Liquid cooling = 12× gains
Dell’s new GB300 liquid-cooled platform promises up to 12× performance gains for certain AI workloads, a sign hyperscalers and enterprises are adopting more aggressive thermal solutions to squeeze greater throughput from existing chips. That kind of platform-level improvement can shift procurement decisions — and alter where buyers prefer to deploy the latest GPUs. (x.com) (x.com)
A modern artificial intelligence rack is now so dense that fans alone are starting to look like window air conditioners trying to cool a steel mill. NVIDIA’s GB300 NVL72 packs 72 Blackwell Ultra graphics processors and 36 Grace central processors into one liquid-cooled rack, and Dell’s PowerEdge XE9712 is built around that design. (nvidia.com) (delltechnologies.com) Liquid cooling works like running cold water pipes past the hottest parts of the machine instead of blowing room air over them. Lenovo’s GB300 datasheet says the rack uses warm-water liquid cooling across processors, switches, and power hardware, with airflow handling the rest, so data centers can remove most of the heat without chilled water. (lenovopress.lenovo.com) (nvidia.com) That matters because newer artificial intelligence jobs are not just “train once, answer fast.” NVIDIA says GB300 is tuned for reasoning and test-time scaling, which means the model spends extra compute while answering, like a student showing more work on a harder exam. (nvidia.com) Once you let a model think longer, heat becomes a performance limit instead of a maintenance detail. NVIDIA says GB300 delivers 2x higher attention performance and 1.5x more dense floating-point compute than Blackwell, and Dell says direct liquid cooling lets the XE9712 run higher rack power density while cutting cooling and energy costs. (nvidia.com) (delltechnologies.com) Dell’s pitch is not a single chip but a whole prebuilt rack. The XE9712 ties those 72 graphics processors together with NVLink, NVIDIA’s high-speed lane system, and Dell says the rack can move data between processors at 1.8 terabytes per second inside one unified pool. (delltechnologies.com) That is why vendors keep talking about “output,” not just raw chip speed. NVIDIA says a GB300 NVL72 artificial intelligence factory can deliver up to 50x more overall output than Hopper-based systems, including a 10x boost in responsiveness per user and 5x better throughput per megawatt in one reasoning benchmark. (nvidia.com) Dell has been moving this from slide deck to shipping hardware for months. On July 3, 2025, Dell said it became the first to ship an NVIDIA GB300 NVL72 system, sending liquid-cooled integrated racks with XE9712 servers to CoreWeave for its artificial intelligence cloud. (dell.com) CoreWeave is the kind of customer that makes the procurement story obvious. Dell said that first rack delivered more than one exaflop of dense artificial intelligence performance and up to 40 terabytes of fast memory in a single rack, which is the sort of density that can push a buyer toward a provider that already has liquid loops, power shelves, and service teams in place. (dell.com) Dell is also aiming this at companies that want their own gear, not just rented cloud time. On March 16, 2026, Dell said more than 4,000 customers were already deploying its Dell AI Factory with NVIDIA, and it framed on-premises builds as a response to companies bringing more artificial intelligence work in-house. (dell.com) The “up to 12x” style claims around these systems usually come from specific workloads, not every task in the building, so buyers will still benchmark their own models before signing nine-figure orders. But the direction is clear: when the bottleneck shifts from chip design to how fast you can move heat and power through a rack, cooling stops being plumbing and starts being product strategy. (nvidia.com) (delltechnologies.com)
