In the rush to integrate artificial intelligence into the economy, fast-moving information technology companies are running into a problem: slow-moving electric utilities. In more ways than one, the electricity supply so vital to running a data center simply doesn’t respond the way Silicon Valley and big investors want.
Broadly speaking, developers of data centers, which house the computers and servers that run AI tasks, are looking for three things: places to build, electricity to buy, and physical connections to the electric grid. The boom in forecasted electricity demand from this development has not been seen in 50 years. All of these proposed new data centers will require electrical hardware and grid improvements built by utilities. The electric power industry and its supply chain of manufacturers are backlogged and overwhelmed by the requests from data center companies. Water supplies used by data centers are stressed, too.
One move by tech companies trying to build data centers faster is to bypass the utilities and attach new data centers to existing power plants. Owners of independent power plants (i.e., plants not owned by the local utility) are making deals to provide data centers with land, the grid connection, and the electricity, with the expectation that they can cut out the local utility entirely and avoid paying the utility to maintain the grid.
You can imagine how appealing such a power grab—pun intended—looks to the companies that will benefit from it. Investment analysts say this is more profitable for the power plant. And they can avoid the queues that take years and years to move through in order to be approved to join the grid. Tech companies are so hungry for new data centers that they are pursuing deals to reopen shuttered nuclear plants, or build new ones with untried designs. The notion that an existing power plant can serve a data center, and not use the grid, is now before regulators for consideration.
Attempts to bypass grid prompt federal inquiry
The Federal Energy Regulatory Commission (FERC) has opened an inquiry into whether there is a justifiable reason to give data centers across the 13-state PJM region a discount by not paying for the electric grid. This financial break is premised on numerous deals with individual generators that claim to bypass the usual arrangement by which utility companies attach customers into the grid. Instead, the data centers want to have an exclusive connection to a single power plant and receive energy only from that one plant (plus whatever emergency power supplies they build on-site for back-up power). This is like showing up at the hospital for surgery, and saying you only want to pay for the few minutes the surgeon held the scalpel. Your surgery won’t happen without the rest of the supporting services.
This inquiry will shed light on the murky claims that data centers are not using the grid if they have contracted to buy power only from an adjacent power plant. The view of PJM, the grid operator for Virginia and surrounding states, is that the data center is still “dependent on the reliable operation of the grid,” with its market monitor calling any claim to the contrary “just non-realistic.”
Fact Check: Data centers will still be reliant on the grid
Frankly, we agree with PJM. At the heart of this dispute over whether the data centers can avoid paying for electric utility infrastructure is the false claim that a direct connection to a generator means the data center is not supported by the grid. Since the generator is connected to the grid, and relies on the grid, this claim is suspect. The details prove that suspicion is correct.
A closer look reveals that the single biggest challenge to this claim is the misconception that a data center has a smooth and unchanging demand for electricity that can be met by the host generator alone. In reality, data center energy demand fluctuates very rapidly from minute to minute, as well as on very short time scales—less than one second. Again, the fast-moving, computer-based data centers collides with the slow-moving, physically constrained electricity supply. Researchers have documented rapid changes in the demand for electricity at the start of computing tasks, and then the gyrations in electricity use while data is processed.
Turning computer processors on and off for large computational tasks causes power demand to fluctuate, unpredictably. Utilities report cases where data centers change demand as much as 300 MW in 24 seconds—the equivalent power usage of Springfield, IL. In an industry-wide survey by the Electric Power Research Institute, 26% of responding utilities described challenges in balancing supply and demand while data centers are operating. Other electric industry representatives describe data center behavior as causing problems for the supply of electricity, even in the usual arrangement where the data center is attached to the grid.
Again, fluctuation in power demand means they can’t ‘go it alone’
Simply by looking at the changing demand, it’s easy to see these variations in demand (called “ramps” by utility operators) move faster than a solo generator can follow. If the change in load is faster than the adjacent generator can match, as these reports are all describing, the imbalance in the power flow will be the absorbed by the grid. This is exactly why a grid-connected generator can’t go it alone as a supply for a data center.
When FERC rules on the issue of data centers avoiding paying for the grid support which is inherent in the proposed “co-location” behind the meter and inside the fence of existing generators, consumers will have physics on their side. The data centers and their host generators might wish they didn’t rely on the grid to handle rapidly changing power consumption, but they can’t operate without the grid. Therefore, FERC will have to conclude that the data centers do not deserve the discount they hope to receive.
Don’t believe the hype. Power hungry computing needs are not exempt from the physical constraints of the real world.