NVIDIA to Debut Fully Liquid-Cooled GB300 AI Servers at GTC 2025, Marking Next Step for AI Infrastructure
NVIDIA is set to revolutionize AI computing once again by announcing its cutting-edge GB300 AI servers, based on the new "Blackwell Ultra" architecture, at the upcoming GTC 2025 conference. These next-generation servers will introduce significantly enhanced performance and will require a major upgrade to fully liquid-cooled solutions.
According to a recent report from Taiwan Economic Daily, NVIDIA's GB300 servers will feature a substantially higher TDP (Thermal Design Power) than the current GB200 series. To manage the considerable increase in heat dissipation, NVIDIA is transitioning completely away from traditional air cooling to implement full liquid-cooling solutions, a move expected to quadruple cooling efficiency.
Given the significantly greater power demands—up to a massive 1400W per GB300 server—the shift to liquid cooling isn't just a luxury, but a necessity. These cooling enhancements allow NVIDIA to push Blackwell Ultra to new performance heights, delivering around 1.4 times the FP4 performance of previous-generation GB200 servers. Additionally, the GB300 will benefit from increased memory capacity, up from 192 GB to 288 GB of HBM3E memory through advanced 12-Hi stacks.
Taiwan's liquid-cooling component supply chains have already seen surging demand, indicating robust market anticipation for NVIDIA’s next-gen AI clusters. As liquid cooling is notably more expensive than air-based solutions, it’s predicted that the pricing of GB300 servers will exceed the existing GB200 NVL72 systems, currently retailing around $3 million each. Despite these higher costs, the market expects robust adoption due to the unprecedented performance gains.
NVIDIA isn't stopping at Blackwell Ultra; it also plans to share updates on the anticipated Vera Rubin series at GTC 2025, although its market debut remains further off.
Stay tuned for the official unveiling during GTC 2025, as NVIDIA sets the stage for yet another leap in AI computing infrastructure.
What are your thoughts on NVIDIA's push towards fully liquid-cooled AI solutions? Let us know below!
