUs India owns. It will be about which sectors run on sovereign AI without anyone noticing it’s sovereign, because it simply works, in their language, at their price,” Dhanda says.
“By 2030, we don’t think the winning story will be about how many GPUs India owns. It will be about which sectors run on sovereign AI without anyone noticing it’s sovereign, because it simply works, in their language, at their price,”
Already, BHASHINI, the government’s language AI platform, now runs entirely on Indian cloud and GPU infrastructure. Also, Sarvam has released a large language model trained specifically for Indian languages.
“These aren’t lab demos; they’re in production today,” he points out.
Dhanda says the sectors to watch closely are the ones global AI labs have the least commercial reason to build for. He includes agriculture, through voice-based AI in regional languages, for farmers who are far more comfortable speaking than typing in English. Healthcare, through diagnostic AI reaching tier 2 and tier 3 cities where specialist access is genuinely scarce. And financial inclusion, through lending models trained on India’s own transaction patterns; UPI alone processes around 22 billion transactions a month, a dataset no foreign model was ever built on.
Read more: India’s AI Paradox: Aggressive Investment Meets a Hard Data-Infrastructure Reality Check
“India’s AI data centre capacity is projected to grow roughly 24 times between 2025 and 2030. The mistake would be measuring success in megawatts. The real measure is how many of those megawatts run models that understand a shopkeeper in Coimbatore as well as they understand someone typing in English. That is what a mature, self-sustaining AI ecosystem looks like: not less foreign dependence for its own sake, but AI that finally fits the country it was built for,” he concludes.
