How Does an AI Factory Drop Power From Four Megawatts to Three Without Stopping Work?

On a hot August evening in Silicon Valley, a utility sent a signal and an AI factory cut its draw from four megawatts to three in under a minute. No operator touched a switch, and every high-priority job kept running.

The Night the Grid Called

Varun Sivaram watched on Zoom with about forty others. His team at Emerald AI sat in a San Francisco conference room. Engineers at the data center and people from the utility itself filled the rest of the tiles. Silicon Valley Power had just sent a demand signal. The factory responded automatically. Power fell from four megawatts to three. Everyone cheered.

Sivaram said they watched with bated breath. It was their first time deploying across thousands of NVIDIA GPUs. His head of product, Mansi Shah, compared the feeling to a SpaceX rocket launch.

Silicon Valley Power has since sent more than 200 demand signals to that same factory. It worked every single time.

What the Software Actually Does

Emerald AI Conductor receives grid signals — load-shedding requests, demand-response events, pricing changes — and rewrites workload priorities in response. Lowest-priority jobs yield. High-priority inference keeps running. The factory that agrees to flex gets to run bigger. That pattern is what NVIDIA DSX Flex is built to generalize. Conductor will integrate into DSX Flex as the platform matures.

NVIDIA's Eos AI factory runs Conductor as a participant in Silicon Valley Power's Flexible Load Interconnect Program. That program is the first commercial utility program designed to treat AI factories as dispatchable resources. Conductor responds in under a minute.

The first dedicated DSX Flex commercial deployment will be a 96-megawatt Vera Rubin AI factory at NVIDIA's AI Factory Research Center in Manassas, Virginia. It builds on five prior demonstrations across two continents.

Lambda's Numbers on the Same Hardware

Lambda, a GPU cloud provider serving more than 10,000 customers, ran the first validation of DSX MaxLPS on NVIDIA HGX B200 GPU Servers. They used a five-rack, 19-node cluster. By running 19 nodes within the same power budget as 16 nodes at full power, Lambda achieved 24 percent more cluster-wide token throughput — from roughly 4 million tokens per second to 5 million. Performance per watt improved by 23 percent.

DSX MaxLPS monitors GPU and rack-level power consumption in real time. It reallocates headroom across nodes based on workload type. Training and inference draw power differently; MaxLPS optimizes allocation in factories running both.

Based on NVIDIA's projections, DSX MaxLPS can enable up to 40 percent more GPU capacity for next-generation Vera Rubin NVL72 AI factories within the same megawatt power budget in suitable deployment environments.

The Next Layer Is Voltage

As AI factories scale, traditional lower-voltage power paths add conversion complexity and distribution constraints. NVIDIA's 800 VDC architecture is designed to reduce conversion complexity, improve power delivery efficiency, and support denser accelerated computing racks. DSX is incorporating 800V DC into its reference designs. The projected gain is 3 to 5 percent end-to-end efficiency versus 54V distribution, available with Vera Rubin NVL72 in 2027.

GB200 NVL72 racks running direct liquid cooling carry roughly 120 kilowatts of heat that has to go somewhere before that power reaches compute. Cooling overhead still diverts electricity from GPUs. A faster GPU still waits on the network. Power can be stranded by bad provisioning.

The only reliable path to more tokens per megawatt is to optimize the whole factory. DSX Sim before the first rack goes in. DSX OS and DSX Exchange once it's running. DSX Reference Designs so builders start from a validated architecture rather than from scratch.

Why the Whole Factory Must Move Together

Jensen Huang has said a one-gigawatt factory will never become a two-gigawatt factory. The answer engineers reach when systems hit physical limits is always the same: stop optimizing the parts and start designing the whole. NVIDIA DSX was introduced at GTC Taipei in May as that answer for the AI factory. Early deployments are already proving it out.

When the grid needed relief, the factory gave it without dropping a job and without asking for more power. With NVIDIA DSX, that is the new baseline for what an AI factory is supposed to do.