On Monday, November 2, two documents about the same problem appeared within hours of each other. The first was a white paper from Grid Strategies, a consulting firm, putting a price on a Senate bill: if it became law, new data centers outside Texas would pay about $55 billion in transmission charges over ten years, roughly $35 billion more than under today's arrangements.21 The second was OpenAI's announcement of what it calls the Grid Protectors Program, a pledge that its future campuses will cut their demand when local grids are strained. Its one firm number comes from Georgia, where a campus expected to need about 3,200 megawatts has agreed to give up as much as 1,000 of them when the utility asks.64
Read together, they look like two answers to one question: what does a data center owe the grid? They are answers to two different questions. One is who pays for the wires: the lines and substations a large new load needs, and its share of the ones already built. The other is who agrees to switch off: which loads will step back during the few hours a year when the grid is at its limit. The first is settled in dollars. The second is settled in hours.
My argument is that the second is a way of paying the first. A grid is sized for its busiest hours, so a load that leaves during those hours asks less of the wires, and the Senate bill, as its analysts describe it, already says that service which can be interrupted should cost less than service which cannot. The $55 billion estimate assumes that no new data center takes that discount. Two things I am not claiming. Switching off does not replace new wires; OpenAI itself says so. And nobody yet knows the exchange rate between an hour of stepping back and a dollar of transmission, because the terms that would set it have not been published.
The estimate assumes every new campus buys firm service.
The other currency has no published price yet.
The bill in dollars
The bill is S. 5653, the Bipartisan American Affordability and Jobs Act of 2026. Senator Shelley Moore Capito introduced it on September 30 with Senators Mike Lee, Sheldon Whitehouse and Martin Heinrich, and it was referred to the Energy and Natural Resources Committee.3 Most of it concerns permitting. The part about data centers sits in its title on electric transmission. As Grid Strategies and TechTarget describe it, it would direct the Federal Energy Regulatory Commission to change transmission pricing so that large computing loads pay for two things: their share of the network that already exists, called the embedded cost, and the incremental cost of the upgrades built to serve them.21 Congress would have to pass it and the president sign it before the commission could act.1
The arithmetic behind the $55 billion is worth seeing, because it shows where the number could move. Michael Goggin, Zach Zimmerman, Rob Gramlich and Gretchen Kershaw of Grid Strategies started from 40.3 gigawatts of new data-center capacity outside Texas in the high-confidence project tracker of Enverus, a data firm, and called that a conservative figure. For the embedded share they applied today's network transmission rates in the zones where those campuses are being built. For the incremental share they took 163 transmission projects that MISO, the grid operator for much of the Midwest, approved to connect 43.5 gigawatts of large loads, about $11.4 billion of work, and applied the cost per megawatt. The result is $35 billion embedded and $20 billion incremental over ten years: about $5.5 billion a year, against roughly $2 billion that data centers pay for transmission now. The PJM and MISO regions carry most of it, about $2.3 billion a year each.2
One assumption carries the weight. The paper assumes that every new data center takes firm service, the kind the grid must deliver in every hour, and it gives its reason plainly: few data centers today are flexible and willing to curtail. The same paper notes that the bill requires a lower rate for non-firm transmission service, and that non-firm service puts less cost on the grid because it avoids upgrades. It does not estimate how much lower.2 So the $55 billion is the price of every campus asking for firm service. It is the bill for a world in which nobody agrees to switch off.
An earlier dispatch here, on the commission's large-load orders in June, argued that a connection request is also a request to allocate risk, and that the payer should be named before the wires go up. The Senate bill names one. Those June orders also asked the regional grid operators about new transmission services for flexible large loads.8 That question is the one this dispatch is about.
The bill in hours
Why would switching off be worth anything? Because a grid is built for its peak. The wires, transformers and generators in a region are sized for the hottest afternoon or the coldest morning of the year, plus a margin. For most of the other hours they run well below that. A load that steps back during the few busiest hours does not add to the number the wires are sized for, even if it runs flat out for the rest of the year.
Georgia is the worked example. In July, Georgia Power announced that it would serve an OpenAI campus in Effingham County, near Savannah, expected to need about 3,200 megawatts, under a 25-year agreement. OpenAI will pay the full cost of the infrastructure and electric service, and it has agreed to up to 1,000 megawatts of flexible demand response: Georgia Power can reduce the energy delivered to the campus at times of high demand. The utility called it among the largest single-facility demand response commitments in the country.4 The Georgia Public Service Commission approved the contract on August 26.5
Notice the order of things. The Georgia campus pays for its wires and agrees to step back. There the two are not alternatives; they are stacked. OpenAI's November 2 program proposes to make that the pattern elsewhere: agree with each utility how much power a campus can give up, how quickly and for how long, with some sites able to feed the grid from on-site generators or batteries. According to The Next Web's account of the announcement, no other site has been named, and OpenAI concedes that flexibility cannot replace new power plants and power lines.6 That concession is correct, and it is the limit of my argument. Stepping back can pay part of the bill. It cannot pay all of it.
What is missing is the price. Neither Georgia Power's releases nor the reporting on the national program say how much notice the utility must give, how long a cut may last, or how many hours a year it may be called.456 Those are exactly the terms that decide how many dollars of wires an hour of switching off is worth.
Figure 2 shows why the hours matter more than the megawatts. It is a model, not data: an invented region whose demand peaks at 20,000 megawatts, with the Georgia campus's real sizes added on top. Drag across the sheet to choose how many of the year's busiest hours the campus gives up its 1,000 megawatts.
The model makes two points. First, how many hours a load must give up depends on the grid, not on the load. Where a region's busiest hours stand well above the rest, a short promise takes the whole 1,000 megawatts off the peak; where they are flatter, it takes many more hours to do the same. In the invented region, about 160 hours does it. Second, the energy given up is small next to the capacity it frees. Those 160 hours are about 0.6 percent of what the campus would use in a year at full load. Beyond that point, more hours buy no more relief: the peak cannot fall further than the 1,000 megawatts the campus agreed to give up, while the energy it forgoes keeps rising. A real region and real terms will give different numbers. The shape of the trade, a little energy for a lot of capacity, is general.
That shape is why the two currencies can be exchanged at all. A firm campus pays for wires that are sized to carry it at the peak. A campus that steps back at the peak needs less of that size, and a tariff with a lower non-firm rate would let it pay less in dollars for what it pays in hours. Whether the discount is worth taking depends on what the campus runs and what it can pause, which nobody outside it can see.
The other way to ration
Britain is approaching the same problem from the queue rather than the tariff. On October 28 its Department for Energy Security and Net Zero published its response to a consultation on connecting what it calls strategic demand. It will let the national system operator move projects the government names ahead in the connection queue. And from 2030 it will align data-centre connections with regional targets that, in the response's words, "will place limits on the volumes that can connect in each region," set by grid capacity and expected demand, as well as water, land and the needs of other users. Work on those targets is under way. Separately, the energy regulator Ofgem is considering flexibility arrangements for data centres and a commitment fee meant to push speculative projects out of the queue.7
The American bill would ration by price: connect, and pay for what you use of the wires. The British response rations by quantity: only so much may connect in each region. Both leave room for flexibility as a release valve, and neither has yet said what it is worth.
Objections, considered
Data centers run around the clock for customers who expect them to. This is the strongest objection, and Grid Strategies' own assumption reflects it: few data centers today are willing to curtail. A campus that serves live requests cannot simply go dark. But the Georgia contract is a signed, approved commitment to give up nearly a third of a campus's load, not a slogan, and a commitment that size suggests the operator thinks some of its work can wait or move. It is one contract. How many others can follow depends on what each campus runs, which is the operator's knowledge, not the grid's.
Making data centers pay more will just slow the building. It may slow some. But the bill does not change what the wires cost; it changes who pays for them, and the alternative is that other customers on the network pay instead. A lower non-firm rate gives a campus a way to pay less that does not move its cost onto someone else.
Switching off only moves the problem to a diesel generator. Sometimes it might. OpenAI's program mentions on-site generation and batteries as ways to give power back.6 What a campus runs on when it leaves the grid has its own costs and its own rules, and a step back that is met by a local generator is a different trade from one met by pausing work. The terms should say which.
It is a bill, not a law. True. S. 5653 sits in committee, and the $55 billion is an estimate of a proposal under stated assumptions.32 The argument does not depend on its passing. Any tariff that prices firm and non-firm service differently, in any region, sets the same exchange rate.
A load that steps back in the grid's busiest hours pays part of its bill in time.
Before you plan around the grid
- Ask which service you are buying. Firm and non-firm transmission may soon be priced differently, and the estimate everyone is quoting assumes firm.
- Price the hours, not only the megawatts. How much notice, how long each cut, how many hours a year: those terms set what stepping back is worth.
- Know what an hour costs you. Which work can pause, which can move to another site, which needs a battery or a generator. Only the operator can answer this.
- Read the assumptions under the headline number. The $55 billion depends on 40.3 gigawatts, today's rates, and firm service for everyone.
- Look at the queue as well as the tariff. In Britain the scarce thing may be a place in line in a given region, not the price of the wires.
Back to the slip in Figure 1. Its first half is priced to the billion. Its second half has two lines marked "not published." The grid sends the bill either way. What a campus can still choose is the currency, and until someone fills in those two lines, nobody can say what the second currency is worth.