NVIDIA is pitching software that links AI servers, cooling, batteries and grid signals so data centres can do more with limited electricity. The most concrete result in The Register’s account is a Lambda test that fitted 19 compute nodes into the power budget normally used by 16, raising cluster throughput by 24 per cent. The catch is that the system also tightens NVIDIA’s grip on the infrastructure stack.
NVIDIA Watch analysis
What happened
NVIDIA’s DSX platform is designed to connect compute hardware with the physical systems around it, including power cabinets, batteries, coolant distribution units, chillers and building-management software. Dion Harris, NVIDIA’s senior director for HPC and AI infrastructure, told The Register that DSX Exchange acts as an API for information from NVIDIA and third-party equipment, including systems from Vertiv and Schneider Electric.
The company’s DSX MaxLPS approach aims to reduce wasted headroom inside data centres. In the Lambda test described by The Register, software coordination allowed 19 nodes to operate within the power envelope normally occupied by 16. Lambda reported 24 per cent higher cluster-wide throughput and a similar improvement in performance per watt. Those are reported test results, not an independent benchmark.
NVIDIA also demonstrated DSX Flex with Emerald AI and Silicon Valley Power. The system can respond when the grid is under pressure by reducing, pausing or moving non-essential workloads while keeping critical work running. NVIDIA says this could give utilities more confidence when allocating scarce capacity, though the demonstration does not establish how often it can operate at scale or what performance users would sacrifice.
Why it matters
AI data centres are increasingly limited not just by chips, but by electricity connections, cooling capacity and local grid constraints. Making workloads flexible could let operators build more computing capacity without treating every facility as a permanently immovable power drain.
For NVIDIA, there is a second prize. DSX works most naturally with validated NVIDIA hardware and software, while support for rival accelerators such as AMD’s Instinct GPUs would require alternative management systems. The efficiency pitch is real-world infrastructure work, but it also looks rather like a very helpful moat with a server rack attached.
Our read
The important shift is from selling faster accelerators to managing the entire environment around them. Better visibility into power and cooling can produce useful gains, but the headline numbers should be treated as NVIDIA and its partners’ claims until independent operators reproduce them across different workloads.
Buyers should ask two questions at once: how much usable compute does DSX add, and how much control over their data centre does it require? A system that makes electricity more productive is valuable. A system that makes switching away harder deserves equally close attention.
What to watch
- Independent results:
Whether operators publish comparable performance-per-watt results outside NVIDIA’s partner demonstrations. - Grid response:
How quickly workloads can be paused or moved, and what happens to latency-sensitive services. - Hardware breadth:
Whether DSX supports competing accelerators with the same level of control. - Commercial terms:
Whether the efficiency gains justify deeper dependence on NVIDIA’s software and validated hardware ecosystem.
Discussion spark: Should data-centre operators accept deeper dependence on NVIDIA if its power-management tools unlock scarce grid capacity, or should utilities require open standards before granting new connections?
Sources and evidence
- Nvidia goes green to keep grid capacity from zapping its revenues – The Register (16 September 2026, 21:37 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.