Experts at India AI Impact Summit urge compute parity and data sovereignty, amid access barriers for SMEs

Did you know that every search query made on a public large language model (LLM) consumes half a litre of kerosene? Drawing attention to this phenomenon of AI guzzling gargantuan amounts of energy were experts who convened on the third day of the India AI Impact Summit 2026, held on 18 February at Bharat Mandapam in New Delhi.
Anne Neuberger, Strategic Advisor at a16z; Gokul V Subramaniam, Intel India President; and Kalyan Kumar, Chief Product Officer at HCL Software, were part of a panel titled Understanding Big Barriers for Bharat’s Growth with Stakeholders, moderated by Sunil Gupta, Co-founder, Managing Director, and CEO at Yotta.

The panellists delved into the irony of India’s massive artificial intelligence (AI) market projections on one hand, contrasting with access disparity on the other hand, especially when it comes to smaller businesses.
India’s small and medium enterprises (SMEs) risk being left out of the global artificial intelligence (AI) innovation race, owing to exceedingly high compute costs, often running into millions for even basic model training, and a glaring infrastructure gap that favours larger corporations with higher capital flows. Indian SMEs have typically been facing a persistent credit crisis, which acts as a deterrent for these small businesses to access hyperscale data centres or affordable GPUs. Hence, they struggle to deploy AI for everyday needs, such as inventory prediction or customer analytics, leaving them lagging behind larger firms already leveraging the technology. This disparity in accessing AI threatens the nation’s broader economic leap, as SMEs form the backbone of job creation and innovation in sectors from manufacturing to agritech.

Compute costs as the SME Chokepoint
Gokul V Subramaniam, Intel India President, addressed engineering advances in AI deployment. He emphasised starting with end-user needs, likening it to choosing the right car for the journey. “Heterogeneous compute is one of the key capabilities,” Subramaniam said. “So, you want to start with what workload, what user experience. Keep that in mind as you architect [AI]. And what I mean by heterogeneous compute is making sure that you can find a way to run it in the most affordable compute capability with the right performance and the right power efficiency.”
Subramaniam advocated for XPUs, that is, a combination of CPUs, GPUs, and NPUs, tailored to deployment contexts, from data centres to edge devices, all within an open ecosystem. This approach promises scalability for resource-strapped Bharat SMEs.

From pilots to enterprise production
Kalyan Kumar, Chief Product Officer at HCL Software, turned to enterprise hurdles, invoking a witty historical nod to explain legacy baggage. “God made the world in seven days because he had no installed base,” Kumar remarked, highlighting decades of layered systems stifling AI. He pinpointed the core issue: organisations built apps first, burying data. “The fundamentals of AI are that it needs access to data, independence of data, so data products, having better metadata, definition of data, and then the ability to get access to LLMs [large language models] or SLMs [small language models] or agents to access them.”

Driving AI skills across talent pools
Subramaniam outlined paths to broaden AI skills beyond elites. Fresh graduates adapt easily, he noted, but veterans require reinvention, especially in chip design. “Our goal is to make sure that for each function, each role, and each capability, in some way, they have an exoskeleton of sorts,” Subramaniam explained. “They walk into the campus, when they badge in, they actually feel like they’ve brought in another assistant with them for what they’re doing.”

Kumar echoed this, urging industry-academia pacts for curriculum revamps valuing grit and multi-disciplinarity. “You can’t be teaching people programming languages that you’re not going to use anywhere,” he said. “Your curriculum has to go through a complete re-jig.”

Black box distrust warrants data transparency
Anne Neuberger, Strategic Advisor at a16z, addressed the elephant in the room, talking about black box distrust, which is vital for high-stakes adoption. “Humans don’t trust black boxes, and we shouldn’t, particularly for things that touch important parts of our lives,” Neuberger stated. She called for data transparency, explainability in model reasoning, such as outlier recommendations in water purification, and continuous training, backed by evolving regulations beyond mere checkboxes.

Reconciling sovereignty with global data flows
Neuberger advocated federated learning to reconcile the tension between AI’s need for vast centralised data clusters and each nation’s demand for sovereignty over sensitive information. “The key will need to be federated data learning,” she said, citing US pilots with the Department of Health and Human Services that trained models on drug discovery and rare cancers while keeping health data local. This approach lets countries pool insights globally without physically moving data, protecting national security and privacy while enabling collaboration.

Kumar reinforced this sentiment, advocating data architecture shifts, such as treating data as independent products through catalogues, metadata discovery, and knowledge graphs to map real flows. He invoked the former Intel CEO’s wisdom on immutable forces: laws of physics, economics, and the land, explaining why pure centralisation remains elusive. “Centralisation is a dream, but it’s never going to happen in reality,” he observed, as data inevitably fragments to the edges in practice.

Half a litre of kerosene per search query
Kumar warned against compute gluttony, pushing small models and FinOps for optimisation. “Please don’t use LLMs as a search engine alternative. I’ll tell you, people are now using these public language models. And you’re going to burn like half a litre of kerosene for every query,” he cautioned.

On democratisation, he hailed the India AI Impact Summit’s real-time Indic language translation as democratisation in action. “One of the things that I, which we should also be very, despite all the challenges of how we got into this room and be sitting here, is the fact that the [India] AI Impact Summit has brought one aspect, which is just the Indic language translation,” Subramaniam said. “Being able to take English and having that in all of our Indic languages... has actually broken that barrier.”

The India AI Impact Summit 2026, being held from February 16 to 20 at Bharat Mandapam in New Delhi, features participation from more than 110 countries and 30 international organisations, including about 20 heads of state or government and almost 45 ministers, alongside 400 exhibitors.