Every data center built in the past decade was designed to do one thing without pause: consume. Grid planners were told to expect a load that never blinks. On 16 September, Google, NVIDIA and a Washington, D.C. startup named Emerald AI launched an alliance built on the opposite premise. The load, they argue, should blink on command.

The coalition calls itself the AI Energy Management Alliance (AEMA). Its headline claim is that AI data centers, taught to throttle themselves, can free roughly 100 gigawatts of capacity on the existing US grid. That is the scale of roughly a hundred large reactors, recovered from software that tells a supercomputer when to wait.

🎯
The bottleneck in AI has moved from chips to grid connections. The interconnection queue, not the power plant, now sets the timeline.

Emerald AI's pitch is that flexibility is worth more than another gigawatt of firm supply. Its $1.05B valuation is a bet that utilities will pay for restraint.

Demand response is decades old. The wager is scale, automation and putting it inside the AI scheduler. The 100 GW figure depends on market rules that do not exist yet.

Emerald AI raised $150M on 25 August at a $1.05B valuation, in a round co-led by Energize Capital and DCVC and joined by twelve Fortune Global 500 companies. Total funding now exceeds $220M. Two weeks later it convened the alliance. The sequence matters. This is a company turning a funding round into a standards position.

Kettles, queues and the binding constraint

In Britain, when a televised football match reaches half-time, roughly a gigawatt of electric kettles switches on within seconds. National Grid calls it the TV pickup. It is the kind of sudden, predictable spike that grid operators have spent a century learning to absorb.

Emerald says its software met 200 grid targets during a UK trial with 100% compliance, shaving load during those kettle surges. The company has run five demonstrations in commercial data centers, in Arizona, Illinois, Oregon, Virginia and London, and published the Arizona result in Nature Energy: a 25% cut in an Oracle facility's draw over three hours on a hot Phoenix afternoon, using NVIDIA A100 accelerators.

The reason this counts is arithmetic. The International Energy Agency projects that data centers will account for nearly half the growth in US electricity demand through 2030. New transmission and generation take a decade to build. The queue to connect is measured in years.

The queue is the product.

What Emerald Conductor actually does

The platform, Emerald Conductor, intervenes in three ways. The company sorts them on its own website, and the sorting is useful for anyone auditing the claim.

Temporal flexibility

Slow or pause batchable workloads when the grid is stressed, then resume them after the event, inside service-level guardrails. Training runs are the natural candidate; a job that will finish tonight can finish tomorrow morning.

Spatial flexibility

Shift inference traffic over fiber from a constrained region to one with headroom, as long as latency budgets allow. In Virginia's 2026 winter peak, Emerald says it geo-shifted inference to Chicago within milliseconds.

Resource flexibility

Dispatch onsite batteries and other energy assets alongside compute orchestration. This is the least novel leg and the easiest to value, because batteries already have markets.

Only the first is classic demand response. The second is a networking trick. The third is a battery with better scheduling. The 100 GW claim rests on all three operating at once, at merchant scale, under contract.

The binding constraint on AI is no longer chips or capital. It is power, and software is the fastest way through it.— Dr. Varun Sivaram, Founder and CEO, Emerald AI

Five demonstrations, one peer-reviewed paper

The evidence base is real, and narrower than the headline. Three of the public demonstrations tell the story best.

ParameterArizona, 2025United Kingdom, 2025Virginia to Chicago, 2026
Site Oracle data center, Phoenix National Grid trial Constrained Virginia site
Intervention ◐ 25% turndown, 3 hours ◐ 200 targets, 100% compliance ◐ Geo-shift of inference
Metric ✔ Peak shave, peer-reviewed ✔ Kettle-surge response ◐ Latency-bound load shift
Partners EPRI, Oracle, NVIDIA, SRP EPRI, Nebius, NVIDIA, National Grid EPRI, Oracle, PJM, ComEd

Demonstrations reported by Emerald AI and its partners; the Arizona result was published in Nature Energy, 2025.

Two columns are the company's own account. The Arizona turndown is the one with an independent, peer-reviewed number attached, and it covers a single site on a single afternoon. That is the evidentiary floor beneath a claim about 100 gigawatts.

25% peak cut, 3 hours

Arizona turndown at an Oracle facility

One hot afternoon, one site, NVIDIA A100 accelerators. Nature Energy, 2025

Where 100 gigawatts comes from

Start with the hour count. The alliance's central arithmetic is that a data center does not need its full import capacity for every hour of the year. If it withdraws less from the grid during a small number of stressed hours, the same wires, transformers and generators can serve more load the rest of the time. The number is a planning estimate, not a metered reading, and it only means anything inside a market that lets flexibility count as capacity.

Google's Tyler Norris, who chairs the alliance, frames the mechanism directly: reducing net grid withdrawal for fewer than 100 hours a year can unlock dozens of gigawatts through compute flexibility, storage or generation. That sentence is the whole thesis in one line. It also names the catch. The unlocking happens in a market design, not in a machine.

Google says it has already integrated 1 GW of demand response across its US utility contracts, adding agreements with Entergy Arkansas, Minnesota Power and DTE Energy to earlier deals with Indiana Michigan Power, the Tennessee Valley Authority and Omaha Public Power District. Enel X has been paying data centers to lean on their uninterruptible power supplies for years. Industrial customers have sold curtailment to grid operators for decades. What the alliance adds is a reference architecture and a lobbying position, not a new physical phenomenon.

Two reports landed with it. The Brattle Group published a technical blueprint for turning recent Federal Energy Regulatory Commission guidance into operational market rules. Aurora Energy Research modelled flexible data centers in ERCOT paired with front-of-meter resources. Both are addressed to regulators as much as to customers, which is what a standards play looks like.

100 GW on the US grid

Capacity flexibility is said to free

An alliance estimate contingent on FERC and utility rules, not an engineering measurement. AEMA, 2026

If grid operators already run demand response, what is new here?

Two things. First, the flexibility sits inside the workload scheduler rather than in a curtailment contract bolted on afterwards, so it can be triggered in seconds instead of negotiated in advance. Second, the counterfactual load is enormous. A single frontier training campus draws like a mid-sized city, and there are dozens planned. Old demand response managed factories and big-box stores. This manages inference.

The demand-response skeptic's case

The hard part has never been curtailing electricity. It is proving it happened. Demand response depends on a baseline: what the site would have consumed had it not responded. Those baselines are gameable, and grid operators know it. Every megawatt Emerald claims has to survive that measurement problem, hour after hour, across dozens of operators.

There is a second problem. Not all AI load is equal. A training run that can be paused was, by definition, not on the critical path. The loads that genuinely cannot move, live inference for paying customers, are the ones that create the grid strain. If flexibility concentrates in the workload nobody is charging full price for, the real relief is smaller than the headline number.

Third, the 100 GW figure assumes a rule change. For flexibility to avoid a new connection, a grid operator has to let a flexible data center interconnect on the basis of its reduced net demand. That is a tariff question, not a software one. The Brattle and Aurora reports exist because the answer today is no.

As we wrote in September, grid congestion has already turned into a software market, with Octave.energy raising €10M to trade around it. On the physical side, as we wrote, the grid-equipment order book is where AI's power build-out shows up first. Emerald's bet is that the software layer, not the copper or the turbines, captures the durable margin.

$1.05B post-money valuation

Emerald AI's Series A mark, August 2026

A $150M round co-led by Energize Capital and DCVC; total funding above $220M. BusinessWire, 2026

What would have to be true

Can flexible data centers win interconnection priority by 2028?

🔮
By the end of 2028, at least one US regional grid operator will grant interconnection priority to data centers that commit to verifiable demand response.

Probability: 55% — the queue math is too painful to ignore, and FERC has already signaled direction, but tariff proceedings move slowly and baseline disputes are unresolved.

✅ Arguments for

Hyperscalers now have a common interest in a rule that lets them connect faster, and Google, NVIDIA and Anthropic carry real lobbying weight.

The engineering is demonstrated, if narrow. A verified 25% turndown is enough to prove the concept to a regulator.

Confirmation criteria: a tariff filing or FERC order that explicitly credits flexible load in an interconnection study.

❌ Arguments against

Utilities have spent decades learning to distrust demand-response baselines, and a flexible interconnection tariff is far harder to police than a factory curtailment contract.

Data centers are politically charged in many jurisdictions, which slows approvals rather than speeding them.

Disconfirmation criteria: regulators decline to credit flexible load in two or more major queue reform proceedings.

Development scenarios

🟢 Optimistic scenario (30%)

Flexible interconnection becomes the default for new AI campuses, and the 100 GW estimate proves conservative.

Implications: Emerald AI's software becomes a toll booth on grid access, and utilities re-rate existing corridors.

🟡 Base-case scenario (50%)

Flexibility is adopted gradually, mostly for training and batch load, and credited in a handful of markets.

Implications: meaningful savings for hyperscalers, but no step-change in how fast AI capacity connects.

🔴 Pessimistic scenario (20%)

Baseline disputes and local opposition stall the tariff reforms, and flexibility remains a marketing line.

Implications: the capacity crunch is solved the slow way, with steel and turbines, and software stays a side business.

What to watch

📊
Key signals to track

FERC docket activity on flexible interconnection and demand-response credit.

Whether the Brattle blueprint becomes a filed tariff or stays a PDF.

The company's next commercial deployments, and whether customers disclose them by name.

A second peer-reviewed demonstration outside Arizona, ideally at merchant scale.

The alliance is a clever piece of positioning. It takes a company with one peer-reviewed afternoon in Arizona and gives it the two things a software startup cannot buy outright: the logos of Google and NVIDIA, and a seat at the table where the rules get written. Whether that converts into 100 gigawatts of capacity is a question for regulators, baselines and time.

Sources

Emerald AI, Google and NVIDIA launch the AI Energy Management Alliance
Primary announcement of the coalition and its flexible-data-center reference architecture, from a founding member's newsroom.
The founding announcement: who is in the alliance and what it is asking regulators to allow.
The full launch-partner list and Google's 1 GW demand-response figure
Trade-press account naming the utilities, hardware vendors and grid-software firms behind the coalition, plus Google's utility contracts.
Useful for the partner roster and the policy framing from the alliance's board chair.
Why demand response is an old idea being resold at AI scale
Skeptical counterweight explaining the 100 GW estimate and the competing approaches from Google and Enel X.
The clearest statement of what is genuinely new here, and what is not.