The Next AI Data Center Will Be Approved for What It Can Switch Off

On September 17 the US House voted 417 to 3 to make large data centers pay for the grid upgrades they trigger. One day earlier, Google, Nvidia and Emerald AI launched an alliance whose whole pitch is that a data center should be paid for turning its power down on request. Those are the same story told from opposite ends. The question of who pays for the wires is being answered with a second question: what can you switch off, and can you prove it?

## The flexibility bet

The AI Energy Management Alliance launched with 18 partners and a push for performance-based standards that reward data centers for verifiable flexibility, according to Nvidia's own announcement. No money was disclosed, and no standard exists yet. The commercial bet behind it is clearer. Three weeks before, Emerald AI raised a $150 million Series A at a reported $1.05 billion valuation for software that throttles AI workloads when the grid is tight (deal record). A billion-dollar valuation for a company whose product is turning GPUs down is a statement about where the constraint sits.

## The harsh version

Europe is writing the same requirement as a rule rather than a standard. Spain's draft decree would force data centers and AI factories to source at least 80 percent of their electricity from new, hourly-matched renewable generation. More than 40 investors at a Goldman Sachs forum in London called it economically unworkable and warned projects would move to other European markets. The same week, the IMF told EU finance ministers that AI could add about 1 percent to European productivity over five years while straining electricity networks through data center growth.

Hourly matching is flexibility by decree: if the wind is not blowing at 3 a.m., the load has to move. The American version lets the operator sell that same capability back to the grid. Either way, the fixed-load data center is becoming the expensive kind.

## The software is already paying

The cheapest evidence that this works comes from Korea. Kakao and LG CNS say AI agents now run the cooling at a leased data center in Hanam and that cooling energy costs fell about 23 percent year on year with no hardware replaced. It is the companies' own figure, for one site and for cooling only, so treat it as a ceiling rather than a benchmark. But it is the shape of the thing: the saving came from control software, and control software is what a flexible grid asset runs on.

The tenants have noticed. Anthropic's Queensland lease anchors a campus planned to reach about 2.16 gigawatts and A$32 billion, and the deal record notes Anthropic covering the grid connection costs for its leased capacity. A tenant that pays for the connection has every reason to want the software that lowers the bill, and every reason to ask the landlord how much load it can shed.

## What 2027 looks like if this holds

Hedged, because the Senate has not voted, Spain's decree is a draft and the alliance has published no standard: the interconnection queue stops being sorted by megawatts requested and starts being sorted by megawatts that can be curtailed on notice. A campus that can prove it will drop 20 percent of load for two hours gets connected ahead of one that cannot, and the proof comes from software like Emerald's, audited against a standard like the one the alliance is promising. The compute deals we track will start carrying that number the way they carry the megawatts today.

## What to do with this

First, ask any AI campus, lease or compute contract for two numbers: how many megawatts it can shed on request, and for how long. If neither is in the document, the grid bill is the tenant's problem and nobody has priced it.

Second, watch whether the alliance publishes a measurable standard before the end of the year. A standard with a number is a market. A standard without one is a press release, and Emerald AI will be the first company whose valuation depends on which it turns out to be.