NVIDIA Launches DSX Ready to Qualify BESS and Cooling Units as AI Factory Power Becomes the Limit

NVIDIA used the week of September 21 to launch DSX Ready, a qualification programme for third-party power and cooling hardware that meets its DSX AI factory reference design requirements, according to ServeTheHome reporting of September 22. The first two categories are battery energy storage systems and coolant distribution units, the two classes of equipment NVIDIA does not supply itself. Qualified BESS vendors at launch are Hitachi Energy, LG Energy Solution and Tesla; qualified CDU vendors are LG Electronics, LiquidStack and Vertiv.

The two categories are qualified differently. CDU vendors can self-qualify against NVIDIA's functional and performance specifications subject to final NVIDIA approval, while BESS vendors must run qualification tests and submit data for review, reflecting the tighter requirements of electrical work. NVIDIA has not disclosed what it charges for participation, and the company stresses the designation applies to a specific product rather than a supplier's whole portfolio. The programme extends the existing NVIDIA product-qualified label for CDUs and is intended to make DGX rack installation closer to turn-key.

The commercial logic is that power and thermal capacity, not chip supply, now set the ceiling on deployment. Coverage of the launch notes NVIDIA's DSX MaxLPS reduces per-rack power requirements and enables up to 40% more GPU capacity within the same facility envelope, with performance-per-watt improvements of 1.3x to 1.5x contingent on optimised infrastructure design, workload profiling and early site-level engagement for 45 degrees Celsius liquid cooling. NVIDIA projects that 100GW of AI factory capacity will come online globally before the end of the decade, which makes the facility power budget a primary constraint on how much compute operators can actually deploy. Suppliers outside the programme risk being excluded from reference-design bids as builders use the catalogue to shorten vendor evaluation.

Interconnect is the other half of the same constraint. GlobalFoundries used the China International Optoelectronics Expo to push into silicon photonics, presenting a palm-sized metal transceiver as emblematic of the shift from copper to light in data centre infrastructure, DIGITIMES reported on September 16. NVIDIA is separately shipping its own MMS1X00-N5400 QSFP112 1310nm 400Gbps optic, reviewed by ServeTheHome on September 20. The sequencing matters for anyone budgeting AI hardware: power and cooling qualification is being industrialised now, while the optical transition inside the rack is still at the standards and early-adoption stage.