Skip to main content

California AI Data Centers: Who Pays for Grid Expansion Costs?

MONEY TRACES | AMERICA'S STRATEGIC ASSETS

California AI Data Centers: Who Pays for the Grid Expansion When AI Demand Comes First?

California's AI Power Problem Is Not Electricity. It's Who Owns the Risk When the Grid Gets Built First.

Published: September 5, 2026  ·  By: Hossam Seif

Artificial intelligence has created a new infrastructure race.

But the biggest financial question behind the AI boom may not be how much electricity data centers will consume. It may be who carries the cost of preparing the grid for demand that can change before the infrastructure does.

Who carries the financial risk when electricity infrastructure has to be built before the future demand it was designed to serve is certain?

California's data center debate is usually framed as an electricity problem. How much power will AI require? Can the grid support new facilities? Will enough generation and transmission capacity be available?

Those questions matter. But California's regulators and energy agencies are already confronting a second problem: the financial exposure created when large, concentrated loads require major infrastructure investment.

The California Energy Commission says data centers accounted for about 1,000 megawatts of California ISO peak demand in early 2026, or roughly 2% of peak demand. Its planning forecast projects about 4,500 MW, or 9% of peak demand, by 2040. The same agency says it began incorporating projected data-center growth into its statewide demand forecast in 2024 using information from utility applications. 0

That creates the central mismatch behind the story: utilities and grid planners have to make decisions years before the final shape of AI demand is known.

The financial question is therefore not simply whether California needs more electricity infrastructure. It is who owns the downside if the infrastructure is built for demand that arrives later, arrives smaller than expected, or does not arrive at all.

The Hidden Financial Chain Behind AI Electricity Demand

The connection between an AI company and an electricity bill is not direct. A data center does not simply appear and immediately transfer its construction costs to a household. The money moves through a chain of planning, interconnection, infrastructure investment, financing, regulation and eventually rates.

AI expansion plans → Large electricity requests → Grid investment → Cost allocation → Revenue recovery → Customer exposure

Every step can change who carries the financial risk.

First, AI Creates a Concentrated New Load

Large data centers are unusual electricity customers because their demand can be enormous, concentrated in one location and dependent on high reliability.

California's Public Advocates Office says developers increasingly propose facilities exceeding 50 to 100 MW at a single site. It also says data centers can consume 10 to 50 times more energy per square foot than a typical commercial office building. 1

A customer of that size can require infrastructure that would not otherwise be needed at the same time, scale or location.

Then the Grid Has to Move Before the Customer Does

Transmission and other grid projects do not operate on the same timetable as an AI expansion plan. The California Energy Commission uses utility application information in its demand forecasting, and those forecasts help inform transmission and other system planning. 2

That means an electricity request can have financial consequences before a data center is fully built and operating.

The Public Advocates Office has warned that speculative or duplicate interconnection requests can inflate forecasts and create a risk of infrastructure being built for projects that are delayed, withdrawn or never fully materialize. 3

Finally, Someone Has to Carry the Exposure

This is where the electricity story becomes a capital-allocation story.

If a large customer pays all of the infrastructure cost created specifically for its connection, the customer carries more of the risk. If some costs are recovered through broader utility rates, other customers can become exposed. If an infrastructure project serves multiple customers and produces wider system benefits, the allocation becomes more complicated.

The important distinction is between paying for electricity that a customer actually consumes and paying for infrastructure whose economics depend on future demand.

California Already Has a Real-World Test Case

This is not merely a hypothetical argument about what might happen someday.

2.5 GW
Concentrated load growth
CAISO approved a South Bay transmission-upgrade cluster intended largely to serve 2.5 GW of concentrated data-center and electrification load growth between 2026 and 2039.
>$2B
Transmission upgrades
The Public Advocates Office says the cost of that South Bay upgrade cluster exceeds $2 billion.

Those figures come from California's own ratepayer watchdog, which has used the South Bay example to illustrate why large-load growth creates a cost-allocation problem. 4

The point is not that the entire $2 billion should be assigned to AI companies. The underlying projects serve a defined set of expected load growth, and the infrastructure can have broader system effects.

The point is that once billions of dollars of grid investment are committed, the financial obligation does not disappear simply because an individual forecast changes.

The Rule That Matters: Cost Should Follow Cause

At the center of the debate is a basic regulatory principle: costs should be allocated in a way that reflects who causes them and who benefits from the resulting infrastructure.

That sounds simple until the infrastructure has multiple users and the future is uncertain.

A transmission line may be justified partly by a data center and partly by broader system growth. A substation may be needed for a particular large customer but later become useful to other customers. A forecast may identify future demand that ultimately arrives at a different location or at a different scale.

So the real policy question is not whether data centers should pay for everything. It is whether the rules can distinguish between:

  • infrastructure uniquely required by a large new customer,
  • infrastructure that produces broader system benefits,
  • infrastructure built on speculative demand, and
  • infrastructure whose costs remain after expected demand changes.

The Little Hoover Commission's March 2026 report specifically examines how California should address data-center energy demand through rates, financing mechanisms and regulatory approaches. 5

California's Current Approach Shows Where the Risk Sits

California has already moved toward making large customers responsible for some of the upfront infrastructure burden.

In July 2025, the California Public Utilities Commission approved an interim version of PG&E Electric Rule 30 for high-energy transmission-level customers such as AI data centers. Under the interim arrangement, applicants must agree to pay for necessary transmission infrastructure work upfront. The CPUC said the future treatment of potential refunds and cost allocation would be determined in a later decision. 6

That is important because it changes the financial sequence.

  1. Demand request A large customer seeks transmission-level service.
  2. Infrastructure requirement The utility identifies transmission work necessary to provide service.
  3. Upfront exposure Under the interim Rule 30 framework, the applicant agrees to fund necessary transmission work upfront.
  4. Future allocation The eventual treatment of refunds and cost allocation remains subject to later regulatory action.

The significance is larger than one utility rule. It shows that the regulatory system is already trying to prevent the entire initial infrastructure risk from automatically sitting with existing ratepayers.

But it also shows why the problem is unresolved. Paying upfront for an interconnection is not the same thing as determining who ultimately bears every system-wide transmission or generation cost associated with long-term load growth.

The Biggest Risk Is Not Construction. It Is Stranded Infrastructure.

The most important financial risk appears when infrastructure survives a demand forecast that does not.

Consider the sequence:

  1. Forecast Planners incorporate expected data-center growth into future electricity demand.
  2. Investment decision Transmission, substations or other infrastructure are planned and approved.
  3. Construction Capital is committed and physical infrastructure is built.
  4. Demand changes A project is delayed, canceled, downsized or uses less electricity than projected.
  5. The asset remains The infrastructure and its associated financial obligations do not disappear simply because the forecast changed.

The Public Advocates Office explicitly identifies this risk. It says that if data-center projects are canceled, withdrawn from interconnection, use less energy than projected or terminate service early, the infrastructure costs incurred to serve them can remain and potentially be passed to existing ratepayers. 7

That is the stranded-asset problem in its most important form for this story.

The uncertainty is not whether AI will require electricity. The uncertainty is whether today's infrastructure assumptions will match tomorrow's actual demand.

The Three-Way Financial Conflict Behind the AI Grid

Stakeholder What It Wants Primary Financial Exposure
AI / data-center developers Fast, reliable access to large amounts of electricity Upfront infrastructure costs, connection charges and potentially longer-term capacity obligations
Utilities Reliable demand and sufficient revenue to support infrastructure investment Underused infrastructure if expected load fails to materialize
Existing ratepayers Reliable electricity without inheriting unnecessary costs Potential exposure to costs that remain after large-load forecasts change

No participant can eliminate uncertainty. The regulatory framework determines where that uncertainty lands.

The Strongest Case Against the Thesis

There is a legitimate argument on the other side.

Large data centers can bring new electricity demand and can contribute revenue to the utility system. Infrastructure built to serve them may eventually support additional customers or broader economic activity. And California cannot simply refuse to build infrastructure because forecasts contain uncertainty.

The California Energy Commission is explicitly refining its demand-forecasting process as it receives more information about proposed data centers and their likelihood of completion. That is evidence that the state is not treating every request as a guaranteed future load. 8

There is also a practical argument for moving quickly. If infrastructure waits until every data center is fully certain, California could make it harder for projects that genuinely will proceed to obtain electricity on the timeline required by the technology industry.

So the strongest counter-thesis is not that the risk is imaginary.

It is that a well-designed regulatory system can manage the risk while still allowing useful infrastructure to be built.

That argument survives the evidence.

What does not survive the evidence is the assumption that the risk disappears simply because the data center industry is growing.

The Problem Is Bigger Than One Data Center

The financial exposure can spread beyond the individual customer because grid infrastructure is a shared system.

The Public Advocates Office says major transmission upgrades within the CAISO system can be socialized across ratepayers, creating a potential pathway through which infrastructure built partly in response to large new loads can affect customers beyond the data center itself. 9

That does not mean every data-center-related investment will raise household electricity bills. It means the possibility is determined by the rules governing cost recovery, the actual use of the infrastructure, the degree to which large customers cover their associated costs, and whether broader system benefits justify broader allocation.

The Money Traces Question
When AI creates the demand, who owns the infrastructure risk if the demand changes?

That is the question hidden underneath the much louder debate about how many gigawatts AI will consume.

What Happens If AI Demand Arrives Faster Than Expected?

The opposite scenario matters too.

If AI demand grows faster than planners expected, California could face the economic cost of not having enough transmission, generation or interconnection capacity ready when customers need it.

In that case, the risk is not stranded infrastructure. It is constrained growth.

Data centers may wait for connections. Developers may shift projects to other states. Utilities may have to accelerate capital spending. And companies building AI infrastructure may place a premium on locations where power can be secured faster.

This is why the financial design cannot simply be "make the data center pay."

If the rules become so restrictive that useful projects cannot obtain service, California can lose economic activity and investment opportunities. If the rules are too permissive, existing customers can inherit infrastructure risk they did not create.

The system has to solve both problems simultaneously.

The Next AI Infrastructure Battle Will Be About Rules, Not Just Growth

The public conversation has focused heavily on electricity consumption. The next phase is more likely to focus on the rules that decide who finances and recovers the cost of serving that consumption.

California's own institutions are already moving in that direction:

  • The CEC is refining demand forecasts using utility application data.
  • The CPUC has approved an interim framework requiring certain large transmission-level customers to fund necessary transmission work upfront.
  • The Public Advocates Office is pushing for safeguards against shifting large-load infrastructure costs to existing customers.
  • The Little Hoover Commission is examining rates, financing mechanisms and regulatory approaches for the data-center buildout.

These are not separate debates. They are different parts of the same financial system.

Forecasting determines what infrastructure appears necessary. Interconnection rules determine who pays at the front end. Rate design determines how costs are ultimately recovered. Regulatory decisions determine how much risk can be shifted between participants.

The electricity system is therefore becoming a financial allocation system for the AI economy.

The Final Ledger: AI Runs on Electricity. Electricity Runs on Risk Allocation.

California's AI infrastructure race is not simply a question of whether the state can generate enough electricity.

It is a question of timing.

Demand can be forecast before it is proven. Infrastructure can be approved before demand is fully realized. Capital can be committed before the economics are certain. And once physical infrastructure is built, its cost does not automatically disappear when the underlying forecast changes.

The California Energy Commission's forecast illustrates the scale of the issue: data centers were already about 2% of California ISO peak demand in early 2026, with the state's planning forecast putting them at about 9% by 2040. 10

But the more important number is not a forecast.

It is the amount of infrastructure capital that ultimately remains supported by actual, durable electricity demand.

The AI power race will be measured in megawatts. Its financial outcome will be measured in who owns the risk when the forecast meets reality.

WHAT WE KNOW: California is planning for substantial data-center electricity growth, and large new loads can require major grid investments.

WHAT REMAINS UNCERTAIN: The final scale, location and timing of AI demand are not yet fixed, and individual projects can be delayed, downsized, withdrawn or canceled.

WHERE THE MONEY GOES: Into transmission, interconnection and other infrastructure needed to serve large loads, with the eventual cost recovery determined by regulatory and utility rules.

WHO BENEFITS: AI developers and the broader economy if reliable electricity enables productive new investment and the infrastructure is efficiently utilized.

WHO BEARS THE RISK: Potentially the large customer, the utility, investors, or broader electricity customers depending on the infrastructure, tariff and cost-allocation structure.

WHAT TO WATCH NEXT: California's decisions on large-load tariffs, upfront infrastructure payments, refunds, minimum demand commitments, transmission planning and the treatment of projects that do not materialize as forecast.

The AI boom will not settle the question by itself.

The rules will.

How We Read the Numbers

This investigation separates electricity demand forecasts from realized demand and separates infrastructure costs from the eventual method used to recover those costs.

The 1,000 MW and 4,500 MW figures are California Energy Commission planning figures for data-center demand, not measured statewide electricity consumption attributable exclusively to AI. The 2.5 GW and more-than-$2 billion figures refer to a specific South Bay transmission-upgrade cluster described by the Public Advocates Office. They are not presented as a statewide total for AI infrastructure.

Likewise, the possibility of stranded infrastructure is treated as a documented regulatory risk, not as a prediction that California ratepayers will necessarily absorb every data-center-related investment.

Sources & Verification

  1. California Energy Commission — Data Centers. Source for California's data-center demand figures, forecasting methodology and the state's use of utility application data in demand forecasting.
  2. Little Hoover Commission — Data Centers and California Electricity Policy, Report #292, March 2026. Source for the Commission's examination of rates, financing mechanisms and regulatory approaches for California's data-center electricity demand.
  3. California Public Utilities Commission — Electric Rule 30 for High-Energy Users, July 24, 2025. Source for the interim requirement that qualifying applicants agree to pay for necessary transmission infrastructure work upfront and for the CPUC's statement that future refund and cost-allocation treatment would be addressed later.
  4. California Public Advocates Office — How Will Data Center Growth Impact California Ratepayers?, October 28, 2025. Source for the 50–100 MW-plus large-load discussion, the 2.5 GW South Bay load-growth example, the more-than-$2 billion transmission-upgrade figure and the documented ratepayer-risk analysis.
  5. California Public Utilities Commission — Electric Transmission Rates and FERC Proceedings. Source for California's ongoing transmission-cost and large-load regulatory context.
Editorial Note: Money Traces does not treat a forecast, an interconnection request, an infrastructure approval and realized electricity demand as interchangeable facts. Where the public record establishes a risk rather than an outcome, this article describes it as a risk. The central question is not whether California should build for AI. It is how the financial exposure created by that buildout should be allocated when future demand remains uncertain.
Written and edited by Hossam Seif, founder of Money Traces.

Comments

Popular posts from this blog

$9.49 Trillion Was Pledged. Where Is the Money?

MONEY TRACES | FINANCIAL GEOPOLITICS $9.49 Trillion Was Pledged. Where Is the Money? The White House headline is enormous. The amount currently tied to named U.S. projects is dramatically smaller. Published: August 23, 2026  ·  By: Hossam Seif The Trump administration says its trade diplomacy has secured $9.49 trillion in new U.S. investment and related economic commitments through its "America First" trade and investment deals. A detailed reclassification published by Steptoe on August 19 tells a much narrower story. Steptoe's line-by-line review found that only about $4.12 trillion of the headline figure can be classified as U.S.-directed investment language — and even that number contains nonbinding, project-contingent, and potentially overlapping commitments. Only $217.7 bi...

Village Roadshow Bankruptcy: How 'Secured' Creditors Got 3.5 Cents

Money Traces · Financial Reconstruction You know the logo. But would you ever expect that logo to sit above a bankruptcy case where a class of “Senior Secured” creditors was projected to recover just 3.5 cents for every dollar of its allowed claim? Village Roadshow helped finance more than 100 films whose combined worldwide box-office receipts exceeded $19 billion. The company was behind films including The Matrix , Joker , Ocean’s Eleven , Mad Max: Fury Road and Wonka . Then, on March 17, 2025, Village Roadshow filed for Chapter 11 bankruptcy protection in Delaware. Its assets were hardly worthless. Alcon ultimately agreed to pay $417.5 million for the company’s 108-film library, while separate transactions brought aggregate headline consideration to about $440.25 million plus assumed liabilities . So what happened? The answer is not that the movies suddenly stopped maki...

Who Is Really Financing the AI Boom? Inside Wall Street’s $500 Billion Risk Engine

MONEY TRACES — FINANCIAL INTELLIGENCE THE BOTTOM LINE The Scale The AI infrastructure buildout is creating a financing demand measured in hundreds of billions of dollars, pushing capital beyond traditional corporate balance sheets and into private financing platforms, structured leases, and asset-backed credit. The Collateral Specialized lenders are financing AI infrastructure with physical compute assets, including NVIDIA GPU deployments, creating a credit question that did not exist at this scale before: how durable is the collateral when computing economics change so quickly? The Exposure Private credit is increasingly accessible to institutional and retirement markets. That does not mean 401(k) investors are directly financing GPUs—but it does create a potential channel through which AI infrastructure risk can enter diversified portfolios. A $500 billion financing pipe...