As a forest of pitchforks approaches America’s data centres, Big Tech can turn to a time-tested strategy: hiding. As in, behind the meter. That phrase refers to on-site power generation that can offer backup for data centres or, at the extreme, let them bypass the electricity grid completely.
The latter seems increasingly attractive because of years-long wait times for a grid hookup [in the US], now compounded by growing general opposition to data centres.
When Governor Greg Abbott of Texas unexpectedly paused grid connections for data centres in the state earlier this month, shares of traditional power plant operators took a hit, but stocks bounced for companies that make the engines and turbines for generating power separately.
As if to punctuate this, it emerged that Amazon has begun work on a giant data-centre campus in West Texas hosting up to 7.65GW of gas-fuelled power, more than the entire generating capacity of some American states.
But going off-grid comes at a steep cost and could actually make the hyperscalers bigger targets, not smaller ones.
Just like hospitals or even houses, data centres routinely install some backup power to keep running if the grid falters. That is different from going off-grid altogether. Amazon’s GW Ranch project, for example, is described as a ‘private grid,’ albeit one that is the size of a city.
The benefit of going it alone is that companies avoid waiting perhaps four years or more for the local utility to connect them, with on-site generation up and running in something more like 18 months—a significant advantage in the accelerating AI arms race. Self-powering also offers a rebuttal to criticism that data centres are raising utility bills for their neighbours.
It is a beguiling argument, but half-baked. Maddening as the power grid can be, it has three advantages: it is the cheapest source of electricity; it is not perfect but still remarkably reliable; and it already exists. There is a reason hyperscalers previously tried to cut deals to hive off some of that valuable grid for themselves, thereby helping to fuel the backlash against data centres.
Going off-grid means building your own; that means obtaining not just the generators but all the associated equipment, too, much of which is also backlogged. You also have to install more than you need to replicate the grid’s big selling point—redundancy—because your turbines can also go down. That reduces utilization, raising unit costs further, which can then be compounded if your data centre eventually does hook up to the regular grid.
Hyperscalers are not known for their sensitivity to energy costs, so maybe none of these inefficiencies matter to them. If the US is truly engaged in the opening stages of a battle for AI supremacy with existential consequences, why not reach for the quickest, if imperfect, solutions? If there’s a war on, you would not complain if an ammo plant was running on diesel engines.
Therein lies part of the wider problem. Most Americans do not view data centres as essential defence assets, but rather as unsightly, resource-hogging nuisances that enrich the super wealthy and make a ‘product’ that could possibly put them out of a job.
Despite AI’s potential, and real tax windfalls for some localities that host data centres, they have an image problem that has festered and inevitably jumped from opinion polls to statehouses. Governor Josh Shapiro of Pennsylvania is the latest of his peers to reverse course and impose strict constraints on data centres.
Hyperscalers fumbled the important selling point that, done right, the extra load that data centres put on the grid could actually reduce bills by spreading the network’s fixed costs across more kilowatt-hours. Withdrawing from the grid would kick away that pitch completely as well as magnify the disconnect between these AI factories and the local communities in which they land.
The pollution aspect is also important. There is already a legal battle over gas turbines installed without an air permit to power a Memphis data centre owned by xAI, a unit of SpaceX. Given the sheer scale of the mostly gas-fired on-site capacity now being planned, such as for Amazon’s project, data centres run the risk of becoming the next Keystone XL.
Other routes to solving AI’s power quandary involve a mix of demand response, batteries and, yes, behind-the-meter generation—but in the context of also being active grid participants.
Retreating to islands and letting everyone around cope with the consequences isn’t going to fly. Rather, developers must work harder to demonstrate the benefits they can offer surrounding communities and integrate somehow. Becoming a good member of society doesn’t begin with retreating from it. ©Bloomberg
The author is a Bloomberg Opinion columnist covering energy.