How AI is Forcing Data Centers to Trade Uptime for Raw Power

Data centres have been designed to be fail-safe. Every facility is built to avoid any downtime in case the application running on its infrastructure is mission-critical, such as for stock exchanges, payment networks, and telecommunications systems. 

Now, the explosive growth of artificial intelligence has introduced a radically different set of demands. The question is no longer simply "how resilient can we make this facility?" but "how resilient does this particular workload actually require us to be?"

The numbers driving this transformation are staggering. ABI Research forecasts that global active data centre IT capacity will grow sixfold between 2025 and 2035 to 147.1 gigawatts. According to JLL, AI and cloud demand are expected to drive the global data centre sector, with nearly 100 GW of new capacity added between 2026 and 2030. 

In India, there is pressure on operators to deploy quickly, efficiently, and intelligently. That requires relooking at how data centres are designed.

One-Size-Fits-All Approach No Longer Works

Anders Fryxell, Chief Sales Officer at the Nordic high-performance computing service atNorth, argues that the industry has already crossed a threshold. 

"A 'one-size-fits-all' approach no longer suffices," he tells AIM. "Workloads are increasingly segmented based on their individual requirements, from processing speed to privacy laws, and it is all in incredibly high demand. At a time when digitalisation is intrinsically linked to business continuity, time to market is a fundamental factor. Those operators that can provide AI-ready, sustainable digital infrastructure at speed will be able to capitalise on the industry's exponential expansion."

Amit Jaju, Senior Managing Director for India at Ankura Consulting, sees the shift as both inevitable and, if managed well, an opportunity. While traditional data centres were built to guarantee maximum uptime for legacy enterprise workloads, the AI era demands a fundamentally different logic.

"Large-scale AI training environments are highly distributed and frequently rely on checkpointing, meaning they function similarly to batch processing.They don’t always require the stringent uptime or multiple layers of failover necessary for latency-sensitive inference tasks or critical cybersecurity operations," he explains.

His proposed solution is a segmented, hybrid model: maintain high redundancy for critical data, while developing purpose-built, high-density AI zones with optimised redundancy specifically for model training. 

"This tailored flexibility prevents the costly mistake of overbuilding and allows compute capacity to scale precisely as demand grows," Jaju notes. "By moving away from a uniform maximum redundancy model, businesses and governments can channel their infrastructure cost savings directly into core AI innovation, robust data privacy measures, and advanced threat detection, thereby enhancing their competitive edge and bottom line."

Data center design is moving from a one-size-fits-all, redundancy-heavy model to a workload-aligned architectures, where infrastructure is tailored to how specific applications behave rather than built on legacy assumptions of uniform power, cooling and uptime requirements, says C R Srinivasan, Chief Executive Officer, Digital Connexion

"This shift is being accelerated by modular, prefabricated building blocks—specifically off-site manufactured skids like electrical and cooling units—that allow operators to deploy infrastructure in a more granular and flexible manner. Electrical skids integrate UPS, switchgear, and transformers into factory-tested modules supporting N, N+1, or 2N redundancy tailored to workload SLAs, while cooling skids provide scalable chillers and heat exchangers for precise thermal management of AI-driven densities. As highlighted in industry analyses such as Deloitte’s Asia-Pacific Data Center Outlook, AI workloads are driving the need for differentiated digital infrastructure environments, from high-density training clusters to distributed inference nodes," says Srinivasan.

Training vs Inference

Data centres now need to tailor the compute based on training and inference. Training involves the intensive, often weeks-long computational process of building an AI model, while inference is the act of running that model to generate a response or output in real-time. 

Deloitte estimated that inference workloads accounted for half of all AI compute in 2025, a figure set to reach two-thirds in 2026. At NVIDIA's GTC 2026 conference, chief executive Jensen Huang declared 2026 the "year of inference," projecting at least $1 trillion in cumulative demand for Blackwell and Vera Rubin systems through 2027, double the prior year's forecast.

Abhinav Sinha, Co-founder and CEO of AI-native cloud infrastructure startup Nava, draws a clear physical distinction between the two. Training, he says, is associated with large, centralised campuses requiring massive power availability. "Inference, by contrast, is driven by user proximity. It is decentralised, closer to users, and latency-sensitive," he explains. 

This also shapes capital allocation: training is capex-heavy with dense power, liquid cooling, and larger reserved power blocks, while inference follows a modular, scale-out model, more locations, smaller increments, and greater flexibility as workloads evolve.

The Cost of Overengineering

The data centre industry has long defaulted to maximum resilience as a precautionary principle. In the AI era, that default is becoming an expensive liability.

A Deloitte survey of data centre operators and power companies found that the primary challenge for infrastructure build-out was power and grid capacity, which 72% of respondents rated as very or extremely challenging.

Harqs Singh, CTO and Co-founder of AI infrastructure company InfraPartners, argues that overengineering resilience across the board locks up capital that could otherwise fund more compute capacity. Layers of redundant power systems, extensive backup infrastructure, and fully duplicated environments drive up capital costs, introduce operational complexity, and slow the time to taking the capacity online.

Sinha, however, offers a measured counterpoint. 

"While for some workloads it may seem excessive, strong resilience and reliability remain the bedrock of the data centre business. Tier-3, 4 standards provide consistency, transparency, and a baseline for trust, especially for customer-facing services," he explains. 

He points to the recent attacks on AWS data centres in the Gulf amidst the ongoing Iran war as a warning that even fully redundant systems can fail when the threat model is wrong. "True resilience is not just within a site; it must extend across geographies, supply chains, and dependencies."

"Power and cooling are increasingly becoming central enablers of data center development, particularly in the context of AI-driven workloads. As AI adoption accelerates, rack densities are increasing significantly, often requiring 10 to 15 times more power than traditional enterprise environments. This shift is fundamentally reshaping both site selection and facility design, where access to scalable, reliable and sustainable power and, the ability to support high-density cooling are now primary decision factors," says Srinivasan.

India: A Different Set of Constraints

In India, the more pressing challenge is closing a yawning capacity gap. The data centre capacity in India is deeply underpenetrated. Despite generating approximately 20% of the world's data, India holds only about 3% of global data centre capacity. 

That gap is set to narrow rapidly over the next five years. The new capacity being built will be designed from the outset for AI workloads, requiring higher load-bearing capacity, liquid cooling compatibility, and designs oriented towards inference rather than training.

"The real issue in India is not redundancy, it's lack of capacity and control," Sinha says. "Data centre design decisions, in any region, should be defined by a combination of customer workload profile and the physical infrastructure realities of that region."

Standardisation

The answer to the efficiency challenge, according to a growing consensus in the industry, lies in standardisation. 

Upgradeable data centres that leverage standard reference designs and are manufactured off-site in modular building blocks can be deployed far more quickly than traditional on-site construction, helping to close the gap between demand and supply. Citigroup estimates that AI-related infrastructure spend could reach around $490 billion as early as 2026, as hyperscalers and enterprises race to deploy AI capabilities at scale.

Physical constraints remain formidable. High-performance GPUs generate heat at densities that conventional cooling infrastructure was not designed to handle. JLL says rack densities above 70 kW increasingly make liquid cooling the only practical option, with operators actively planning and deploying direct liquid cooling as AI infrastructure expands.

The industry is now converging on a portfolio approach— that no single design philosophy can serve the full spectrum of AI workloads, and that the facilities of the future will need to be as varied as the applications they support.