Maximizing AI Factory Performance per Watt with NVIDIA DSX MaxLPS
Aug 24, 2026, 8:00 AM · NVIDIA Developer
DSX MaxLPS treats stranded rack power and cooling overhead as capacity, aiming for more inference output without a larger grid connection.
Why it matters
NVIDIA’s 100 MW example leaves 60 MW for AI load after overhead, rack losses and operational inefficiency. Dynamic Power Software reallocates unused rack headroom, while workload profiles tune power and frequency for different phases.
The post reports 170 kW reclaimed in a 540 kW example and projects up to 40% more Rubin capacity in 100 MW. It links 45°C liquid-cooling inlet design to more free cooling.
From the desk
This is NVIDIA’s own technical blog, so projections and validation results are vendor-reported. Dynamic Power Software and DSX Exchange are Developer Preview.
Signal Desk's read: MaxLPS is a facilities-control story disguised as GPU efficiency. Coordinating power and cooling can create real gains, but synchronized bursts must not break latency, redundancy or policy.
The lifecycle idea is strong: build for eventual capacity, populate gradually as workloads shift. It requires planning before the software is proven.
Context
The source distinguishes technology cooling from facility cooling and says 45°C operation does not require agentic control; predictive control is optional.
Who feels it
- Data centers
- Telemetry and cooling design become prerequisites.
- Finance
- More capacity per megawatt may improve returns.
- Reliability
- Emergency policies need testing.
What to watch
- Developer Preview validation
- Annual chiller savings at sites
- DPS behavior during failures
Companies: NVIDIA