AI’s Power Problem Is Also a Timing Problem

The missing variable is time

AI’s electricity debate usually begins with scale: larger data centers, more accelerators and the power stations needed to keep them running. But an electricity system also has a clock. A kilowatt demanded during a stressed evening can be harder to accommodate than the same kilowatt used when spare capacity is available. That difference creates an opportunity for the AI industry to improve how existing infrastructure is used while new supply is being built.

Google’s October 22 announcement with Voltus puts that opportunity in focus. Google describes an expanded agreement to build a one-gigawatt virtual power plant in the PJM region, bringing together flexible equipment in homes and businesses. The announcement concerns a planned capacity portfolio, not a newly completed conventional generator. Its significance is the business model: a large technology company is funding flexibility elsewhere on the grid. Understanding that model requires separating electricity production, reductions in peak demand and changes to when computing happens. Each can help, but they solve different problems.

A power plant made from many smaller decisions

A virtual power plant coordinates resources that would otherwise operate separately. A battery can discharge at a useful moment. A building can adjust cooling within an agreed comfort range. An industrial customer can reschedule a suitable process. Aggregated and dispatched together, these actions can provide a service that the electricity system can plan around. The underlying equipment remains physical; the virtual part is the coordination and commercial arrangement connecting it.

Voltus says the expanded Google-funded program will recruit capacity from local homes and businesses and use assets including batteries, smart thermostats and flexible industrial processes. Participating customers are intended to receive financial benefits through payments and bill savings. The earlier agreement, announced in June, covered up to 100 megawatts over three years. The expansion indicates a larger commitment, but announcement language should not be mistaken for proof that every resource has enrolled, passed testing and begun delivering.

That distinction matters because a portfolio is assembled one customer and one device at a time. A promising location still needs an eligible resource, a workable control arrangement and a participant willing to accept the operating conditions. The useful question is therefore how much dependable capacity becomes available, by what date, and on what terms. A large headline number is the beginning of that inquiry.

The idea itself predates the current AI boom. The US Department of Energy’s 2025 virtual-power-plant update described commercially available deployments and identified enrollment, operational standardization and integration with utility planning and markets as important scaling challenges. AI-driven electricity demand gives those existing tools a new buyer and a stronger reason to expand.

Two kinds of flexibility, with different owners

It is easy to confuse this arrangement with a data center turning down its own power. The Voltus announcement describes flexibility aggregated from other customers. Google has also pursued a separate approach inside its computing operations. In March, it said it had incorporated one gigawatt of data-center demand-response capacity into long-term contracts with utility partners. That capability limits or shifts a portion of machine-learning workloads during selected periods. These are different programs and should not be combined casually into a single claim about operating capacity.

The distinction changes who must do what. In a third-party portfolio, the aggregator needs to coordinate participants and demonstrate their response. In a computing program, the operator must decide which jobs can wait or change execution patterns without breaking its service commitments. One approach pays for flexibility outside the data center; the other changes demand within it. They may complement each other, but neither automatically establishes the performance of the other.

Google’s August 2025 account of its data-center work described earlier flexibility for non-urgent tasks, such as processing a YouTube video, and a demonstration that reduced machine-learning-related power demand during three grid events with Omaha Public Power District. It also emphasized limits: essential online services and customer obligations constrain what can be moved, and availability depends on location.

For readers evaluating an AI infrastructure announcement, this is a useful first filter. Ask where the promised adjustment physically occurs. A household battery, an industrial process and a training cluster may all respond to the same system need, but their constraints, incentives and failure modes are different.

Peak capacity and annual energy answer different questions

A gigawatt is a measure of power, not a promise of continuous electricity production. To understand a flexible resource, the duration of its response matters as much as its maximum size. So do advance notice, the number of events it can support and the time needed to recover. A battery that delivers briefly and a process that can pause for several hours belong in different operating plans, even if their advertised power ratings match.

Consider a simple illustration rather than a forecast for this agreement. If a load shifts two hours of electricity use from a constrained period to a quieter one, it may reduce the peak while leaving total daily energy consumption broadly unchanged. If a battery supplies that period, it must have been charged and will eventually need charging again. If an industrial process is deferred, its operator may later catch up. The next peak matters too.

PJM’s explanation of demand response describes a voluntary program in which customers are compensated through participating service providers for qualifying reductions when requested. Ordinary changes in a customer’s normal operating pattern do not count as demand response. The principle is straightforward: the system needs a real, attributable change, rather than a payment for something that would have happened anyway.

This also limits the environmental conclusion. A reduction during a particular hour does not, on its own, establish a reduction in annual emissions. That assessment depends on the energy used before and after the event, the equipment involved and the electricity displaced. Peak management, efficiency and decarbonization deserve separate measurements.

The cheapest flexibility may be outside the server room

For an AI operator, leaving an expensive accelerator idle can have a meaningful opportunity cost. A job might be technically interruptible yet commercially awkward to delay. Training schedules, customer deadlines and the availability of other resources all affect the decision. The energy bill is only one part of the calculation; the value of the work displaced is another.

The International Energy Agency’s October 14 report on integrating data centers into electricity systems highlights this tension. It describes active management of computing loads as still uncommon and notes that operators serving third-party customers may have limited visibility or control over those workloads. The report also identifies a reason some operators explore flexibility despite its cost: it may enable faster grid connections. That is a potential benefit under specific arrangements, not a guarantee for every project.

Seen in that light, buying flexibility from other customers can be economically sensible. A factory with an adaptable schedule or a building with a suitable battery may be able to provide a grid service at a lower total cost than pausing high-value computing. The comparison must include payments, equipment, administration and the effect on each participant, rather than assuming that shifting demand is free.

This is also why flexibility should be treated as a product with defined terms. How much notice is required? How long can the adjustment last? Who has the right to opt out, and what happens then? The answers determine whether the buyer has acquired a dependable capability or merely a list of possible responses. Software can coordinate those commitments, but it cannot erase them.

What a credible performance report would show

The next useful milestone for the Google–Voltus program is evidence of delivery. An informative report would separate contracted capacity, enrolled resources, tested capability and actual event performance. Those figures need not be identical. Reporting them separately would make it easier to see whether the bottleneck is customer recruitment, installation, verification or reliable dispatch.

A second requirement is a convincing baseline. To claim a reduction, an operator needs a defensible estimate of what electricity use would otherwise have been. Weather, production schedules and building occupancy can all change demand independently of an event. A performance claim becomes much more useful when its measurement method, uncertainty and treatment of unusual days are explained.

The portfolio also needs to work under the conditions that make it valuable. A test on an ordinary afternoon may not establish what happens during prolonged heat, equipment outages or repeated requests. Resources that appear independent in a spreadsheet can share a constraint: similar weather, similar customer schedules or the same local network. A credible evaluation should examine those relationships rather than treating every enrolled device as interchangeable.

Participant experience belongs in the same report. Payments can be attractive, but a program also needs to make sense after enrollment. Households need clear control over comfort and participation. Businesses need confidence that flexibility will not disrupt their main activity. Transparent terms, understandable event records and a workable way to resolve errors would help make participation durable.

These are practical tests to apply to the program, not allegations that it has failed them. The public announcements establish the intended arrangement and its ambition. They do not yet provide enough operating detail to calculate a verified net saving per participant or an independently demonstrated reduction in system costs.

The local grid remains the unit that matters

Annual electricity statistics can make the challenge look smoother than it is. The IEA’s 2026 outlook projects global data-center electricity consumption rising from about 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. That is a projection across many facilities and services, not a reading of AI alone. A particular connection can still be constrained even when the global total appears manageable, because electricity must reach the right location at the right time.

A portfolio across a large grid region therefore needs to be assessed against the actual services it provides. Reducing a regional peak and relieving a specific local constraint are related possibilities, but they are not automatically equivalent. Location, network conditions and the rules governing the service determine what an adjustment can accomplish. Procurement should follow that physical reality.

There is a similar distinction between participant benefits and benefits for everyone else. A direct payment to an enrolled business is observable. A claim that all customers’ bills will fall requires a wider analysis of avoided costs, program costs and who pays for new infrastructure. It should not be inferred from the existence of a customer incentive alone.

For technology companies, this makes energy strategy increasingly operational. Choosing a site, deciding how quickly to ramp computing demand and negotiating what happens during constrained periods can influence the infrastructure needed to serve a project. The opportunity is to make those decisions together, early enough that flexibility can affect the design, rather than adding it later as a communications exercise.

Build more, and use the available system better

Demand flexibility will not remove the need for additional generation, storage and networks. Nor does this announcement show that AI’s electricity requirements have stopped growing. It points to a narrower, useful possibility: some of the pressure can be managed by changing the timing and coordination of demand, instead of treating every new load as equally inflexible.

The commercial lesson extends beyond one company. An AI business may have more than one way to obtain room to grow: improve the efficiency of its computing, arrange a more adaptable operating schedule, support new supply or fund flexibility elsewhere. These options have different costs and constraints. A serious strategy compares them against the service the business actually needs and the grid conditions it actually faces.

The Google–Voltus expansion is worth watching for that reason. The next chapter should be judged through enrolled resources, verified response, participant economics and the infrastructure benefits that can be demonstrated. If those pieces come together, the program could illustrate how technology demand finances useful capabilities beyond the data-center fence. The important outcome would be a more dependable relationship between computing and the electricity system, measured hour by hour rather than only in headline gigawatts.

• The AI Slowdown Isn’t About Intelligence. It’s About Economics.
• The Hidden Cost of GenAI: Why "Tokenomics" is Breaking Enterprise Budgets