Big Computing's New Deal with the Power Grid

By Eric Kamande

The agreement Georgia Power struck with OpenAI in July says a lot about where the relationship between computing and electricity is headed, and it goes well beyond a utility signing up a large customer.

The planned OpenAI facility in Effingham County is expected to need around 3,200 megawatts of power, roughly the output of three large coal plants. But the headline figure isn’t the most interesting part. Under the terms of the agreement, OpenAI will cover the full infrastructure and electric-service costs for the facility, provide financial assurances protecting other Georgia Power customers, and here’s the shift, make up to 1,000 megawatts of its load available for demand response. Georgia Power can reduce electricity delivery to the facility during high-demand periods.

A customer this size offering to be controllable is genuinely new territory.

From anchor tenant to grid resource

The old model was simple: a large industrial customer states its electricity needs, the utility plans around them. The customer pays the bill. The utility plans the wires and generators. Neither party has much reason to think about the other’s operational constraints.

Data centres complicate that arrangement because they sit in an unusual position. They require enormous, sustained power, but the work they do is not always time-sensitive in the way that, say, a hospital’s electricity use is. A hospital cannot defer surgery because the grid is stressed at 6pm. A data centre running batch training jobs or archival processing may be able to. The machines are still on, but the precise timing of some workloads has slack in it.

This is the premise behind demand response for computing infrastructure, and it explains why utilities are starting to negotiate flexibility provisions rather than simply capacity commitments. A large data centre that can shed or shift load on request starts to look less like a passive customer and more like a resource the grid operator can call on.

A clutch of energy and technology companies have begun organising around this idea. Several major grid operators and AI infrastructure providers, including Google and NVIDIA, have been exploring formal frameworks for data centres that can dynamically adjust electricity use in response to grid conditions. The basic goal is to turn what has been treated as fixed demand into something more responsive, a distinction that matters a great deal when a regional grid is stretched on a hot afternoon and generation reserves are thin.

Who picks up the bill for the new wires

There is a second, less discussed dimension to how these agreements are being structured: who funds the infrastructure needed to connect a massive new load in the first place.

In August, Equinix announced an arrangement with Central Georgia Electric Membership Corporation for its planned Hampton facility. Equinix agreed to pay for a new high-voltage substation and two transmission lines, committing to a 20-year take-or-pay arrangement that covers the utility’s costs for the contracted demand. The utility gets the infrastructure it needs to serve the customer. The customer gets the connection. Existing ratepayers do not bear the cost of upgrading the grid for a customer they had no say in adding.

This model is worth watching. Electricity infrastructure is built in 30 and 40-year increments. The assumptions used to justify building a new substation can become outdated within a decade. When a single customer is the reason the infrastructure gets built, making that customer responsible for its costs, over a long enough horizon, reduces the risk that other ratepayers subsidise a facility that later relocates, scales down, or fails.

Georgia Power’s OpenAI agreement takes a similar approach. The long-term financial commitments and customer protections are as significant as the raw megawatt figure, even if they get less attention.

What this means for where data centres get built

Site selection for large computing facilities has historically focused on land, water, connectivity and power availability. That calculus is getting more complicated.

A location with abundant generation but weak transmission connections presents a different proposition now than it did five years ago. A developer willing to fund grid upgrades changes the economics for a local utility in ways that a developer expecting the utility to carry those costs does not. And a facility designed from the outset to offer demand response flexibility may be easier for regulators and grid operators to accommodate than one with completely fixed consumption.

None of this removes the underlying pressure. Flexible demand response helps utilities manage the grid they have, but it cannot substitute for building new generation and transmission when demand continues to compound. A data centre that can dim its consumption by 30 percent during a grid emergency is still consuming 70 percent of an enormous amount.

But the conversation has moved. Access to power and access to the grid infrastructure that delivers it are becoming inseparable considerations. For large computing operators, electricity is no longer simply a cost to be managed, it is a constraint to be engineered around, and increasingly, a system to be actively participated in.

Get innovation insights from The FutureList weekly. Subscribe to our newsletter here