From energy headroom to AI-ready campuses: What high-density compute actually needs

By Sunil Gupta, CEO and MD, Yotta Data Services

The global narrative around AI infrastructure often assumes that countries with surplus renewable energy are naturally positioned to lead the next wave of AI data centers, but energy headroom alone does not translate into AI-ready infrastructure. What matters is not availability at the grid level, but the ability to deliver stable, high-quality, and continuous power at the point of compute – across the campus, facility, and rack. AI workloads are fundamentally different from traditional enterprise computing, requiring specialised GPU infrastructure deployed in tightly coupled, high-performance clusters. A conventional data center rack historically operated at 5-15 kW, whereas AI-optimised racks now exceed 60–140 kW. Next-generation systems are pushing even higher as individual GPUs reach 1,000W to 1,200W TDP, requiring tens of megawatts to support clusters running across thousands of parallel nodes.

Power Delivery at Machine Scale

At these densities, power systems are being engineered to respond in near real time. GPU clusters can introduce sharp load shifts within seconds during scheduling or training cycles, which is driving changes in electrical design – from low-impedance distribution paths and high-speed switching systems to UPS architectures that can absorb both sustained draw and short-duration spikes without derating. Voltage tolerances are tighter, harmonic distortion is more actively managed, and power conditioning is now embedded into how cluster stability is maintained. Battery storage is increasingly integrated into the power layer itself, not just as contingency but to absorb short bursts of demand, support frequency stability, and smooth variability from the grid.

Redundancy design is moving beyond traditional N+1 models toward limiting the impact of failures across large compute environments. In tightly coupled AI clusters, localised power disturbances can propagate beyond a single rack, particularly during synchronised training jobs, making fault isolation a priority at the cluster level rather than just the component level.

This is driving more segmented power architectures, where distribution is aligned with workload groupings and failure domains are contained more deliberately. Power infrastructure is going through a subtle but important shift. UPS systems and power paths aren’t just there as backup, instead they’re being designed to handle sudden spikes in demand, especially when workloads ramp up, without causing voltage fluctuations that could throw off GPU operations. The emphasis is moving toward maintaining continuity under dynamic load conditions, rather than only protecting against complete outages.

This extends to how power is delivered at the campus level. While generation capacity is scaling, consistent delivery is shaped by transmission readiness, substation capacity, and last-mile connectivity. In India, data center capacity has expanded from around 375 MW in 2020 to over 1.5 GW in 2025, with projections over 9 GW by 2030.
India’s challenge, however, is not linear capacity expansion, rather it is a multi-gigawatt shift driven by the nature of its AI adoption curve. Unlike many economies, India is not building AI for a narrow enterprise base, but for population-scale deployment. The country has already demonstrated this pattern through digital public infrastructure such as UPI, Aadhaar, ONDC and CoWIN, where systems designed for scale saw rapid and widespread adoption.

AI is expected to follow a similar trajectory, placing simultaneous demand on three fronts: training sovereign and domain-specific models, enabling inference at population scale across sectors such as governance, healthcare, and financial inclusion, and supporting global AI companies that will need to host and serve Indian workloads locally due to latency, data residency, and trust requirements. Taken together, this shifts the requirement from incremental megawatt additions to sustained, multi-gigawatt infrastructure buildouts.

As a result, the constraint is less about generation and more about how efficiently power is routed to high-density campuses. Dedicated high-voltage connections, long-term power purchase agreements, and integrated storage are becoming standard, particularly when aligning renewable energy with continuous, high-utilisation workloads.

Cooling and Compute Begin to Converge
Cooling architecture is undergoing a structural shift. At rack densities beyond 50–60 kW, air cooling becomes thermodynamically inefficient due to airflow constraints and heat recirculation. This is accelerating the move toward liquid cooling, particularly direct-to-chip approaches, where heat is managed at the component level rather than across the entire rack. Instead of relying on ambient air, these systems use a circulating fluid interface in close contact with high-heat components, allowing heat to be drawn away more efficiently and handled through dedicated cooling loops.

This shift supports significantly higher rack densities and improves overall energy efficiency compared to conventional air-based designs. As a result, cooling is increasingly intertwined with compute design, influencing rack layouts, floor planning, and even chip-level packaging. With cooling accounting for a substantial share of total energy use, achieving lower PUE, often targeting 1.2 or below, depends heavily on how well these systems are integrated into the broader infrastructure.

Scaling Campuses, Not Just Capacity
At scale, these requirements are reshaping infrastructure planning. Hyperscale AI deployments are crossing 100-300 MW per campus, forcing a transition from sprawling facilities to hyper-dense architectures governed by the ‘physics of proximity.’ Because high-speed GPU interconnects require minimal physical distance to maintain low latency, power and cooling must be concentrated into a tightly integrated campus fabric, where multiple buildings operate as a single logical machine rather than as isolated facilities. This is driving ‘Substation-to-Chip’ electrical designs that bring high-voltage utility feeds directly to the white space to minimize distribution losses. Furthermore, as liquid cooling becomes the campus standard, the focus shifts toward thermal circularity, where the concentrated heat from 140 kW racks is harvested for industrial reuse. Regions that can align these elements – ensuring not just access to energy, but its precise, high-voltage delivery and the recovery of its thermal output – will lead the next phase of AI growth.