NVIDIA GPUs push 1,200W power draw

NVIDIA’s newest AI hardware is not just faster. It is physically harder to host. That’s the real story here. Blackwell-era systems are pushing power and cooling requirements into territory that looks less like a normal server refresh and more like industrial infrastructure planning. The chips still matter, obviously — but now the room around the chips matters almost as much. (docs.nvidia.com) ### What actually got bigger? The jump is easiest to see in NVIDIA’s own system docs. HGX B200 is configurable up to 1 kW per GPU. In an eight-GPU DGX B200 box, NVIDIA’s deployment guidance lists estimated system power at 14.3 kW max. DGX B300 — the newer Blackwell Ultra system with eight B300 GPUs — is listed at 14.5 kW. So yes, the headline number is huge, but the more important point is that the whole server has become a high-density power appliance. (docs.nvidia.com) ### Why does that change the data center? Because a GPU server does not live alone. It sits in a rack with networking, storage, power conversion, and neighboring servers all competing for the same electrical and thermal budget. NVIDIA’s B200 best-practices guide says a scalable unit of up to 32 DGX B200 systems reaches about 458 kW. That is half a megawatt f(docs.nvidia.com) the GPUs?” and becomes “can my facility feed and cool them?” (docs.nvidia.com) ### Why is liquid cooling showing up everywhere? Air still works for some B200 deployments — DGX B200 is explicitly air-cooled. But rack-scale Blackwell systems have already crossed into liquid territory. GB200 NVL72 is described by NVIDIA as a liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs. GB300 NVL72 is fully liquid-cooled too, again with (docs.nvidia.com)at with fans alone stops being the elegant option. (nvidia.com) ### So is the problem the GPU or the rack? The rack. Or really, the whole power chain. NVIDIA’s DGX B300 SuperPOD reference architecture says a B300 rack uses four power shelves, each able to deliver up to 33 kW. That design choice tells you what changed. Vendors are no longer optimizing only the server motherboard. They are redesigning busbars, rack power distribution, and cooling layouts so the cluster can physically exist at useful density. (docs.nvidia.com) ### Does memory still matter? Yes — maybe more than ever. B300 increases memory per GPU to 288 GB HBM3E, versus 180 GB on B200, and pushes node memory to 2.3 TB across eight GPUs. GB300 NVL72 also leans on 1.5x larger HBM3E memory than the prior Blackwell generation. That is great for reasoning models and larger context windows(docs.nvidia.com)p if the rest of the system can keep them fed. (docs.nvidia.com) ### Why does this matter beyond NVIDIA customers? Because AI infrastructure decisions now lock in building-level choices. If a cloud provider designs around air-cooled 10 to 15 kW servers, that is one path. If it designs around liquid-cooled rack-scale systems and 30 kW-plus rack power components, that is another. Those choices affect retrofit costs, depl(docs.nvidia.com)ce — it is the full stack from silicon to facility engineering. (docs.nvidia.com) ### What’s the bottom line? The scary number is not “1,200W GPU.” The scary number is what happens when you multiply that across eight GPUs, then across 32 servers, then across rows of racks. NVIDIA’s Blackwell generation is turning AI compute into a power-and-cooling game. The winners will be the companies that can deploy the whole machine, not just buy the chip. (docs.nvidia.com)