NVIDIA Opens DSX Ready Power and Cooling Qualification as GPU Scarcity Shifts the AI Bottleneck to Racks
NVIDIA launched a DSX Ready qualification program covering battery energy storage and cooling distribution units, formalising the power and thermal supply chain for rack-scale AI factories while GPU availability remains tight for smaller buyers.
NVIDIA announced a new qualification program called DSX Ready covering two classes of data center hardware it does not supply itself: battery energy storage systems and cooling distribution units, ServeTheHome reported on September 22. The program gives customers clear guidance on which ancillary parts are confirmed to work with NVIDIA DSX racks, extending the company's existing 'NVIDIA product qualified' label for CDUs into a broader, more heavily branded certification tied to the DGX AI factory blueprint. NVIDIA indicated it expects to add further component categories over time.
The qualification paths differ by component type. CDU vendors may self-qualify against NVIDIA's functional and performance specifications, though they still need final NVIDIA approval, while BESS vendors face a more involved process that requires running qualification tests and submitting the data for NVIDIA to review and sign off. At launch, LG Electronics, LiquidStack and Vertiv are the first qualified CDU vendors, and Hitachi Energy, LG Energy Solution and Tesla are the first qualified BESS vendors. NVIDIA has not disclosed what, if anything, it charges for participation.
The commercial logic is that deployment speed at rack scale is increasingly gated by power delivery and heat rejection rather than by chip availability at the largest buyers. Formalising a vendor list for CDUs and batteries shortens the integration cycle for DGX-class racks, turning what has been a bespoke engineering exercise into something closer to a turn-key configuration. It also gives NVIDIA a lever over a layer of the stack that sits outside its own bill of materials but determines how fast customers can energise new capacity.
At the same time, accelerator supply remains tight for everyone outside the hyperscale tier. DIGITIMES reported on September 29 that South Korea's AI sector is facing a growing GPU shortage, with research institutes and smaller companies struggling to secure enough compute for model training and deployment as government support ends. Suppliers are responding with higher rental pricing and long-term contract demands, a pattern that shifts bargaining power away from smaller buyers even as the largest operators move their constraint from silicon to facilities.