At the recent CERAWeek 2026 energy summit, NVIDIAAnnounceIn collaboration with Emerald AI and six major energy companies, Google launched a new "Flexible AI Factory" architecture, claiming to release up to 100GW of grid flexibility capacity. At the same time, Google also announced that its "Demand Response" contracts with power companies for its data centers have officially surpassed 1GW, signifying that data centers are transforming from simple "electricity consumers" into "intelligent assets" capable of stabilizing the grid and even feeding back electricity.
NVIDIA Vera Rubin DSX: Enabling "Grid Sensing" Capabilities in Computing Power
The core technology NVIDIA launched this time is an AI factory reference design called Vera Rubin DSX. This architecture has a built-in DSX Flex software library, and its biggest function is to give AI computing systems "grid awareness" capabilities.
In short, through Emerald AI's Conductor platform scheduling, AI factories can dynamically adjust computing power based on grid load:
• Grid stress peaks:Automatically reducing non-urgent AI training tasks (such as long-term model optimization) will feed the saved electricity back to the community power grid.
• Ample power supply:Improve computing power utilization and accelerate the production of AI Tokens.
The alliance with NVIDIA this time includes six major U.S. energy suppliers, such as AES, Constellation, and NextEra Energy. The project is expected to conduct its first commercial-scale validation this year at an AI research center in Virginia, aiming to solve the most pressing problem in current AI deployments: excessively long grid connection waiting times, allowing data centers to obtain power permits more quickly.
Google's Real-World Performance Report: 1GW Demand Response Contract Officially Secured
While NVIDIA outlines its future blueprint, Google is delivering solid results in practice. Michael Terrell, Google's head of advanced energy, pointed out that Google has signed contracts with several power companies, including TVA and DTE Energy, to integrate a total of 1GW of data center load into the demand response system.
Google's approach involves using machine learning to accurately predict when non-real-time computing tasks can be offloaded. This "demand elasticity" not only helps power companies stabilize peak loads but also shortens the infrastructure expansion time required to launch new data centers.
Google emphasizes that this flexibility can effectively reduce the infrastructure burden on power companies, ultimately reflected in electricity bill reductions for all users.
A strategic shift from "co-located generation" to "feedback grid"
Over the past year, due to aging power grids in some parts of the United States, many large AI campuses have been forced to adopt a "co-location" strategy, building their data centers directly next to nuclear power plants or large solar farms to obtain electricity.
However, NVIDIA and energy industry players agree that isolating these power generation facilities from the main grid is not a long-term solution. Constellation CEO Joe Dominguez points out that the current power crisis is essentially a "peak problem."
Through a flexible architecture, AI factories can utilize on-site power generation equipment to maintain operations initially (transitional power), and then become grid assets that can be dispatched at any time after grid connection is completed. Nscale's Monarch campus in West Virginia even plans to expand its power capacity from 2GW to 8GW, aiming to become a "super power supplier" for the grid.
Analysis of viewpoints
NVIDIA's high-profile move at CERAWeek is essentially about defining the "infrastructure specification rights" for the AI era.
In the past, when we discussed AI laptops or mobile phones, we focused on the TOPS computing power of the NPU; however, at the level of massive data centers, power efficiency (PUE) and grid interaction capabilities are the key factors determining operating costs. NVIDIA is trying to tell energy companies through the Vera Rubin DSX architecture that instead of viewing AI data centers as a threat to the grid, they should be viewed as a super-large "virtual battery" or "controllable load".
For Google, achieving the 1GW milestone proves that the training tasks of large-scale language models are indeed highly schedulable. This "blurring of the boundary between computing power and electricity" foreshadows that the future battleground for tech giants will extend from software code to high-voltage power towers. When AI factories can reduce the marginal cost per kilowatt-hour by adjusting computing power, and even profit from "saving electricity" or "discharging" during peak electricity prices, the competitiveness of AI will no longer be just about the quality of algorithms, but also a demonstration of energy dispatching capabilities.
