An MTS research project — May 2026
The Economics of Data Centers.
As AI companies build hundreds of new data centers across the country, each costing more than the budgets of most countries and consuming more electricity than mid-sized cities, more and more thought leaders on X have made the claim that they are an excellent investment, almost akin to free money. We took a look at the public documents filed about each of these and found that this framing is heavily dependent on AI companies hitting lofty goals, and if they don't, it's not the financing companies that suffer.
If AI demand goes great: the AI company keeps most of the upside.
If AI demand falls short: a lot of the loss has already been pre-allocated to people who never saw the cap table.
Start here
If this is new to you.
Four questions to get oriented before the numbers.
- What's a data center?
- A warehouse full of computers. The new AI versions are bigger than anything that came before — a single campus can use as much electricity as a city of a million people.
- Why are they so expensive?
- Most of the cost is the Nvidia chips inside ($25–50 billion per gigawatt-sized facility). The building, power, and cooling adds another $20 billion. Call it ~$50B per facility, all in.
- Why is there an argument?
- One side (David Sacks, AI bulls) says these pay for themselves in ~2 years. The other side (commercial real-estate professionals) says ~10. Both numbers are arithmetically correct. They're answering different questions.
- Why should I care?
- Because if the answer is wrong and the buildout goes south, the loss doesn't land on Sacks or the AI companies. It lands somewhere else — and the buildout has been structured to make sure of that.
The evidencePublic documents, read carefully.Eight primary sources. Click to expand.
Each card is a regulator's order, a state report, a corporate filing, or a court case. Most coverage cites them; we read them and pulled out the load-bearing quotes. Click any card to see the document, annotated.
189 FERC ¶ 61,078 — the order that constrained behind-the-meter as a queue-bypass.
"We find PJM has failed to meet the high burden to demonstrate the non-conforming provisions are necessary." (P 86)
Report 598 — 156 pages, most cited and least read paper in the debate.
"Building enough infrastructure to meet unconstrained energy demand will be very difficult to achieve." (Summary, p. iii)
PUR-2025-00058 — first state to write a dedicated data-center rate class.
"GS-5 customers contributed an overall per books rate of return of 6.28% ... the lowest return for all rate classes other than outdoor lighting."
Every Base Residual Auction. 2027/28 cleared at the $333.44 cap.
Independent Market Monitor: ~45% of $47.2B cleared cost across the last three auctions attributable to data-center growth.
Item 9A: internal controls not effective at Dec 31, 2025.
Microsoft 67% of FY25 revenue. $21.6B total indebtedness. $60.7B RPO. Useful life extended 5→6 years.
$27.3B notes at 6.581%, A+ S&P. The off-balance-sheet template.
4-year initial lease coupled with a 16-year residual value guarantee. Strip the Meta backstop and the bonds are sub-IG.
Two reports. The 2016 forecast was correct until GPUs ate the model.
2024 report projects 325–580 TWh by 2028 — an 80% spread that is itself the key finding.
Permit 01156-01PC. NAACP v. X.AI Corp., 3:26-cv-00074-MPM-JMV.
Federal complaint quotes Brent Mayo email chain documenting 27-turbine unpermitted buildout. "Move fast and break things."