AI data center electricity consumption is financed by energy producers, cloud operators, and, indirectly, their customers. The Google-Constellation accord announced in 2026 illustrates this shared responsibility: Constellation will invest more than $4.3 billion, while Google secures long-term purchases. For a company, the practical impact is mainly reflected in the price of AI resources, their regional availability, and contractual commitments.
Why do AI data centers consume so much electricity?
An AI data center’s electricity consumption primarily comes from computing accelerators, their cooling systems, and the equipment that distributes power. According to the International Energy Agency, data centers consumed approximately 415 TWh worldwide in 2024, or 1.5 % of global electricity demand.
An accelerator is a specialized chip, such as a GPU or TPU, that performs the computations required for artificial intelligence. Training a model uses substantial power over a limited period. Inference—that is, the daily use of the model to produce a response, image, or analysis—repeats these computations on a large scale.
The International Energy Agency projects global consumption of approximately 945 TWh in 2030, or slightly less than 3 % of electricity demand. This average conceals a more immediate challenge: capacity is concentrated in a few areas, where it can place significant pressure on the local grid.
In the United States, data centers consumed approximately 176 TWh in 2023, or 4.4 % of the country’s electricity. Berkeley Lab estimates that they could reach 325 to 580 TWh by 2028, representing between 6.7 % and 12 % of U.S. consumption. This wide range reflects uncertainty about the pace of AI deployment and efficiency gains.
What does the 3.59 GW accord between Google and Constellation actually finance?
The Google-Constellation accord announced on October 6, 2026 covers 3,590 MW on the PJM grid, but only 890 MW correspond to planned additional nuclear capacity. The remaining 2,700 MW come from facilities already in operation. Presenting the 3.59 GW as entirely new electricity generation would therefore be incorrect.
A PPA, or long-term power purchase agreement, sets the terms under which a buyer commits to purchasing energy. Google entered into a 20-year PPA corvering the additional 890 MW, as well as a 15-year supply accord covering 2,700 MW of existing generation. The commercial prices and total amount paid by Google were not disclosed in 2026.
| Perimeter | Capacity | Duration | Type of electricity | Announced availability |
|---|---|---|---|---|
| Increase in nuclear capacity | 890 MW | 20 years | Additional generation from 11 existing units | First commissioning expected in 2028 |
| PJM supply | 2,700 MW | 15 years | Generation already in operation, with no precisely allocated new source | Existing capacity in 2026 |
| Total under contract | 3,590 MW | Two contracts | Approximately 24.8 % of additional capacity | Not fully available in 2026 |
The 890 MW do not come from new reactors. Constellation plans equipment and efficiency improrements at 11 existing nuclear units in Illinois, Pennsylvania, and New Jersey. The first increase is expected to enter service in 2028, followed by the others progressively.
Constellation announced an investment of more than $4.3 billion for this program. Google apporte benefits from commercial visibility through its long-term commitment, reducing the producer’s financial risk. In practice, the energy operator finances the work, while the cloud buyer contributes to its viability through future purchases.
Who will pay for the increase in AI data center electricity consumption?
The cost of AI data center electricity consumption will be distributed among producers, cloud operators, grid managers, and business customers. Long-term contracts finance certain capacity, but no source published in 2026 demonstrates that the Google-Constellation accord will automatically lead to an increase or decrease in the price of inference on Google Cloud.
Part of the financing appears in energy companies’ investments. Another portion is absorbed by cloud providers before being incorporated, according to their commercial decisions, into the prices of GPUs, TPUs, storage, or usage-based AI services. Grid infrastructure may also require reinforcements, whose financing depends on the local regulatory framework.
The trap is confusing the price of electricity with the cloud bill. The price of an AI request also depends on the model used, the number of tokens processed, the service level, the hardware deployed, and server utilization. A smaller model, correctly suited to the need, can reduce the bill without changing energy providers.
In the projects we lead, we often see teams monitoring a model’s unit price while allowing unnecessary processing, test environments, or overly long requests to continue running. At this stage, it is better to measure usage before negotiating a multiyear commitment. A structured adoption of AI in the enterprise helps distinguish profitable use cases from ongoing experimentation.
Does the choice of cloud region change cost and carbon footprint?
The choice of a Google Cloud region affects accelerator availability, service pricing, latency, and the carbon intensity of electricity. In 2025, Google Cloud reported 96 % hourly carbon-free energy and 16 gCOâ‚‚e/kWh in Paris, compared with 57 % and 343 gCOâ‚‚e/kWh in Northern Virginia.
A cloud region is a geographic area containing multiple data centers. Choosing Paris can reduce the distance to French users and facilitate certain data localization decisions. However, it does not guarantee the presence of the desired GPU, the lowest price, or immediate capacity.
Google Cloud states in 2026 that accelerator prices and availability vary by region and zone. GPU commitments can accorder discounts of up to 37 % for one year or 55 % for three years, but capacity remains tied to a specific region. A high discount becomes costly if the application changes technology or actual demand remains below forecasts.
The choice should therefore begin with the product’s constraints. An interactive application will often prioritize latency and availability, while a deferrable workload can be shifted in time or to a less carbon-intensive region. Architectures relying on AI agents connected to multiple tools also require measuring cumulative calls: one user action can trigger several cloud requests.
Which indicators should an SMB track before entering into a contract?
An SMB should track seven indicators before reserving artificial intelligence resources: regional availability of GPUs or TPUs, service price, commitment duration, latency, share of carbon-free energy, grid carbon intensity, and the ability to interrupt processing. These data directly connect energy, budget, deadlines, and operational risk.
AI data center electricity consumption generally does not appear as a separate line item on a cloud bill. It is therefore necessary to manage the variables accessible in the provider’s console and in the application. Since the September 2026 publication, Google Cloud Carbon Footprint has notably used version 17 of its carbon model and revised regional coverage.
Here are the checks to perform before making a capacity commitment:
- verify that the required GPU or TPU is available in the target region and ask whether a reservation is necessary;
- compare pay-as-you-go pricing with the savings from a one- or three-year commitment, then test an underutilization scenario;
- measure latency from actual users rather than from the technical team’s offices;
- check the percentage of decarbonized energy hour by hour and the carbon intensity published for each region;
- classify workloads as immediate, deferrable, and interruptible so that those that can be shifted are moved;
- track the cost per business operation, such as per document analyzed, rather than only the cost per request.
Another risk deserves attention: regional commitment reduces flexibility in the event of a capacity shortage. From the agency’s perspective, the instinct is to test the service with representative volumes before making a commitment for several years. Honestly, a 55 % discount is justified only if the technical requirements and volume are already stable.
Security also needs to be budgeted. A cost reduction achieved by multiplying services and connections can increase the attack surface, that is, the number of exploitable points. Regular management of correctifs, described in this guide on the patch management without production downtime, is part of the true cost of cloud infrastructure.
Do energy contracts guarantee electricity dedicated to the cloud?
An energy contract signed by a cloud operator does not necessarily mean that a power plant physically supplies a specific data center. The 2026 Google-Constellation accord operates at the PJM grid level: additional and existing generation continues to supply the regional electricity market, where electricity is pooled.
This distinction prevents an overly literal reading of the announcements. The contract provides financial visibility for certain facilities and brings contracted consumption and generation closer together, but the electrons do not follow a reserved path to Google’s servers. Electricity consumption by an AI data center remains dependent on the grid’s overall operation at any given moment.
The joint 2026 announcement also mentions load management and demand response, that is, the temporary reduction of non-prioritary consumptions when the grid is under strain. This mechanism can improve resilience without interrupting critical services, provided that workloads have been classified and designed to supporter a report.
For a customer company, the useful information is therefore not limited to the announced number of megawatts. It includes the available region, the price charged, reservation terms, measured emissions, and service continuity. Defining these parameters before architecture and contracting avoids most unpleasant surprises; an outside perspective can alors save time without requiring a particular provider.
FAQ about electricity for AI data centers
What is the difference between an MW and a TWh?
A megawatt, or MW, measures the power available at a given moment. A terawatt-hour, or TWh, measures the amount of electricity consumed or produced over a period of time.
Will artificial intelligence increase my cloud bill?
Artificial intelligence can increase a cloud bill if the number of requests, their length, or the use of accelerators increases. No mechanism published in 2026 automatically links Google-Constellation accord to the inference price charged to customers.
Does a low-carbon cloud region really cost more?
A low-carbon cloud region does not necessarily cost more, because pricing also depends on the service, hardware, and local capacity. Google Cloud confirms in 2026 that prices and availability may vary by region and zone.
Can the electricity used by an AI application be reduced?
An AI application can reduce its computational workload by choosing a smaller model, limiting unnecessary requests, caching certain responses, and deferring non-urgent processing. The right metric remains the energy or cost per useful business operation.