Emerald AI, a grid software company that reached unicorn valuation last month, announced on 17 September 2026 that it has formed a coalition with Google and Nvidia to use its technology to secure space on the electricity grid for more data centres. The group is called the AI Energy Management Alliance, or AEMA, and its founding trio is joined by Anthropic and by a set of utilities that includes AES, Constellation, National Grid and NRG Energy.
The alliance’s pitch is a single number. By pausing noncritical tasks and shifting some compute loads between sites, AEMA says an additional 100 gigawatts of data centres could connect to the existing grid — capacity that is already there for most of the year, sitting unused because the network is built for peak demand rather than average demand. No new power plants, no new transmission lines, no waiting.
That is a large claim from a young company, and it deserves examination rather than repetition. This article sets out what Emerald AI’s software actually does, how it differs from the demand response programmes utilities have run for decades, how the 100 gigawatt figure compares with an independent estimate published last year, and where the company’s own chief scientist says the approach runs out of road. For anyone planning data centre operations against a multi-year connection queue, the distinction between the two numbers is the part that matters.
Table of contents
- What Emerald AI Actually Built
- The AI Energy Management Alliance in Full
- How Demand Response Worked Before Emerald AI
- The 100 Gigawatt Claim, Checked Against an Independent Estimate
- Why Nvidia and Anthropic Signed Up to Emerald AI
- Emerald AI Against Google’s Own Tools and Enel X
- The Money Behind Emerald AI
- Why the Timing of the Emerald AI Alliance Makes Sense Now
- What Emerald AI Does Not Fix
- What This Means for Data Centre Operators
- Frequently Asked Questions About Emerald AI and AEMA
- References
What Emerald AI Actually Built
The product is coordination software, not hardware, and that framing explains both its promise and its limits.
The core function
Emerald AI’s software coordinates requests from utilities with data centres. When the grid is under strain, the utility sends a signal; the software decides what can be paused or moved. Noncritical tasks stop. Some compute loads shift to other data centres where the grid still has headroom.
Why that is different from a generator
Today, a data centre participating in a demand response programme usually responds by turning on backup generators. It does not use less power from the grid because it needs less; it uses less because it is burning diesel instead. Emerald AI proposes an alternative to dirty diesel generators, reducing the draw rather than relocating its source.
The direct connection
Emerald AI’s specific advantage, according to TechCrunch’s reporting, is that its software connects utilities directly to data centres. That directness is what allows a site to respond quickly to a request — quickly enough, the argument goes, that a data centre behaves much like a battery on the network.
Why the battery comparison is the whole pitch
Utilities do not value a resource that responds in an hour the way they value one that responds in seconds. Framing Emerald AI’s controllable load as battery-like is a claim about response time, and response time is what determines whether a grid operator will count the flexibility when deciding how much new load to approve.
| Approach | How the load drops | What it burns |
|---|---|---|
| Traditional factory demand response | Production pauses | Nothing, but output is lost |
| Data centre with backup gensets | Grid draw shifts to on-site fuel | Diesel |
| Enel X uninterruptible power supplies | Stored power rounds off the peak | Nothing, for a short window |
| Emerald AI software | Noncritical tasks pause or move site | Nothing; work is deferred |
The AI Energy Management Alliance in Full
AEMA is not a product launch dressed as a coalition. The membership list is doing specific work.
The founding trio
Emerald AI, Google and Nvidia formed the alliance. Google operates data centres at global scale and has been developing its own flexibility tools. Nvidia sells the hardware whose power draw created the problem in the first place.
Anthropic’s presence
Anthropic joined the founding trio. It is the only member that is purely a model developer — it is neither a chipmaker, a hyperscaler nor a utility — which makes it the clearest signal that frontier labs now treat grid access as a strategic constraint rather than a landlord’s problem.
The utilities
AES, Constellation, National Grid and NRG Energy are in. Without utilities, an Emerald AI alliance would be a group of customers agreeing among themselves to be flexible. With them, it is a set of organisations that can actually approve an interconnection.
The second, quieter promise
The coalition also promises to help tech companies and utilities find new sites for data centres — an effort that has not been easy for either party. Siting is a slower, more political problem than demand response, and putting it in the same announcement suggests AEMA intends to be a standing forum rather than a one-off statement.
| Member | Role in the alliance | What it needs from it |
|---|---|---|
| Emerald AI | Supplies the coordination software | Utility-side adoption at scale |
| Founding operator | Faster interconnection for new sites | |
| Nvidia | Founding hardware vendor | Grid capacity to keep absorbing GPUs |
| Anthropic | Frontier model developer | Compute that can actually be plugged in |
| AES, Constellation | Generation and retail utilities | Load that behaves during peaks |
| National Grid, NRG Energy | Network and generation operators | Deferred network reinforcement |
How Demand Response Worked Before Emerald AI
The concept AEMA is built on is not new. Utilities have used it for decades, and understanding the old version explains what is genuinely different about the new one.
The structural fact it exploits
The grid is designed with peak loads in mind. For much of the year, demand sits well below the maximum the network can support. That headroom is real, permanent and currently wasted — which is the entire opportunity Emerald AI is selling.
The original customers
Traditionally, demand response covered large electricity users such as factories. In times of high demand they would agree to reduce usage, typically by pausing production or switching to backup generators. In exchange, the utility paid them for the service, often handsomely.
How data centres joined
Data centres can already take advantage of those programmes today. They usually do it the same way a factory did: by turning on backup generators. The mechanism was inherited wholesale, diesel and all.
What Emerald AI changes about the transaction
The old deal traded output or air quality for grid stability. The Emerald AI version trades latency on work that was never urgent. If the software correctly identifies which tasks are genuinely noncritical, nothing visible is lost — and that conditional is where the engineering risk sits.
The 100 Gigawatt Claim, Checked Against an Independent Estimate
AEMA’s headline figure has a useful comparison point published last year, and the two do not match.
What AEMA says
Pausing noncritical tasks and shifting some compute loads could allow an additional 100 gigawatts of data centres to connect to the grid.
What Goldman Sachs found
A study by Goldman Sachs published last year concluded that limiting maximum grid usage to 90 percent for a few hours at a time could let data centres free up 76 gigawatts of capacity in the United States.
The gap between them
The two figures are 24 gigawatts apart. AEMA’s number is roughly 1.3 times the Goldman Sachs estimate — about 32 percent higher. That is not a contradiction: the Goldman study modelled a specific, conservative rule, while the Emerald AI approach adds load-shifting between sites on top of simple curtailment.
How to read the difference
Treat 76 gigawatts as the figure derived from a published methodology and 100 gigawatts as a coalition’s target. Both are estimates of the same underlying headroom; the higher one assumes more aggressive coordination actually happens in practice, across many operators, reliably, during the hours it is needed.
Why Nvidia and Anthropic Signed Up to Emerald AI
Each founding member has a different bottleneck, and the alliance addresses all of them at once.
Nvidia’s interest
Nvidia’s constraint is no longer manufacturing. It is whether customers can energise the racks they have ordered. Grid capacity that arrives through software rather than through a decade-long transmission project converts directly into deliverable hardware.
Anthropic’s interest
For a model developer, an interconnection queue is a training schedule. Anthropic’s presence in an Emerald AI alliance indicates that access to power has become a first-order planning input for labs, alongside chips and data.
Google’s interest
Google has been developing its own flexibility tools, which makes its membership the most interesting. A company that could go it alone is choosing a shared forum instead — usually a sign that the hard part is the utility relationship, not the software.
The utilities’ interest
A utility that can count on a large customer reducing demand during a handful of peak hours can approve that customer’s connection sooner and defer network reinforcement. The flexibility is worth money to them, which is why demand response programmes have historically paid participants handsomely.
Emerald AI Against Google's Own Tools and Enel X
The startup is not alone in this market, and the competing approaches reveal what is genuinely hard.
Google’s internal tooling
Google has been building its own demand-flexibility tools. As an operator with global scale, it can shift work between its own regions without needing an intermediary — an option unavailable to a single-site colocation customer.
What Enel X does differently
Enel X lets data centres tap their uninterruptible power supplies to round off peaks in power demand. That draws on stored energy rather than deferring computation, so it is limited by battery capacity and duration rather than by workload flexibility.
Where Emerald AI sits between them
Emerald AI’s approach requires neither a fleet of sites nor a large battery estate. It requires knowing which workloads can wait — a scheduling problem rather than an energy-storage one, and one that scales to operators of any size.
The unsolved part
All three approaches depend on the same unproven assumption at scale: that a meaningful share of AI compute is genuinely deferrable. Training runs, including long reinforcement learning stages, have checkpoints; inference serving a live product does not. The mix determines the real number.
| Provider | Mechanism | Main limit |
|---|---|---|
| Emerald AI | Pause and shift noncritical compute | How much work is deferrable |
| Google internal tools | Move work between owned regions | Requires a global fleet |
| Enel X | Discharge uninterruptible power supplies | Battery capacity and duration |
| Diesel gensets | Switch the source, not the load | Emissions and local air quality |
The Money Behind Emerald AI
The funding round is what turns the alliance from a position paper into a deployment plan.
The Series A
Emerald AI recently raised $150 million in a Series A led by Energize Capital and DCVC, at a reported $1.05 billion valuation. That gives it the kind of funding needed to roll out the technology on a wider scale.
What the round implies
A $150 million raise against a $1.05 billion post-money valuation means the round represents roughly 14 percent of the company. That is a conventional structure, not a distress signal or a frothy outlier — the notable part is the Series A stage at a unicorn price.
Why a software company needs that much
Grid software is sold to utilities, and utility sales cycles are measured in years. The capital is buying time in front of regulators and network operators far more than it is buying engineering.
Why the Timing of the Emerald AI Alliance Makes Sense Now
The coalition arrives late by one reading and early by another, and both readings are informative.
Overdue, given how AI loads behave
AI compute is unusually peaky, and data centres can ramp up and down quickly. Those two properties are exactly what a demand response programme wants, and they have been true for years. On that basis the creation of something like AEMA is overdue rather than opportunistic.
Early, given how little is standardised
There is no common protocol for a utility to signal a data centre, no agreed definition of a noncritical workload, and no settled way to verify that a promised reduction actually happened. An alliance is the usual vehicle for writing those definitions before a regulator writes them instead.
The interconnection queue is the real clock
In several US markets the wait to connect a large new load is now measured in years. Every month of that queue is a month of idle capital for an operator and a month of deferred revenue for a chip vendor. Emerald AI is selling a way around a queue, which is a far more urgent product than a way to save on an electricity bill.
The political weather has changed
Data centre power demand has become a local political issue in a way it was not two years ago, with siting decisions drawing organised opposition. A consortium that can show flexibility, rather than simply requesting more supply, has a better story to tell a public utility commission — and that story is a large part of what Emerald AI’s members are buying.
What would prove it works
The test is not another announcement. It is a published event log: a named utility, a named site, a dated curtailment request, the megawatts actually shed, and the workload class that absorbed it. Until an Emerald AI deployment produces that record in public, the 100 gigawatt figure remains an argument rather than a measurement.
What Emerald AI Does Not Fix
The company’s own chief scientist set the limits publicly, which is more candour than these announcements usually contain.
The chief scientist’s caveat
Ayse Coskun, Emerald AI’s chief scientist, told TechCrunch that the company’s technology promises to blunt the industry’s need for new generating sources, but that it will not eliminate it entirely. Blunt, not remove.
Why that caveat is structural
Demand response monetises the gap between peak and average. It cannot raise the peak. Once AI load grows to the point where the average approaches the old peak, the headroom is gone and the only remaining answer is new generation.
The deferral is not free
Paused work still has to run. If a data centre defers four hours of training every evening during a summer heatwave, that work reappears later — and a fleet running closer to its ceiling has less room to absorb it.
Reliability is a policy question too
A utility approving a connection on the strength of promised flexibility is taking a commercial promise as an engineering input. Whether regulators let them do that, and on what terms, will decide how much of the 100 gigawatts is ever realised.
What This Means for Data Centre Operators
For anyone outside the founding members, the practical questions are narrower than the headline.
Classify the workload first
The value any operator can extract from an Emerald AI style scheme is set by the share of compute that can genuinely wait. That number should be measured before it is negotiated, not after.
Read the tariff, not the press release
Demand response pays under a utility’s programme terms, which vary by market. The alliance does not change those terms; it changes how easily a site can meet them.
Expect the queue to move before the price does
The near-term benefit of grid flexibility is a faster interconnection, not a cheaper bill. In markets where the queue is the binding constraint, that is the more valuable of the two.
Where it fits a wider plan
Grid flexibility belongs in the same conversation as siting, cooling and hardware refresh cycles. Teams building an AI strategy around large-scale training should treat power availability as a design constraint rather than a procurement detail, and our AI models and tools hub tracks the infrastructure announcements alongside the model releases.
Frequently Asked Questions About Emerald AI and AEMA
What is the AI Energy Management Alliance?
A coalition founded by Emerald AI, Google and Nvidia, joined by Anthropic and by the utilities AES, Constellation, National Grid and NRG Energy, aiming to make demand response an integral part of data centre development.
Is the 100 gigawatt figure a forecast or a target?
It is what AEMA says could be connected if noncritical tasks are paused and compute loads shifted. It is an ambition attached to a mechanism, not an outcome anyone has delivered.
How does this differ from a data centre using backup generators?
Generators change where the power comes from. The Emerald AI approach reduces the amount of power drawn at all, by deferring or relocating the work itself.
Does demand response reduce a data centre’s total energy use?
No. It moves consumption out of peak hours. The energy is still consumed, which is why the chief scientist’s caveat about new generation still applies.
Who pays the data centre for the flexibility?
The utility does, under its demand response programme. Utilities have paid large industrial participants for this service for decades, often handsomely.
Is Emerald AI the only company offering this?
No. Google is building its own tools and Enel X uses uninterruptible power supplies to shave peaks. Emerald AI’s differentiator is a direct utility-to-data-centre link and a battery-like response time.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.