On Oct. 21, NVIDIA announced that European partners are adopting its DSX data center platform to expand AI computing capacity within power and cooling constraints.
The Associated Press put Europe’s share of global AI computing capacity at about 5%, compared with roughly 75% for the United States. NVIDIA’s reference design coordinates computing, networking, storage, software, power and cooling. Its DSX MaxLPS software shifts available power among participating equipment within limits set by operators.
NVIDIA says MaxLPS can produce an equal volume of tokens while using up to 30% less power. A separate evaluation with Nscale measured how much more work the system could handle within a fixed power budget.
At Nscale’s Iceland facility, MaxLPS managed 192 graphics processing units, or GPUs, against 140 in a static baseline. Both setups had the same provisioned 264.4-kilowatt power budget. Aggregate output rose 49.2%, from 1,084,503 to 1,618,443 tokens per second, after the test added a third high-throughput inference instance. Output from each existing instance changed little.
Measured power consumption increased from 166.2 to 198.9 kilowatts as the larger GPU fleet ran. The 99th-percentile time to first token rose 17% from a 15.7-second baseline; median and 75th-percentile latency remained within 5% of baseline. The test measured higher total output under a fixed provisioned budget. It did not test the separate claim of 30% less power for an unchanged workload.
For cooling, DSX pairs closed-loop liquid cooling with dry coolers that transfer heat to the air without evaporating water. NVIDIA estimates that a hypothetical two-gigawatt deployment would use 54 million fewer liters of water per day than conventional cooling towers. It says facility cooling-water use could approach zero, although hot conditions may require supplemental mechanical cooling.
Lambda reported that its own proof of concept ran 19 nodes within the power budget used by a 16-node baseline. It said total token throughput increased about 24% and performance per watt improved 23%. NVIDIA also reported that an early DSX Flex trial reduced cluster power draw by up to 40% during simulated grid events while critical workloads continued.
NVIDIA lists several cloud providers as exploring MaxLPS. Three DSX OS components are in developer preview; the company gave no dated rollout milestone.
