Google, Nvidia and Anthropic are behind AEMA, an alliance that wants electricity demand management written into how data centers are planned and built, and the startup at its center, Emerald AI, has just raised $150 million in a Series A led by Energize Capital and DCVC. Emerald AI's software sits between the utility and the data center: when the grid needs load taken off, a site pauses non-critical jobs or moves the computation to another data center that still has headroom. Three of the largest buyers of AI compute are funding a company whose product is making AI compute stop.
The logic starts with how grids are built. A network is sized for its worst hour, so for most of the year demand sits noticeably below the maximum the system can carry. That unused margin is the asset everyone in this alliance is trying to get at.
Selling it back is an old trade. Utilities have paid large industrial customers for decades to cut consumption when demand peaks — halt a production line, switch to backup generators — and the payments have often been generous. Data centers can already join those programs, and the usual way in is the same one a factory uses: start the diesels.
Emerald AI's pitch is that the diesels are unnecessary. Its software connects the utility's request directly to the workload, which the company says should let a data center answer the grid almost as fast as a battery system. The startup is not alone in trying. Google is building tools of its own, and Enel X lets data centers lean on their uninterruptible power supplies to shave peaks.
The size of the prize comes from a Goldman Sachs study published last year: holding a data center's draw from the grid to 90% of its maximum for a few hours at a time would free up 76 GW of capacity. The number sitting opposite it is 100 GW of additional data centers.
Put those two figures next to each other and the case looks weaker than it reads. 76 GW is a modeled ceiling, not delivered capacity — it assumes the cap holds, that operators accept it, and that the hours the grid needs are hours the workload can spare. Even taken at face value it does not cover the 100 GW being built. This is a technology that buys time on the interconnection queue, not one that removes the need to add generation. It is worth noticing that Google is simultaneously a participant in the alliance and a competitor building the same capability in-house, which is what companies do when they think the function matters and the vendor is replaceable.
The more revealing part of AEMA's agenda is its second goal: helping technology companies and energy organizations find new sites for data centers. Both sides have struggled to match one to the other. An alliance that needed only software would not have that item on its list. Siting is the bottleneck; flexible load is the argument you bring to a utility to get through it faster.
What the announcement does not settle is who pays, and for what. Factories got paid well to curtail, and the whole demand-response market runs on that payment. Nothing here says what a data center receives for deferring work, or which AI jobs count as non-critical — and that definition is the entire product. A job that can wait a few hours is a very different asset to a grid operator than one that cannot, and the distinction is left to the operator making the promise.
Emerald AI's chief scientist, Ayse Coskun, told TechCrunch that the company's technology can reduce the industry's need for new generation but cannot remove it. That is an unusually honest framing from a company that just raised $150 million, and it sets the ceiling on the whole exercise. The load AI creates moves in jumps, which is precisely why data centers make good demand-response participants and precisely why nobody should expect flexibility to substitute for steel in the ground.