California wants data centres to pay their way

California signed seven data-centre bills on September 21, bringing electricity costs, water use and local approvals further into the economics of the AI buildout. The governor's announcement presents the package as a way to stop infrastructure costs being passed to households and give communities more information about proposed projects.

That is a sensible direction. A data centre should be able to explain who will pay for the infrastructure it needs before anyone treats its announced capacity as a dependable addition to supply. Leaving those costs unresolved makes a development look easier to finance than it may prove to be.

One of the new laws, SB 1168, gives the California Public Utilities Commission a specific job: assess rate structures that make data centres bear a reasonable share of transmission and distribution costs and the resources needed to serve their demand. It does not announce a single new electricity price for every facility. The consequential work will happen in how regulators turn that instruction into terms operators can price, negotiate and finance.

A chip order does not secure a working site

The commercial attraction of an AI data centre rests on its ability to sell usable computing capacity. A machine awaiting power cannot do that, however advanced the hardware or credible the customer.

This creates a problem with the way infrastructure announcements are often read. A proposed site's capacity describes an ambition. It does not, by itself, establish when a customer can run work there or the full cost of making that possible. The distinction becomes important when several expensive commitments have to arrive together: equipment, a completed building and a reliable electricity connection.

Our view is that an AI infrastructure investment case should identify which of those commitments is least secure. A project with an agreed route to power and a credible completion date can be more useful than a larger proposal with a cheaper headline cost. The value lies partly in the work a customer can actually start, and when.

This is also why the allocation of grid costs matters before construction. Imagine a utility building additional infrastructure for a planned customer. If the customer's demand arrives later or at a lower level than expected, the infrastructure still has to be paid for. A contract that assigns that risk clearly is worth more than an optimistic demand forecast that leaves the question open.

That example is an economic test, not a claim about a particular California development. It shows what an investor or customer should be able to establish before relying on an announced completion date.

More demand does not mean every power bill rises

The International Energy Agency's April 2026 analysis estimates that global data-centre electricity consumption reached 485 terawatt-hours in 2025. Its central projection rises to 950 TWh in 2030. The chart below shows those two endpoints, including ordinary data-centre workloads as well as AI. The 2030 value is a projection, not contracted demand.

The IEA also makes a distinction that gets lost in much of the argument about AI and electricity. Additional demand can improve the use of existing infrastructure where spare capacity is available. In a constrained system, it can require expensive new assets. The effect on prices depends on those conditions and on who is required to pay.

That is a better basis for policy than assuming that every data centre is either a burden or a public benefit. The relevant questions belong to the proposed site. What additional equipment is needed? Who finances it? What payment remains due if the customer uses less electricity than expected?

A developer can reasonably ask for predictable answers. Households can reasonably ask not to guarantee an investment whose upside belongs elsewhere. A workable tariff has to address both. Rules that make a customer bear costs it causes can improve the quality of the investment decision, because the project has to support more of its own economics.

There is a limit to this argument. Regulators still have to distinguish costs caused by a new facility from upgrades that would have been needed anyway. Charging one customer for unrelated work would make the calculation less accurate. The aim should be a defensible allocation that survives scrutiny from both sides.

The opportunity is in getting projects built

For companies selling into the AI buildout, this puts a premium on work that helps a customer reach operation. A useful product might establish a more credible load forecast, reveal unused capacity, or help demonstrate how a facility will behave under a utility's operating conditions. Its value would come from a better project decision or a shorter, more dependable route to service.

The purchase case needs to be specific. A dashboard that displays an energy number may be easy to demonstrate and hard to justify. A system that supplies evidence a utility or project financier will accept has a clearer role. The buyer needs to know which decision it changes, who accepts its output and what remains to be resolved after it is installed.

The same discipline applies to claims about efficiency. A lower energy requirement for one computing task is useful. Turning it into a saving for a facility requires assumptions about workload, utilisation and the customer's service requirements. Those assumptions should be visible. Otherwise, a precise-looking calculation can conceal the part of the business case that needs the most attention.

California's package brings another practical obligation into view: water. The governor says the measures require more information on expected use, supply, efficiency and drought planning, with developers responsible for necessary supply upgrades. For an operator, that means the proposed site's relationship with its local infrastructure deserves attention alongside the technology inside the building.

Judge the next announcement by its commitments

Making infrastructure costs explicit is good for the credibility of the AI buildout. It gives communities a clearer basis for evaluating a development and forces projects to compete on economics that are closer to the cost of delivering them.

Implementation will determine how useful California's approach becomes. An assessment process that never produces predictable terms would leave developers with another uncertainty. Clear responsibilities and a credible timetable would give them something they can work with.

For the next large data-centre announcement, our first question is who has committed to supplying the power, on what terms, and by when. The headline capacity becomes much more informative once those answers are available.