Skip to content

Power

The AI Energy Management Alliance: What a Flexible Data-Center Commitment Must Mean

AEMA's September 2026 launch changes the buyer question: which loads can respond, under what conditions, and what evidence supports the commitment?

Leviathan SystemsPublished 2026-10-205 min read

Analysis checked October 20, 2026.

On September 16, 2026, Emerald AI, Google and NVIDIA announced the AI Energy Management Alliance (AEMA) to advance grid-responsive data centers. NVIDIA's launch post emphasizes response speed, duration, predictability and emergency behavior. Its principles include defining ride-through, curtailment and contingency-response obligations before connection, plus common metrics and operational data sharing. These are announced principles, not an enacted tariff or approval for a particular facility. [1]

Our reading: the useful change for buyers is a move from a flexible label to an assessable response commitment. Ask what the facility promises to change at its grid interface while preserving the specified workload obligations. Announced principles, contracted capability and verified performance are separate stages of evidence.

A grid response and a workload pause are different commitments

The launch identifies shifting compute, discharging storage and using paired generation as possible mechanisms. [1] Their effects cross different boundaries, so a buyer should ask different questions about each.

Moving work changes when or where eligible computation runs. Ask what happens to its deadline, data access and destination resources. Deferring a batch is a different proposition from moving it elsewhere: the latter needs evidence about the destination as well. A description of movable work does not establish that every job can pause.

Storage discharge can substitute stored electricity for grid draw while computation continues. EIA distinguishes power capacity in MW from energy capacity in MWh and relates battery duration to their ratio. [3] A battery power figure alone therefore leaves the offered interval unresolved. Ask which stored energy is available for the commitment and how readiness for another event would be restored.

Paired generation can supply part of the facility load while reducing its draw from the grid. Ask which source can support the promised response under its applicable operating permissions and availability conditions. A backup asset's presence does not establish its eligibility for this purpose. These are conceptual purchase questions; the appointed utility, facility and platform parties must establish the actual scheme and operating procedures.

For a combined proposal, request an explanation of how the contributions fit together without counting the same change twice. A facility could claim a grid-import reduction while maintaining computation, or promise a reduction that depends on deferring work. Those offers expose the buyer to different workload consequences even if their headline power figures match.

Google's 1 GW is a contract milestone

On March 19, 2026, Google reported integrating 1 gigawatt (GW) of demand-response capacity into long-term energy contracts with multiple U.S. utilities. It described limiting or shifting a portion of machine-learning workloads and said flexibility would be available only at certain locations. [2]

This is reported contracted response capacity. It is not 1 GW of new generation, firm electricity supply, continuously available GPU IT capacity or capacity available to another customer. The announcement also does not supply an event record demonstrating delivery of that entire response. A portfolio milestone cannot settle the service characteristics of one campus.

For the separate question of what a site power figure allocates to a customer, use the power-capacity claims guide. Here the purchase question concerns the promised change in electricity demand and its effect on the buyer's work.

One campus, two proposals

Consider a fictional proposed campus sharing a training batch whose owner permits deferral and an inference service with a latency requirement. These are stipulated requirements, not universal properties of training and inference. Neither proposal below includes a measured response result.

Proposal A says every workload can pause when requested. It leaves the inference requirement unexplained and supplies no distinction between participating and protected work. Calling the whole campus flexible conceals the central conflict: the proposed response may depend on interrupting the very service the buyer needs to keep available.

Proposal B identifies only the stipulated training batch as potentially eligible. It leaves inference outside that workload action and asks the platform owner to demonstrate the training deferral and recovery assumptions. It also asks the facility owner to establish what change would appear at the relevant grid interface. No response amount, speed, duration or successful event is assumed.

B is more assessable, but its narrower wording is not proof of performance. The buyer can now request evidence against a particular claim: under the stated conditions, can the eligible work change enough to satisfy the proposed obligation while the inference service still meets its requirement? If the answer depends on storage or generation, that contribution needs its own supporting basis.

Require evidence of the obligation, including recovery

Start with the response boundary. Ask which facility, connection point and eligible resources the commitment covers, what conditions trigger it, and who can request and acknowledge the response. Request the agreed measurement basis: which meter or record establishes the change, and how ordinary demand variation is distinguished from the claimed action. A falling GPU-utilization chart alone does not reconcile the facility's grid draw.

Then examine the complete event, rather than a successful instant. The evidence request should connect the request time, response onset, sustained interval and end of the obligation with the corresponding facility and workload records. The relevant parties must define acceptable performance and evidence for their actual arrangement; the buyer should not import an unexplained response target from another project.

Recovery belongs in the comparison. For the fictional batch, ask whether deferred work can still finish within its permitted window and what resources that recovery assumes. For storage, ask about restoration of the energy available for the next event. A proposal that shows only the initial reduction leaves the post-event demand and readiness assumptions unresolved.

Keep emergency behavior explicit. Staying connected through a disturbance, reducing demand on request and responding to a contingency are distinct obligations in the alliance's framing. [1] Evidence for one should not silently stand in for the others. Identify which behavior the proposed facility is actually promising.

The purchasing decision becomes clearer when a supplier can connect a scoped obligation to applicable records and a workload consequence the buyer accepts. Where evidence is missing, retain a specific open question. Neither an alliance announcement nor a large contracted number answers it on the supplier's behalf.

Carry the resolved inputs into the installation scope

Once the responsible parties establish the facility and workload basis, carry the relevant physical dependencies into site-readiness planning. For customer-furnished equipment in a third-party facility, Leviathan's colocation installation service covers the agreed physical IT package. Share the equipment plan, requested tasks and known facility dependencies when discussing that installation scope.

Contact

Discuss Your Infrastructure Project

Tell us about your project, and our team will follow up to discuss the scope, requirements and availability.