The Tokenomics Behind Enterprises Ditching the Cloud and Moving On-Prem

India's enterprises are increasingly moving away from a cloud-first approach as AI usage scales and enters production. As per a recent global survey by data platform Cloudera, 68% of Indian respondents had moved at least some AI workloads from public cloud to private cloud or on-premises infrastructure over the past year. 

Dell Technologies believes this shift is now happening "at scale" across sectors ranging from banks and government to fast-growing AI-native startups, even as the country's underlying data centre stock races to catch up with the demands of next-generation AI hardware.

In an exclusive interview with AIM, Chris Kelly, Senior Vice President, Infrastructure Solutions Group Specialty Sales, APJC, Dell Technologies, and Venkat Sitaram, Senior Director and Country Head, ISG, Dell Technologies India, say the changing economics of AI workloads are likely to shape where enterprises choose to run them.

According to Sitaram, enterprise AI adoption in India moves through a fairly consistent four-stage arc—experimentation, prototyping, piloting and full production—and the economics of where that workload should physically live shift sharply as it progresses through each stage. 

Early experiments, he says, almost always start on the public cloud, where entry costs are low, and infrastructure is instantly available without capital outlay. But as underlying data volumes grow and a model moves from proof-of-concept to production, the cost of running that workload in a hyperscale cloud environment escalates quickly enough to push enterprises back toward on-premises deployment.

"That cost arbitrage is pushing things back on-prem," Sitaram says, adding that this pattern is especially visible among India's emerging neo-cloud and GPU-as-a-service providers, many of whom are built on Dell's stack to offer sovereign AI capacity locally. 

For data-sensitive and heavily regulated sectors such as banking, financial services, and government, there often isn't a choice to begin with, as these organisations tend to build AI infrastructure natively on-premises from day one rather than passing through a cloud-first phase at all.

Kelly frames the tipping point less around company size and more around the predictability of demand. Enterprises able to reasonably forecast their compute requirements over a six-to-twelve-month horizon, he notes, can typically justify moving to on-prem immediately and see a better return on investment than staying on rented cloud infrastructure—a calculus he says holds true "whether you're small or large." 

Public cloud, in his framing, serves mainly for short, unpredictable bursts of compute demand rather than as a permanent home for steady-state AI workloads. Also, he pushes back on the idea that this requires deep pockets. "You don't have to invest the same capex... as a small company just to get things going," he says.

The Tokenomics Behind the Shift

Further, Kelly offers a cost comparison that is reshaping conversations with IT leaders. 

According to BenchLM Token Price Index, the median blended cost for flagship frontier model sits around $3–5 per million tokens for a standard tier. So, if a single user runs 10 AI agents burning a billion tokens a day through a frontier model in the cloud, it could cost an enterprise $3,000 to $5,000. Kelly compares running an equivalent workload on an open-source model on a Dell workstation, where, he claims, the ongoing marginal cost is "next to zero" once the upfront hardware cost is absorbed. 

That math, he says, is why enterprises are increasingly building internal governance frameworks that route routine tasks to smaller, cheaper open-source models running locally, reserving expensive frontier models only for workloads that genuinely need that capability. 

“I think there needs to be harness or governance frameworks around your AI models that help to put some processes in place regarding workload allocation,” he says.

Data Centre Hardware

Much of India's existing data centre stock was built over the past decade for conventional enterprise computing, with racks drawing between 5–15 kW of power. Generative AI workloads, however, require denser configurations of up to 250 kW per rack that need liquid cooling.  

Kelly notes the shift extends to structural fundamentals like floor-loading capacity, elevated doorway heights, and staging areas designed for far heavier equipment. He notes that a typical enterprise rack in the past weighed a couple of hundred kilograms; newer liquid-cooled AI racks can weigh over two tonnes, a load many older facilities were never engineered to bear.

Also, Sitaram notes that AI is reshaping the basic architecture of enterprise computing itself—pulling it away from a centralised data-centre-only model toward a far more distributed one. 

Not every AI workload needs to run in a data centre anymore, he says, elaborating that some can run at the edge, some on a desk-side workstation, and some centrally, with enterprises increasingly needing a strategy that spans all three tiers rather than defaulting to one. 

Dell's own portfolio, he notes, is built to support this spread—from AI PCs and precision workstations at the desk, through node-based Power Edge servers for mid-scale enterprise deployment, up to full rack-scale, liquid-cooled AI factories for the largest training and inference workloads.

Sovereignty as Infrastructure Opportunity

Both executives point to data sovereignty as an increasingly concrete infrastructure line item rather than a compliance afterthought. 

Sitaram clarifies that sovereignty concerns in India are less about where data physically sits and more about how it moves across jurisdictions and is protected throughout its lifecycle, creating demand for data pipeline management, storage, and cyber-resilience solutions. Dell is positioning to capture that opportunity, particularly as agentic AI systems multiply potential points of vulnerability inside an enterprise. 

"Every agent could become a point of risk," Sitaram says, framing resilience and immutable environments as a growing opportunity in their own right.

That opportunity, the executives say, is being pursued across an unusually broad customer base simultaneously, including large enterprises, global capability centres (GCCs), public sector bodies building sovereign AI infrastructure at national scale, and a fast-growing cohort of AI-native startups building agentic products from the ground up.