Bloomberg: AI data centers need more

The easy version of the AI boom is: buy Nvidia chips, plug them in, print intelligence. But that was never really true, and it is getting less true with every new rack generation. The hard part now is the rest of the machine — the plumbing, wiring, voltage conversion, networking, and heat removal that let those chips run flat out without melting down. Bloomberg’s story on May 12 lands on that gap: the bottleneck is spreading outward from GPUs into the whole data-center stack. ### What changed? The shift is tied to the move from simpler GPU clusters to denser AI racks built around systems like Nvidia’s GB200 NVL72. Nvidia pitches that platform as a liquid-cooled rack-scale design with 72 Blackwell GPUs, fifth-generation NVLink, and much tighter GPU-to-GPU communication. That is great for performance. But it also means the surrounding infrastructure stops being generic. Cooling loops, power distribution, and rack integration become custom, high-stakes parts of the build. (bloomberg.com) ### Why does density make everything harder? Because watts turn into heat, and AI racks now pack an enormous amount of both into a small footprint. Older data centers could lean more heavily on air cooling and standard power layouts. New AI systems increasingly need direct-to-chip liquid cooling, coolant distribution units, and heavier-duty electrical gear close to the rack. Once you cross that density threshold, the data center starts to look less like a room full of servers and more like an industrial plant. (nvidia.com) ### What are the missing pieces? They are not glamorous, which is exactly the point. You need interconnect hardware to move data between accelerators at very high speed. You need power shelves, busbars, transformers, backup systems, and voltage-conversion gear to feed unstable, bursty loads. And you need thermal hardware — cold plates, manifolds, pumps, CDUs, and control systems — to carry heat away continuously. Bloomberg’s point is that a shortage or delay in any one of those can stall the whole deployment, even if the GPUs are already spoken for. (nvidia.com) ### Why are power parts suddenly such a big deal? AI servers do not draw power like ordinary enterprise boxes. They spike hard, run hot, and increasingly push the industry toward higher-voltage architectures. Delta, for example, has been showing 800V HVDC designs, 90kW power shelves, and rack-level capacitance systems aimed specifically at GPU-heavy deployments. Vertiv has been pushing full reference architectures for Nvidia systems because customers do not just need components — they need a blueprint that actually works end to end. (bloomberg.com) ### Why can’t builders just swap in standard cooling? Because Blackwell-class systems are designed around liquid from the start. Nvidia explicitly says liquid cooling helps raise compute density and reduce floor-space needs in GB200 deployments. Supermicro says its GB200 systems rely on direct-to-chip liquid cooling. In other words, cooling is no longer an optional optimization. It is part of the product architecture. (delta-americas.com) ### Who benefits from this? Not just Nvidia. Suppliers that handle the “boring” layers — Vertiv, Delta, and other power-and-thermal specialists — become much more strategic when the deployment problem shifts from chip procurement to systems integration. That is why investors have spent so much time on infrastructure names tied to AI buildouts. The spend is broadening from silicon into everything needed to make silicon usable at scale. (nvidia.com) ### What is the real takeaway? The AI race is becoming a construction and utilities story as much as a semiconductor story. GPUs still matter most at the center. But around that center sits a growing ring of constraints — electrical, thermal, and mechanical. Basically, the next wave of AI capacity will be won not only by whoever gets the chips, but by whoever can actually make the building behave like a machine. (bloomberg.com) (vertiv.com)