A new AI data center arrives with a 500 MW power request.
The utility may need a dedicated substation. The regional grid may need stronger transmission. New generation or capacity may have to be secured.
Those are three different costs.
And they do not automatically have the same payer.
The real question is not simply “Should the data center pay?” It is “Which cost did the data center cause, who else benefits, and who carries the risk if the project changes?”
That distinction has become much more important since this article was first published.
In late September, federal regulators told PJM—the largest U.S. regional grid operator—to revisit parts of a reliability plan partly because of how new large-load costs could fall on existing customers.[5][6]
At the same time, utilities such as AEP Ohio are using collateral, minimum-demand charges and stranded-cost protections to make large data-center customers put money behind the capacity they reserve.[2]
Quick Answer
The emerging principle is cost causation.
In plain language:
If a new customer causes a major grid cost, that customer should generally carry an appropriate share of the cost.
But “appropriate share” is the difficult part.
A substation built almost entirely for one campus is different from a transmission line that also improves regional reliability and connects future generation.
So a better framework separates:
dedicated cost
→ shared-system cost
→ generation / capacity cost
→ reservation risk
→ stranded-cost risk
Original Asset 1: The Cost Causation Stack
1. Direct connection
Dedicated substations, transformers, feeders or lines built mainly to connect one data center.
2. Local network
Upgrades in the nearby utility system needed to move the new load safely.
3. Regional transmission
Higher-voltage upgrades that may serve the data center but can also support other customers or new generation.
4. Generation and capacity
Additional power plants, storage or capacity commitments needed because the system must meet the new peak demand reliably.
5. Reservation and stranded risk
What happens if the customer reserves 500 MW, the utility commits capital, and the project later shrinks, delays or leaves?
The farther down this stack you go, the harder it becomes to say that one customer either should pay everything or nothing.
Why Households Ask, “Why Aren't They Paying for Their Own Infrastructure?”
This question appears repeatedly in public discussions.
The confusion comes from the fact that utilities historically separate customer-specific facilities from shared network investment.
A new factory, home or data center may pay direct connection costs.
But a larger transmission upgrade can enter a regulated system where costs are spread across many customers if regulators conclude the asset provides wider benefits.
Data centers make that old distinction harder because one new customer can be hundreds of megawatts.
A single project can materially change the timing and size of regional investment.
Original Asset 2: The Dedicated vs. Shared Benefit Test
When a grid upgrade appears, ask six questions.
- Would this asset be needed without the data center?
- Was it sized specifically for this customer?
- Can other customers use it?
- Does it improve wider system reliability?
- Does it unlock new generation or future industrial load?
- Would the asset still be useful if the data center left?
The more customer-specific the answers are, the stronger the case for assigning the cost directly to the large load.
The broader the benefit, the stronger the case for some form of shared allocation.
AEP Ohio Shows How Financial Commitment Can Protect the Grid
AEP Ohio's Data Center Tariff is one of the clearest examples of the new approach.
For customers that do not meet specified credit and liquidity tests, the tariff can require collateral equal to 50% of the total minimum charges for the full contract term.[2]
The tariff also uses a minimum-demand charge.
That means a customer cannot simply reserve a large amount of capacity and then pay only for a very small actual load.
For many customers, monthly billing demand is tied to high percentages of contract capacity or prior peak demand.[2]
The idea is simple:
If the grid builds for your reservation, your financial commitment should resemble the capacity you asked it to reserve.
Original Asset 3: Stranded Cost
Stranded cost is the part of an infrastructure investment that remains unrecovered or underused after the customer or expected demand disappears.
A simple diagnostic is:
Stranded-cost exposure
≈
committed grid capital − recoverable value if the customer leaves
This is not a tariff formula.
It is a way to ask where the risk sits.
AEP Ohio's transfer rules make the issue unusually explicit. A customer can seek to reassign some unused contract capacity, but the transfer cannot leave stranded investment for other ratepayers unless the assigning customer pays that cost.[2]
Why Minimum-Demand Charges Matter
Imagine a utility builds around a 500 MW contract.
The data center later uses only 150 MW.
If the customer pays only for the 150 MW it happens to consume, other customers may be left paying for the capacity that was built and reserved but remains unused.
A minimum-demand rule shifts some of that risk back to the party that made the reservation.
This connects directly to the ghost-demand problem.
Grid planning becomes safer when requested capacity is backed by financial commitment.
Microsoft's “Pay Our Way” Principle
Microsoft made its position explicit in January.
Its Community-First AI Infrastructure Initiative says it will ask utilities and regulators to set rates high enough to cover the electricity costs created by its data centers and that it will pay for transmission and substation improvements required by its expansion.[1]
Microsoft's argument is not that every grid asset should automatically be charged 100% to one company.
Its stated principle is that residential customers should not be left paying the added electricity costs caused by Microsoft's growth.
That is a corporate commitment, not a nationwide rate rule.
Google's Minnesota Deal Shows Another Structure
In February, Google and Xcel Energy announced a Minnesota arrangement tied to a new data center.
Reuters reported that Google would fully fund a 1.9 GW clean-energy buildout—1.4 GW wind, 200 MW solar and 300 MW long-duration storage—and invest another $50 million in battery storage support. The arrangement was structured so existing customers would not see bill increases from the project.[4]
This is a project-specific solution.
It illustrates a broader direction:
large loads are increasingly expected to arrive with a credible plan for the power system they require.
Why “Make the Data Center Pay 100%” Can Also Be Too Simple
A transmission line may start with one large customer.
But it can also:
- improve reliability,
- connect new generation,
- serve future industrial customers,
- relieve congestion,
- or reduce the need for other upgrades.
If one customer is charged 100% for an asset that provides large long-term benefits to many users, the allocation can overstate that customer's responsibility.
This is the other side of cost causation.
Original Asset 4: Two-Sided Allocation Error
Over-socialize the cost:
one large load causes most of the investment
→ cost is spread widely
→ existing customers subsidize the new load
Over-assign the cost:
one large load triggers an upgrade
→ upgrade creates broad system value
→ one customer pays for benefits used by many
Good rate design tries to avoid both errors.
FERC and PJM: The Question Has Reached Regional Grid Planning
On September 29, a FERC commissioner wrote that customers driving new costs should bear appropriate responsibility and that existing customers should not be left paying costs attributable to new demand.[5]
The immediate issue involved PJM's proposed reliability backstop procurement as the region faced a capacity shortfall linked in part to rapid load growth.
Reuters reported on September 30 that FERC directed PJM to delay and revise parts of the plan and conduct further proceedings on cost allocation. Current reporting put the relevant capacity shortfall at roughly 6.8 GW.[6]
This does not create one universal U.S. rule.
It shows that the data-center cost question has moved from individual utility tariffs into regional market design.
What About “Bring Your Own Power”?
Some developers are trying another path: build generation behind the meter or next to the data center.
Behind-the-meter generation means generation located on the customer's side of the grid connection rather than relying entirely on power delivered through the public network.
Reuters reported in late September that developers were increasingly using smaller gas turbines because they can be deployed faster than large plants and can bypass some grid-connection delays.[7]
This can reduce pressure on some grid investments.
It does not make the cost question disappear.
The project still has fuel, emissions, local infrastructure, reliability and possibly backup/grid-service requirements.
Original Asset 5: The Cost Allocation Receipt
When you read a new AI data-center announcement, imagine a receipt with these lines:
| Cost line | Question |
|---|---|
| Direct facilities | Which assets exist mainly for this customer? |
| Transmission | How much is dedicated and how much creates wider system value? |
| Generation / capacity | What new supply must be procured because of the load? |
| Studies / connection | Who pays the planning and engineering costs? |
| Reservation | What financial commitment backs the requested MW? |
| Exit risk | Who pays if the customer leaves or downsizes? |
| General rate base | What amount is ultimately spread across other customers? |
That receipt is much more useful than simply asking whether “Big Tech” or “the public” pays.
The Household-Level Question
A household does not need to understand every transmission tariff.
The useful question is:
Which part of my bill is paying for infrastructure that benefits the system generally, and which part exists mainly because a new large customer arrived?
That is why transparency in rate design matters.
What to Watch Next
- PJM cost allocation. How will the regional process ultimately assign backstop and new-load-driven costs?
- AEP Ohio tranche two. How much new demand is willing to accept current collateral and minimum-demand rules?
- Transfer / exit rules. Do utilities increasingly create markets or mechanisms for unused reserved capacity?
- Behind-the-meter power. Does private generation reduce grid-delay risk without creating larger local costs?
- Household-rate evidence. Do regulators publish clearer breakdowns of how much large-load growth changes customer bills?
- Shared-benefit tests. Do commissions develop more explicit methods for separating dedicated from system-wide value?
The Simple Idea to Remember
There is no single correct payer for every AI grid upgrade.
There is a sequence of questions.
What cost?
Who caused it?
Who benefits?
Who carries the risk if the load disappears?
Cost causation is not “make one company pay everything.” It is matching cost, benefit and risk as closely as possible.
Key Vocabulary
cost causation
The principle that customers who cause new system costs should bear an appropriate share of those costs.
rate base
Regulated utility investment on which the utility is generally allowed to earn a return and recover costs through customer rates.
minimum-demand charge
A billing rule that requires a large customer to pay for a minimum amount of reserved demand even if actual usage is lower.
collateral
Financial security that protects the utility and other customers if a large-load project fails to meet its commitment.
stranded cost
Investment that becomes underused or unrecoverable because the expected customer or demand does not materialize.
behind-the-meter generation
Generation located on the customer's side of the grid connection and used primarily to serve that site.
Related Articles
- The AI Power Boom Has a Forecasting Problem: “Ghost” Data Center Demand
- AI Data Centers Need More Than Power. They Need Permission.
- Cheap Power Is Not Always Cheap: The Grid Delay Behind AI Data Center Economics
- What Must Be Built to Power the AI Data Center Boom?
Sources
- Microsoft — Building Community-First AI Infrastructure, January 13, 2026.
- AEP Ohio — Data Center Tariff, checked October 4, 2026.
- AEP Ohio — Data-center load update, February 13, 2026.
- Reuters — Google signs AES, Xcel supply deals to meet data-center energy needs, February 24, 2026.
- FERC — Commissioner See's concurrence in PJM Reliability Backstop Procurement, September 29, 2026.
- Reuters — FERC asks PJM to revise plan to shield homes from data-center costs, September 30, 2026.
- Reuters — Small gas turbines and behind-the-meter data-center power, September 29, 2026.
Update History
October 4, 2026 — Updated with current AEP Ohio tariff mechanics, FERC/PJM cost-allocation developments, and new cost-causation / stranded-risk frameworks.
Sources checked through October 4, 2026. Electricity cost allocation varies by utility, state, regional market, tariff and project. The article explains analytical principles rather than a universal rate rule.