Chapter 7 · Power & CoolingGTC Taipei (DSX Flex and “DSX Max-Q” first named at GTC, March 2026)
NVIDIA DSX MaxLPS and DSX Flex: more work per megawatt, and power that follows the grid
NVIDIA
Two pieces of power software. MaxLPS moves unused power headroom between racks so more GPUs fit in a fixed power budget; Flex lowers an AI factory’s draw when the grid asks it to.
What was announced?
Both are part of NVIDIA’s DSX platform, announced on 31 May 2026. NVIDIA describes MaxLPS as technologies to “maximize token performance per megawatt within a fixed power budget”; LPS stands for the site’s land, power and shell (the building). NVIDIA’s March 2026 release used the name “DSX Max-Q” for what appears to be the same library.
NVIDIA describes DSX Flex as “grid-responsive power orchestration for AI factories”: it adjusts workload power in response to grid signals such as demand-response and pricing events, and coordinates grid power with on-site renewables and storage.
Why does it matter?
Power, not chips, is now the binding constraint on most AI sites: a site gets a fixed grid connection, and each new generation of rack draws more. Conventional planning reserves enough power for every rack to hit its peak at once, so much of the connection sits unused most of the time (“stranded” power).
MaxLPS tries to put that stranded power to work, and Flex tries to make an AI factory a load the grid operator can call on. Flexibility can help a site get connected sooner, because a load that can be reduced on request is easier for a utility to accept.
What has changed?
| Aspect | Conventional | MaxLPS / Flex |
|---|---|---|
| Rack power budget | Each rack provisioned for its peak | Racks capped near their most efficient point; spare headroom moved between racks (Dynamic Power Software) |
| Racks per megawatt | Set by nameplate peak | NVIDIA’s examples: GB200 NVL72 rack 125 kW → 90 kW, “39% more racks” |
| Grid events | Load is constant; backup generation covers emergencies | Workload power reduced on a grid signal while critical jobs keep running |
How does it work?
MaxLPS works at four levels: the GPU (power limits per chip), the rack, the cooling plant and the workload. Its Dynamic Power Software watches actual draw across a group of racks and lends headroom from racks that are not using it to racks that can. A GPU run slightly below its maximum power usually loses much less performance than the power it saves, which is why capping can raise output per megawatt.
Flex sits above that. When a utility signals a peak or a price event, it lowers power on flexible jobs (training can slow down; latency-sensitive inference should not), and brings it back when the event ends. Emerald AI is integrating Flex with its Conductor platform; NVIDIA describes a trial with Emerald AI, EPRI, Nebius and the UK’s National Grid in which a Blackwell Ultra cluster cut its draw by up to 40% during simulated grid events.
Figures and claims, and what backs them
- Manufacturer claimNVIDIA says MaxLPS can produce “the same number of tokens with up to 30% less power” (October 2026).
- Manufacturer claimNVIDIA says operators can run “up to 40% more GPUs” at their most efficient point; its August blog presents this as a projection for Rubin.
- Manufacturer claimNVIDIA reports GB200 NVL72 rack power cut from 125 kW to 90 kW (“39% more racks”) and Vera Rubin NVL72 from 136 kW to 101 kW (“35% more racks”).
- Manufacturer claimNVIDIA quotes Nscale (four GB300 NVL72 racks in 264 kW): “49% higher inference throughput and 37% more managed GPU capacity”.
- Manufacturer claimNVIDIA says a Blackwell Ultra cluster cut power “by up to 40%” in simulated grid events (Emerald AI, EPRI, Nebius, National Grid trial).
Impact on AI infrastructure delivery
- Power & Cooling
- Capacity planning changes from nameplate peak to managed draw. Electrical and cooling designs must still be safe at the true peak, so the gain is in how many racks share a connection, not in smaller cables or plant.
- IT Infrastructure
- Rack counts in a design can rise for the same grid connection, which changes floor space, network ports and procurement quantities.
- Operations
- Power becomes a scheduling policy. Operations must agree with customers which workloads may be slowed or capped, and record it in service terms; a grid event that slows a customer’s job is a commercial question, not only a technical one.
Commercial availability
- Announced (current stage)Described publicly; no date for customers to buy it.
- PlannedA stated availability window that has not arrived yet.
- ShippingIn production and reaching early customers or partners.
- Generally availableOrderable by any customer through normal channels.
- DeployedRunning in customers’ production environments, confirmed by them or the manufacturer.
NVIDIA’s August 2026 blog says Dynamic Power Software is “currently in Developer Preview”. DSX Flex is described only through pilots and trials, including a “commercial, multi-megawatt pilot” with Emerald AI and Silicon Valley Power. NVIDIA’s DSX release says its products “will be offered on a when-and-if-available basis”.
Not yet confirmed
- That “DSX Max-Q” (March 2026) and MaxLPS are the same product is inferred from NVIDIA’s naming; NVIDIA does not state a rename.
- All figures are NVIDIA’s or its partners’, mostly “up to”; none has been measured independently. The size of the Silicon Valley Power pilot is not given.
Learn more in BSS/OSS Academy
Official sources
- NVIDIA DSX Gives Infrastructure Builders the Playbook for AI Factories(opens in a new tab)NVIDIA Newsroom · 31 May 2026 · Manufacturer
- Maximizing AI Factory Performance per Watt with NVIDIA DSX MaxLPS(opens in a new tab)NVIDIA Technical Blog · 24 August 2026 · Technical documentation
- European Ecosystem Builds World’s Most Efficient AI Factories With NVIDIA DSX(opens in a new tab)NVIDIA Newsroom · 21 October 2026 · Manufacturer
- NVIDIA Releases Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Digital Twin Blueprint(opens in a new tab)NVIDIA Newsroom · 16 March 2026 · Manufacturer
Figures are as the manufacturer states them unless marked otherwise. None has been independently tested by BSS/OSS Academy.