AI data center power: Nvidia and Google's alliance says flexible AI could free 100 gigawatts, but the number is a pitch

Nvidia, Google and Emerald AI launched the AI Energy Management Alliance on September 16, 2026. Its pitch to utilities and regulators is simple: let AI data centers connect to the grid faster, and in return they will cut their power draw when the grid is under strain. The alliance says flexible AI could unlock an extra 100 gigawatts from the existing US grid.

That figure matters because electricity, more than chips, is the next limit on how fast the frontier labs can scale. If you are searching for how AI data center power is actually going to be found, this is the first organized industry answer. It is also a claim, not a measurement.

What the alliance is

The Nvidia launch post, dated September 16, describes an effort to "build AI infrastructure that doesn't just connect to the grid but works with it." The alliance would set technology-neutral, performance-based requirements, measured by response speed, duration, predictability and emergency response, that a facility must meet before it connects. The aim is to convene computing platforms, data center operators, power producers, utilities and regional grid operators around those requirements.

SiliconANGLE, reporting on September 17, defines the concept: "A flexible data center is a facility that cuts its power draw when the electrical grid is under strain." The founding members are Nvidia, Google and Emerald AI. Eighteen more companies joined at launch, including Anthropic, National Grid, AES and RWE.

The numbers, and where they come from

Four figures carry the case. Each needs a source attached.

• 100 gigawatts. AEMA states on its website that "flexible AI could unlock an extra 100GW from the existing grid," as reported by The Next Web. Emerald AI chief executive Varun Sivaram said, per TechBriefly: "America could unlock 100GW on our existing grid for flexible data centers." He noted the grid runs at roughly 50 percent of capacity on average.
• 1 gigawatt. Google has committed to reducing "1GW of power demand...when needed, through utility agreements across the US," said Tyler Norris, its head of advanced energy market innovation, according to The Next Web.
• About 100 megawatts. Nvidia, Digital Realty and Emerald AI plan to switch on a power-flexible AI factory in Virginia, which SiliconANGLE describes as the world's first, later this year.
• A 5 to 10 year backlog. The Next Web puts the wait to connect new AI facilities at that range, and says Texas alone holds 474 gigawatts of data center requests.

Put the first number against the last one. The grid has room for a lot of new load in most hours. What it lacks is the ability to serve that load in its worst hours. The alliance's answer is that a data center willing to throttle for those hours deserves a place at the front of the queue.

Why a superintelligence paper cares

The race to build systems at or beyond human level is a race to turn more electricity into more compute. Our Safe Superintelligence tracker follows one lab whose strategy rests on access to a very large machine. The question for every lab is the same: where does the power come from, and how soon? A five-to-ten-year interconnection wait does not fit a business that expects its models to improve every few months.

That is why Anthropic's presence among the launch partners is telling. A lab that has asked for the industry to pace itself is also lining up to get more power faster. Both can be true. Pacing is about capability and oversight. Power is about capacity, and capacity is what every lab is short of.

What the 100 gigawatts is and is not

The claim has three limits that the launch material does not resolve.

1. No published method. The Next Web reports that the alliance's website does not attribute the 100 GW figure to a specific individual or provide its methodology. We could not find a supporting study in the sources we opened. 2. It depends on behavior. TechBriefly notes the projection rests on facilities staying flexible "during the grid's worst hours of the year." That is a promise about future operations, not capacity built. 3. Training and serving are different loads. Throttling a batch job is easy. Throttling a service people are using is not. TechBriefly lists workload-shifting software, batteries and on-site generation as the tools, and none of that is yet proven at scale. The Virginia site, around 100 megawatts, is the first test, and it is not live.

There is one number with independent backing: the alliance cites a Brattle Group estimate that each 10 percent gain in utilization lowers electricity rates by about 3.4 percent. That is an argument for why utilities and the public might welcome flexible load, separate from whether it unlocks 100 gigawatts.

Hype check

The claim is that flexible data centers can free 100 gigawatts. The honest status is an estimate from Emerald AI's chief executive, restated on the alliance's website, resting on average grid utilization. It is plausible as a direction and unproven as a size. We rate it mixed.

Our take

This is a sensible deal in a hard market, and it is a better answer than asking households to pay for new generation first. But it is a trade: speed to connect in return for a promise to curtail. The first real data will be the Virginia facility and Google's 1 gigawatt in practice. Until a utility or regulator accepts the terms and a flexible site rides through a real grid emergency, treat 100 gigawatts as a pitch. What we would watch next is which utilities and regional grid operators adopt the requirements, because a promise to flex only counts when the party running the grid agrees to count it.