On October 27, the US Department of Energy offered Constellation Energy a conditional loan commitment of up to $5 billion to uprate 11 reactors at eight existing nuclear plants — six in Illinois (Braidwood, Byron, Clinton, Dresden, LaSalle, Quad Cities), plus Calvert Cliffs in Maryland and Limerick in Pennsylvania. The headline writes itself: $5 billion for nuclear power for data centers.
I read the DOE fact sheet, Constellation's press release, and the TechCrunch and World Nuclear News coverage, and the headline is wrong in the specific way headlines usually are. This is not $5 billion buying new power. It is cheaper financing for power that was already sold.
What is actually happening
First, the loan is conditional, not closed. Constellation still has to meet technical, legal, environmental, and financial conditions before DOE signs definitive documents and releases any money. "Loans $5B" in a headline is, at this stage, "offers up to $5B if everything checks out."
Second, the capacity is an uprate, not a new reactor. An uprate raises the licensed output of a reactor that already exists — bigger turbines, more enriched fuel, upgraded equipment. The plan adds roughly 1.1 gigawatts across the eight plants, about what DOE says powers 930,000 homes. No new site, no decade-long permitting fight for a greenfield plant.
Third, and most important: most of this power already has a buyer. TechCrunch reports that most or all of the added output from seven of the eight plants has already been bought by Google, Meta, or Amazon. Google is the confirmed buyer at the Pennsylvania plant and four Illinois plants, Amazon at the Maryland plant, and Meta effectively takes the whole Clinton plant in Illinois. Constellation had already committed more than $7 billion to add 1,080 megawatts for those three buyers under long-term agreements.
So the sequence is: hyperscalers signed up first, Constellation committed the capital, and now the federal government is offering to make that capital cheaper.
I did the per-kilowatt math
This is the part I could check myself, with nothing but the reported numbers:
- The loan alone works out to $5 billion / 1.1 GW = about $4,545 per kilowatt.
- Constellation's own committed spend works out to $7 billion+ / 1,080 MW = about $6,481 per kilowatt.
- DOE's program allows up to 80% financing of eligible project costs, which implies eligible costs of up to about $6.25 billion behind a $5 billion loan — in the same ballpark as Constellation's $7 billion-plus commitment, not on top of it.
- The loan is up to about 71% of that committed spend.
Read that way, the federal money is not creating a new project. It is refinancing most of an already-committed one at (presumably) better terms than Constellation would get privately. That is a real benefit to Constellation and, indirectly, to the buyers — lower financing costs on firm power — but it is not 1.1 GW of new supply appearing because Washington wrote a check.
There is also a scale check worth doing. The DOE's UPRISE program, launched in March 2026, targets 5 GW of added nuclear capacity by 2029, and says it has now identified about 4.7 GW of it. Constellation's 1.1 GW is roughly 23% of that identified total. The rest includes Vistra's 433 MW of uprates backed by a $4.2 billion conditional commitment from October 5, plus restarts like Palisades (800 MW) and Crane (835 MW). Uprates and restarts, not new builds, are doing almost all the work in this program — because they are the only nuclear capacity that can plausibly arrive on a data-center timeline.
The take for people building AI infrastructure
If you run infrastructure, the lesson is not "nuclear is back." It is narrower and more useful: the binding constraint on AI power is no longer whether the megawatts exist, but who has already contracted them and at what financing cost.
1.1 GW sounds enormous until you compare it to demand: a single gigawatt-scale AI campus eats roughly that much. This entire eight-plant, eleven-reactor federal program adds about one campus worth of power, and that power was already spoken for before the loan was announced.
So do not read this as supply relief for the grid, and do not read DOE's claim about "affordable, reliable power to American homes" at face value either — TechCrunch asked, and neither Constellation's announcement nor the government would confirm that residential rates fall. The 930,000-homes figure is a scale comparison (1.1 GW divided by about 1.18 kW of average draw per home), not a promise that 930,000 households get cheaper electricity.
What this actually signals is the new shape of AI power procurement: hyperscalers lock up existing firm capacity years ahead, the government then subsidizes the financing of the upgrades, and everyone else — smaller operators, actual households — competes for whatever is left on the same grids (PJM and MISO, in this case). If you are planning capacity and you are not Google, Meta, or Amazon, the relevant question is not "will nuclear save us" but "whose contracted megawatts am I standing behind in the queue."
The uprate itself is genuinely good engineering economics — squeezing ~10% more out of plants that already exist beats building new ones on any timeline that matters. Just don't mistake a cheaper loan for a new power plant. The plants were already there. The buyers were already there. Only the financing is new.
Sources: US DOE fact sheet on Project UPRISE (October 27, 2026); Constellation Energy press release (October 27, 2026); TechCrunch, "Trump DOE loans $5B to nuclear plants that will power data centers" (October 28, 2026); World Nuclear News, "US government loan to support Constellation uprates" (October 28, 2026); POWER Magazine on the UPRISE portfolio.
Top comments (0)