At a glance
- Author: Sumati Sahgal, VP, Data Centre & Secure Power, Greater India India’s sovereign-compute procurement does not explicitly mandate liquid cooling
- However, cooling is increasingly tied to bid readiness for data centre operators and the partners designing their infrastructure
- IndiaAI increasingly evaluates outcomes influenced by cooling..
Sovereign GPU Procurement in India Elevates Cooling Readiness
India’s sovereign-compute agenda is reshaping how data centre operators approach infrastructure design. Under the IndiaAI programme, procurement increasingly evaluates outcomes such as power usage effectiveness (PUE), availability, domestic hosting, and sustained operating performance—all of which are directly influenced by cooling design. For operators bidding on high-density GPU capacity, cooling readiness is no longer a secondary engineering consideration; it is becoming a core component of bid competitiveness.
Procurement Outcomes Tied to Thermal Performance
IndiaAI’s evaluation framework demonstrates how cooling has moved into the foreground of infrastructure competitiveness. In 2024, the programme briefly introduced an operating data-centre PUE of 1.35 or below as an eligibility requirement. That threshold was later removed following industry feedback, but graded PUE scoring remained, with progressively higher technical scores awarded for lower PUE and the highest score available below 1.2.
Operational efficiency carries financial consequences as well. The continuous empanelment framework requires quarterly compliance reporting, and PUE improvements can reduce certain penalties. IndiaAI also applies service credits when availability falls below 99.95%. Because pumps, controls, coolant distribution, leak detection, and heat rejection all affect sustained operating performance, cooling infrastructure is now under greater scrutiny during both procurement and ongoing operations.
Domestic hosting adds another layer of constraint. IndiaAI requires AI services to be delivered from data centres located in India, meaning infrastructure planning must account for local climate, utility, and water conditions rather than relying on generic design assumptions.
GPU Platform Selection Dictates Thermal Architecture
Hardware selection can significantly narrow cooling choices. In the fourth IndiaAI procurement round, reporting reviewed for the research identified an offer of 17,000 NVIDIA B300 GPUs. This offer provides a current link to IndiaAI procurement, while NVIDIA’s Vera Rubin platform illustrates how quickly thermal requirements are increasing. Vera Rubin NVL72 combines 72 Rubin GPUs and 36 Vera CPUs in a rack-scale system, significantly raising the cooling capacity that GPU data centre operators in India must plan for as newer platforms enter deployment.
At these densities, GPU procurement creates requirements beyond the servers themselves. Cold plates, manifolds, coolant flow, CDU capacity, controls, leak detection, and facility heat rejection must all be planned around the selected compute platform. The GPU choice can effectively lock a project into a specific thermal architecture early in the deployment cycle. Data centre operators must confirm that the proposed hall can provide the required cooling capacity, redundancy, and facility connections at the intended rack locations.
This also limits the value of broad “AI-ready” claims. Customers need specific information on supported rack densities, available liquid loops, residual air-cooling requirements, heat-rejection capacity, and commissioning procedures. Because IndiaAI’s empaneled infrastructure spans several GPU generations and vendors, cooling requirements will vary across deployments. Designing around the server architecture provides a more reliable basis than applying a single cooling approach across every GPU environment.
Liquid Cooling Extends from Chip to Heat Rejection
Direct-to-chip cooling begins at the processor, but its performance depends on the full thermal path. Cold plates capture heat from high-power components. A Coolant Distribution Unit (CDU) provides the interface between the technology cooling system serving the IT equipment and the facility water system, managing heat exchange, temperature, pressure, flow, filtration, and coolant condition.
In facilities with suitable water infrastructure, liquid-to-liquid CDUs can connect the secondary loop to the building mechanical plant. Brownfield sites may require other approaches. Heat Dissipation Units (HDUs) or liquid-to-air configurations can introduce liquid-cooled IT where a usable facility-water connection is unavailable, transferring heat back into the room or air-side cooling system. The facility still needs sufficient capacity to remove that heat.
Rear-door heat exchangers can support retrofit and hybrid environments, particularly where residual air heat remains, though they should be evaluated against the thermal requirements of the specific rack architecture. High-density rack-scale systems may still require direct-to-chip cooling even when rear-door heat exchangers are used elsewhere in the hall.
Facility-side chillers and heat rejection complete the thermal chain. Their design is especially important in India because high ambient temperatures reduce cooling efficiency and limit chiller-less operation. Water treatment also requires careful engineering. A closed liquid loop serving IT equipment can recirculate coolants without significant ongoing consumption, while the upstream heat-rejection system may use dry, adiabatic, evaporative, or hybrid methods, each with a different energy and water profile.
Rack cooling and facility cooling therefore need to be engineered together. Direct-to-chip accessories, CDUs, HDUs, rear-door heat exchangers, chillers, and heat rejection all influence the final combination of rack density, PUE, water use, and availability.
Bid Readiness Requires More Than Installed Capacity
IndiaAI’s PUE, availability, and reporting requirements increase the importance of resilience and measurement. Design specifications should address CDU pump and control redundancy, isolation, leak detection, alarm sequences, and recovery procedures. An efficient thermal design can still create availability risk if a critical cooling component becomes a single point of failure.
Coolant management needs defined ownership. Specifications should cover coolant type and concentration, filtration, chemistry, material compatibility, and responsibility for maintaining fluid quality through operations. Telemetry supports the quarterly reporting requirement. Operators need consistent measurement boundaries and visibility into IT load, cooling energy, CDU thermal capacity, temperatures, and flow.
Commissioning should validate the complete thermal system. Factory acceptance testing, flushing and cleanliness checks, pressure and leak testing, integrated controls testing, and staged-load testing can identify problems before production workloads are introduced.
Clear ownership across the ecosystem is equally important. Server OEMs, rack integrators, cooling suppliers, MEP contractors, controls teams, commissioning agents, and facility operators all influence system performance. Compute requirements must be translated into defined thermal operating conditions. Mechanical contractors need to ensure cleanliness, hydraulic balance, and installation quality. Commissioning teams need to verify performance from the cold plate through facility heat rejection.
Implications for Indian Data Centre Operators
For operators evaluating a GPU data centre in India, design and procurement requirements often boil down to practical questions about cooling architecture, efficiency, and deployment readiness. Air cooling can still suit some deployments, but higher-density GPU racks may require direct-to-chip liquid cooling or a hybrid architecture combining liquid cooling with residual air cooling.
Liquid cooling can reduce energy used for server fans, air movement, and mechanical cooling, thereby lowering PUE. Actual improvement depends on rack density, liquid heat capture, local climate, operating temperatures, and facility-side cooling design. Before deploying high-density GPU racks, operators should confirm rack-level cooling requirements, CDU capacity and redundancy, coolant quality and leak detection, residual air-cooling needs, facility-water connections, heat-rejection capacity, telemetry, commissioning procedures, and the local climate and utility conditions that will shape sustained operation.
Cooling readiness is becoming part of sovereign-compute readiness. Dense GPU architectures are increasing thermal requirements in India’s procurement environment, and for a GPU data centre in India, evaluating cooling alongside the compute platform is essential. Chip-level heat capture, coolant distribution, facility
MegaWatt Intelligence Desk
Hyperscale & Grid Analyst. `More stories by $MegaWatt Intelligence Desk →`
