EXECUTIVE SUMMARY:
- By 2027, it is estimated that a single rack of NVIDIA GPUs will need nearly 1 MW of power density to fire up – and with such growth, an inevitable bottleneck is emerging in current cooling infrastructure
- As liquid cooling emerges to become the de facto solution for data centres, operators will increasingly demand easy integration
- Predictability, repeatability, and scalability is key, particularly when designing or retrofitting existing infrastructure
- LiquidStack’s GigaModular CDU solution is built for scalable design, up to potentially 10 MW, with operators reporting 25% savings in capex and floorspace
With the advent of commercial Gen AI models, progress has been both drastic and wide-reaching in its applications. However, industry insiders are more and more concerned about the rapid growth in infrastructure requirements and the enormous power demands that accompany these developments.
Prompts that appear simple in nature trigger a ‘butterfly effect’ of computational processes and data transfers, consuming electricity at every stage to generate a response; primarily on NVIDIA GPU processors.
New GPU alternatives with lower power consumption are entering the market; however, within current data centre ecosystems, NVIDIA remains dominant due to its superior architecture for AI workloads. This defines the thermal profile of modern data centres, where, without advanced cooling protocols, GPUs cannot reach peak performance or density.
Current estimates from the International Energy Agency predict that data centre consumption will reach 1,000 TWh by 2030; more than double the 2024 rate. Energy use is growing 12 percent year on year, with data centres now accounting for 1.5 percent of total global consumption.
Computational power is only one part of this energy-intensive journey. For every kilowatt powering these chipsets, ABI Research reports that 37 percent of total energy consumption is dedicated solely to cooling.
By 2027, powering a single rack of NVIDIA GPUs will require nearly 1 MW of power density; a capacity once associated with entire flagship facilities. Today, facilities demand hundreds of megawatts to operate, and ABI predicts that the number of public data centres will quadruple by 2030.
With such explosive growth, an inevitable bottleneck is emerging in current cooling infrastructure due to thermal management constraints, requiring industry leaders to drive scaling, sustain growth, and move the industry forward. While computational efficiency for chipsets has improved year after year, large-scale cooling remains a persistent challenge. Traditional air cooling can no longer efficiently handle the heat densities produced by next-generation AI infrastructure.
Cooling advancements are a clear necessity; not only to accommodate existing infrastructure but also to shape the future capabilities of data centre hubs. This is where liquid cooling becomes essential, with the Uptime Institute reporting that 22 percent of organisations are already using some form of direct liquid cooling (DLC).
Liquid cooling is not a foreign technology; it is already used within the data centre industry. However, outside GPU farms and hyperscale environments, it requires custom-designed technology, which is expensive, difficult to distribute, and not easily scalable.
Liquid cooling for all?
As liquid cooling moves toward becoming the de facto solution for data centres, operators will increasingly demand easy integration. Designing or retrofitting existing infrastructure will require liquid cooling solutions that are predictable, repeatable, and scalable.
Cooling systems must keep pace with AI scaling without requiring custom designs or complex management, while remaining easy to service.
Across industries; from hyperscalers supporting SaaS platforms to enterprise data centres managing financial systems, continuous uptime is essential for data centre reliability and investment.
Cooling, Platformised
LiquidStack, a provider of next-generation cooling solutions, develops technologies tailored to the specific needs of data centres. Their solutions include two-phase liquid immersion; a technically demanding form of thermal management and expand across the full spectrum of liquid cooling, with particular focus on CDUs (coolant distribution units) for DLC systems.
Their most recent solution, the GigaModular CDU, is built for scalable design. It is a single-phase DLC platform that scales from 2.5 MW to 10 MW, with centralised control, modular pump architecture, and full front accessibility for simplified servicing and flexible placement.
Operators implementing these platforms have reported 25 percent savings in both capex and floorspace; providing a critical advantage when rapidly deploying solutions or retrofitting legacy environments. The “pay-as-you-grow†model further aligns capital expenditure with capacity expansion.
Scaling efforts extend beyond the rack, incorporating built-in resilience measures to address supply chain unreliability for global operators. Two factories are currently operational in the US, with active plans to expand manufacturing capacity worldwide. The global service network already provides consistent SLAs across regions.
Data centre operators cannot accelerate chip development without equally serious investment in cooling capacity.
LiquidStack offers a comprehensive platform solution that is scalable, serviceable, and globally deployable; critical for the future of data centre infrastructure.
