Google, Nvidia, and Emerald AI Launch Alliance to Make Data Centers Grid Assets

The researcher who mathematically proved that US data centers could unlock roughly 100 gigawatts of grid capacity without building a single new power plant is now chairing the industry coalition formed to make that happen. Tyler Norris, who published that research as a Duke University PhD student in February 2025 and then joined Google as head of advanced energy market innovation — where he closed 1 gigawatt of demand-response contracts with five US utilities — was named inaugural board chair of the AI Energy Management Alliance on Wednesday as Google, Nvidia, and AI grid-orchestration startup Emerald AI launched the new body alongside eighteen launch partners.

The AEMA's stated mission is to standardize demand-response capabilities across the AI data center sector — establishing common practices, policy frameworks, and technical specifications so that data centers routinely function as what the industry calls "flexible loads" rather than fixed, grid-stressing draws. Its launch partners span compute and power: alongside the three founding companies, the coalition includes Anthropic, The AES Corporation, Constellation, Generate Capital, National Grid, NRG Energy, Analog Devices, Calibrant Energy, Camus, ClearPath, Encoord, Fluence, GridUnity, PassKey Inc., RWE, Splight, Verrus, and Voltus. Frank Lacey, an experienced energy industry executive, will serve as AEMA's executive director.

The three companies are not starting from scratch. AEMA builds on the Advanced Energy Management Alliance, originally founded in 2014 to advance energy solutions and grid efficiency — and on a year of commercial-scale demonstrations that Emerald AI and Nvidia conducted at data centers in Phoenix, Chicago, London, and Portland before the coalition's formation.

Read more: Anthropic commits $45B to Nvidia Vera Rubin compute capacity

Google Paired the Launch with Two Policy Blueprints Targeting FERC and ERCOT

Google simultaneously released two commissioned research reports framing actionable regulatory pathways. The first, from The Brattle Group, translates recent bipartisan direction from the Federal Energy Regulatory Commission into operational specifications for utilities and grid operators under the title "Solutions for Timely Interconnection of Large Loads: Enabling Non-Firm Transmission Service." The second, from Aurora Energy Research, models how flexible AI data centers in ERCOT — the Texas grid — could be paired with front-of-meter resources to improve system economics.

The regulatory context matters. On June 18, 2026, FERC issued six tailored show-cause orders to CAISO, ISO New England, MISO, NYISO, PJM, and the Southwest Power Pool, directing each operator to justify or immediately reform its interconnection rules for large loads — or offer flexible loads an expedited 60-day study track. The agency had been directed by DOE Secretary Chris Wright in October 2025 to accelerate the process. In ERCOT alone, 198 gigawatts of new large-load applications arrived in Q1 2026. The policy machinery now exists to reward data centers that commit to flexibility; what has been missing is an industry-wide standard for what that commitment means and how to measure it. The AEMA is designed to fill that gap.

How Emerald Conductor Modulates GPU Workloads Without Breaking SLAs

At the operational center of the coalition is Emerald AI's Conductor platform, which Nvidia has invested in and helped validate at multiple commercial data centers. Conductor is a software-only system — no hardware modification required — that sits between the power grid and the compute stack, operating as what CEO Varun Sivaram describes as "an AI for AI."

The platform works through a closed-loop architecture combining two components: an autonomous agent that makes real-time scheduling decisions, and a digital twin simulator that models the downstream compute consequences of any proposed adjustment before executing it. This allows the system to act quickly without blindly degrading performance.

When a grid operator signals the need to reduce load — for example, during a hot summer afternoon when air conditioning demand peaks — Conductor tags the data center's running workloads by priority and curtailment tolerance, then dispatches one of three types of response. Temporal flexibility means delaying or slowing batch-oriented tasks — AI model training, fine-tuning, non-urgent inference — to a later window. Spatial flexibility means migrating workloads to a geographically different facility in a region where grid stress is lower. Resource flexibility means drawing on on-site batteries or other behind-the-meter generation to avoid pulling from the shared grid entirely.

Real-time inference — requests requiring immediate responses — sits outside what the system can safely curtail without SLA violations, which represents a meaningful portion of production AI workloads. The platform's value is highest for training and batch processing, where flexibility is structurally available.

In a commercial-scale demonstration in Phoenix, Arizona on May 3, 2025, Conductor reduced power consumption by 25 percent over three hours while maintaining AI workload performance — a company-reported result conducted at an Oracle Cloud Infrastructure data center in partnership with Nvidia, Salt River Project, and the Electric Power Research Institute. A later trial in London, conducted through EPRI's DCFlex initiative with National Grid and cloud provider Nebius, cut electricity demand by up to 40 percent while critical workloads continued without interruption — with the system shedding more than 30 percent of its load within 30 seconds of receiving a grid signal. In Chicago, under harder conditions with unknown random workloads, some AI jobs malfunctioned mid-run — and the Conductor system adapted automatically, keeping the cluster stable. The Chicago demonstration is the closest thing available to an adversarial test, and it passed.

The Grid's Structural Problem Makes Software-Based Flexibility Unusually Attractive

The economic logic behind AEMA rests on a mathematical insight from Norris's 2025 study, published through Duke University's Nicholas Institute for Energy, Environment and Sustainability. The US grid is designed to meet rare peak demand events — the hottest summer afternoons — which means it carries significant reserve capacity that sits idle most of the year. Norris and co-authors calculated that if large new loads agreed to curtail their draw for an average of about 44 hours annually — roughly 0.5 percent of uptime — the grid's existing headroom could absorb approximately 98 gigawatts of new large loads without any additional generation construction.

This is not a small number. The International Energy Agency projects that US data center power consumption will represent nearly half of total US electricity demand growth through 2030, with AI as the primary driver. Building that much new generating and transmission capacity fast enough is structurally impossible — interconnection queues in regions like PJM and Virginia already stretch five to ten years. Software-based flexibility, deployable in months, competes against timelines that cannot be shortened by investment alone.

Google has already demonstrated the commercial viability of this model. Its contracts with Indiana Michigan Power, Tennessee Valley Authority, Entergy Arkansas, Minnesota Power, and DTE Energy collectively represent 1 GW of demand-response capacity — the largest hyperscaler demand-response portfolio announced to date. The Brattle Group estimates that flexible load adoption at scale could produce more than $110 billion in system-wide customer savings by reducing the need for peak-only infrastructure.

Real-Time Flexibility Has Limits, and Near-Term Emissions Risks Are Real

The technology's proponents have been careful not to overstate it, and independent analysts have added important caveats. Peter Hirschboeck, founder of energy infrastructure analysis firm Impact ECI, told Latitude Media that while the story "makes sense," real-world implementations frequently underperform their models, and that electricity prices often rise even when efficiency improves systemically.

A more specific technical risk is the near-term emissions profile. Research cited by Duke University's follow-up work acknowledges that demand response, by shifting load timing, can alter the mix of generation resources serving a facility — and that in some scenarios this means drawing more from fossil baseload plants rather than less, at least until more renewable capacity is available. Flexibility reduces peak stress and may delay expensive peak-generation investment, but the carbon calculus depends on regional grid composition in ways that vary significantly by utility and time of day.

There is also the question of which workloads can actually flex. Inference serving for production applications — where users are waiting for responses — cannot be easily curtailed without immediate, measurable performance degradation. The commercial value of the Conductor platform scales with the proportion of training and batch workloads in a given facility; as hyperscale data centers shift increasingly toward serving deployed models rather than training new ones, the curtailable fraction may shrink.

Emerald AI's software-only claim is technically accurate but strategically important. Installations that add battery energy storage systems gain a third mode of flexibility — drawing from stored energy rather than shedding load — but batteries add capital cost and carry their own capacity and durability limits. Front-of-meter renewable generation paired with flexible data centers, as modeled in the Aurora ERCOT report, provides more durable flexibility at the cost of co-siting complexity. The AEMA's work will partly consist of defining which combinations of hardware, software, and contractual arrangements qualify as meaningful demand response under the new FERC frameworks.

What the Coalition Has to Prove Next

The immediate next milestone is Emerald AI and Nvidia's planned 96-megawatt commercial-scale power-flexible AI facility in Virginia, the NVIDIA AI Factory Research Center, developed in partnership with Digital Realty, PJM Interconnection, and EPRI through the DCFlex initiative. This would be the first facility designed from the outset to operate as a grid asset — not a demonstration added retroactively — and its economic and technical performance will carry significant evidential weight for the AEMA's claims about what the coalition can actually deliver.

The Brattle Group blueprint's value depends on FERC codifying flexible-load incentives in a final interconnection rule. That rule's content and timeline remain uncertain following the June 2026 show-cause orders. The AEMA's ability to shape those proceedings — with Google, Nvidia, Anthropic, and major utilities collectively presenting a unified industry position — represents the coalition's most direct regulatory leverage.

For AI infrastructure builders, the arithmetic is changing. Power availability has overtaken GPU supply as the primary constraint on where and how fast AI compute can be deployed. The AEMA's formation marks the moment the industry stopped treating that constraint as someone else's problem to solve with transmission investment, and started treating it as a software and standards problem it intends to solve itself.