<\/p>\n\n\n\n
The data center industry is undergoing a fundamental transformation. As Artificial Intelligence and High-Performance Computing (HPC) applications proliferate, traditional air-cooled infrastructure is reaching its physical limits. Enter High Density (HD) colocation, a paradigm shift that delivers power densities of up to 150kW+ per cabinet, around ten times that of conventional deployments.<\/p>\n\n\n\n
At nLighten, we’ve moved beyond theory to practice, deploying HD solutions across Europe and developing standardised services that make this advanced infrastructure accessible at scale. Here’s what you need to know about this critical evolution in data center capabilities.<\/p>\n\n\n\n
The Physics of the Problem<\/strong><\/p>\n\n\n\n Modern AI workloads rely on Graphics Processing Units (GPUs), predominantly from NVIDIA, that are optimised for the parallel processing required by machine learning algorithms. A single GPU-based server can consume 10kW or more, compared to 1-2kW for traditional CPU-based systems. Pack multiple servers into a cabinet, and you’re suddenly dealing with heat loads that overwhelm conventional air cooling approaches.<\/p>\n\n\n\n The solution lies in liquid cooling. Liquid conducts heat approximately 3,000 times more effectively than air, making it the only viable path forward for high-density deployments. By extending chilled water systems directly to customer equipment through specialised heat exchangers, data centers can support these high power densities while improving overall efficiency and sustainability.<\/p>\n\n\n\n The Market Reality<\/strong><\/p>\n\n\n\n The numbers driving HD adoption are staggering. According to McKinsey, approximately 70% of new data center demand growth is concentrated on AI-capable infrastructure. Goldman Sachs projects that the five highest-spending US hyperscale enterprises will invest a combined $736 billion in capital expenditure in 2025 and 2026 alone.<\/p>\n\n\n\n Major technology providers such as Amazon, Google, Microsoft, Meta, and emerging players like OpenAI and Anthropic, are racing to deploy AI infrastructure. Amazon added 3.8 gigawatts of capacity in a single year, with capital expenditure of $125 billion in 2025. Meta’s AI infrastructure spending reached $70 billion in 2025, with “notably larger” investments planned for 2026.<\/p>\n\n\n\n But it’s not just the hyperscalers. Research institutions, financial services firms, healthcare organisations, and enterprises across sectors are deploying AI for applications ranging from drug discovery to fraud detection to predictive maintenance. Each requires the computational density that only modern HD infrastructure can provide.<\/p>\n\n\n\n From Training to Inference: Different Workloads, Different Needs<\/strong><\/p>\n\n\n\n An important distinction exists between AI training and AI inference workloads. AI training, the process of building foundation models, requires massive clusters consuming multiple megawatts. AI inference, deploying those models to generate real-time results, typically requires smaller deployments closer to end users.<\/p>\n\n\n\n This is where nLighten’s distributed European footprint becomes particularly valuable. Our data centers sit in proximity to major population and business centers, delivering the low latency essential for inference applications. As enterprises move from experimentation to production AI deployment, this edge advantage grows increasingly critical.<\/p>\n\n\n\n Real-World Implementation<\/strong><\/p>\n\n\n\n nLighten already operates production HD deployments across our portfolio. In Germany, our first direct liquid cooling installation serves a public research institution with a deployment combining GPU and CPU-based servers, network equipment, and storage, all with different cooling requirements addressed through a unified infrastructure approach that can support up to 100 kW per cabinet.<\/p>\n\n\n\n In the UK, a 3MW deployment using Rear Door Heat Exchangers (RDHx) supports a GPU-as-a-Service provider, with 72 cabinets each consuming 42kW.<\/p>\n\n\n\n These projects have taught us valuable lessons about the operational realities of HD colocation: the importance of commissioning processes, the need for flexible demarcation points between provider and customer responsibilities, and the value of modular approaches that allow customers to start small and scale as requirements evolve.<\/p>\n\n\n\n