India’s AI bottleneck is moving beyond GPUs to data readiness: NetApp

India’s enterprise AI ambitions may be running into a problem that has little to do with the availability of GPUs. According to Premalakshmi Ramakrishnan, Managing Director & Area Vice President – India & SAARC, NetApp, and Vasanthi Ramesh, Vice President of Engineering & Engineering Site Leader, NetApp India, enterprises are increasingly finding that the harder part of scaling AI is getting their data ready, governed and available to the systems that need it.

“When we go and talk to customers in the market and especially in India, the first few four or five points that we can consistently hear across many of the organisations is, first thing is, my data is not AI ready,” Ramakrishnan said in an interaction with Fortune India.

The issue is emerging even as companies spend heavily on compute. Ramakrishnan said customers have built GPU clusters and fine-tuned models but are still struggling to scale AI initiatives because data may not be in the right place or format, or adequately governed. Preparing and integrating that data can itself require significant engineering work.

This is increasingly reflected in industry research. A Dun & Bradstreet survey of Indian businesses found that while AI adoption is widespread and 69% of respondents plan to increase AI investments, only 4% said their enterprise data was fully ready to support AI at scale. Another 58% described their data as partially ready. 

ServiceNow’s Enterprise AI Maturity Index 2026 similarly found that 74% of Indian enterprises identified data accuracy, access and management as an area requiring strengthening for AI, while 56% pointed to legacy-system integration. Despite enterprise AI investment in India rising 119% in a year, only 22% of organisations had AI testing, auditing and risk-assessment processes in place.

As AI scales, storage and data movement become the new bottlenecks

Ramakrishnan also pointed to a less visible constraint in AI infrastructure, as the the gap between the speed of GPUs and the ability of the underlying storage and networking infrastructure to feed them data.

“There’s a huge amount of CapEx investments already made on the GPUs, but they are underutilized. That’s largely because the data IO need to catch up with what the GPU demands are,” she said, calling the storage layer the “silent performance killer” when data movement between workloads and environments has not been adequately planned.

Vasanthi Ramesh made a similar point from an engineering perspective, arguing that the constraint is increasingly moving down the infrastructure stack. “Compute has gone like X number of times in speed and capabilities, but our networking and our IO has not evolved at all,” she said. “So now you have like super good GPUs that can process fast, but you don’t have the lower layer that will actually work that much faster. So you’ve created a bottleneck and you need to break that bottleneck.”

The implication is that AI infrastructure cannot be viewed simply as a race to add more GPUs. NetApp is pushing a disaggregated architecture approach, where compute and storage resources can be scaled more independently. Ramesh said this can shift the focus towards scaling the input/output and networking layer rather than continually adding compute.

The other challenge is where enterprise data actually resides. Ramakrishnan said data is increasingly being created across hyperscaler clouds, AI infrastructure and on-premise systems, making it difficult for enterprises to move everything between environments. She described AI as inherently hybrid, with enterprises needing to extract value from data wherever it resides.

Cloud economics is also complicating the equation. Ramesh said enterprises increasingly factor in the cost of moving large quantities of data out of the cloud when deciding between cloud, sovereign cloud or more isolated infrastructure. “Customers look at cost. One of the most important things is cost of their data,” Ramesh said, adding that cloud egress costs can become significant because AI models require large volumes of data for training.

The infrastructure build-out around AI is nevertheless accelerating in India. JLL estimates that the country’s operational data-centre capacity will rise from 1.6 GW in mid-2026 to 6 GW by 2029, while Wood Mackenzie expects overall capacity to reach 12 GW by 2030, with AI-dedicated capacity expanding nearly 24-fold to 6.55 GW. 

Ramakrishnan said NetApp expects nearly 8 GW of investments in India by 2030 linked to AI adoption.  While estimates differ depending on how data-centre capacity and AI-specific infrastructure are defined, the broader trend is clear, that India is entering a period of rapid infrastructure expansion as enterprises and hyperscalers prepare for higher AI workloads.

The executives also see adoption moving beyond technology companies. Ramakrishnan said banking, financial services and fintech are among the more mature sectors, with applications spanning fraud detection, credit underwriting, KYC, customer service and personalisation. IT services, telecom, retail and e-commerce are also adopting AI rapidly, while manufacturing, automotive and logistics are beginning to use it for robotics, fleet management, warehouse automation and digital twins.

Governance is becoming as important as AI capability

For Ramesh, however, the biggest enterprise challenge is not merely making data available to AI systems but deciding what AI systems should be allowed to do with it. “The biggest thing block right now is governance,” she said, pointing to uncertainty among enterprises around what data can be used, which processes should govern it and how those decisions should work across jurisdictions. This becomes particularly complicated for Indian organisations that operate global centres of excellence and have to comply with multiple regulatory environments.

The issue becomes more consequential as enterprises move from experimentation to agentic AI, where systems can increasingly take actions rather than merely generate recommendations. Ramesh said the defining concept for enterprise AI could increasingly become “trustworthy applications”, with organisations needing to ensure that data use is auditable, traceable and governed. “It is very important that they do what is right and they are auditable, they are traceable, there is some lineage for every information that comes,” she said.

This concern is consistent with the wider enterprise AI landscape. ServiceNow’s research found that only 22% of Indian enterprises currently have AI testing, auditing and risk-assessment processes, even as investment is accelerating. 

Ramesh said governance also has a direct bearing on cybersecurity. The more data is moved between environments, the more difficult it becomes to track and protect it, potentially expanding the attack surface. She argued that data readiness and governance therefore need to be embedded closer to the storage layer rather than treated as separate processes. “Your last defense is going to be your storage. And that’s the place where you need to be extremely strong,” Ramesh said, pointing to capabilities such as ransomware detection, recovery, isolated environments, encryption and data observability.

That makes the next phase of India’s AI infrastructure race less straightforward than simply adding more computing power. As enterprises move towards production-scale AI and agentic systems, the competitive advantage may increasingly depend on whether they can organise the data underneath those models—while controlling where it sits, who can access it, how it moves and whether every AI-driven decision can be traced back to a governed source.