> ## Content Index
> Fetch the complete content index at: https://www.malektakieddine.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI’s electricity deals: what is actually being bought?
- URL: https://www.malektakieddine.com/ai-electricity-deals-what-is-actually-being-bought/
- Published: 2026-09-16T00:00:00.000Z
- Updated: 2026-09-17T18:59:09.000Z
- Description: AI companies are securing electricity through acquisitions, power-purchase agreements, utility programmes and computing-capacity contracts. The legal structure determines the degree of control — and the risks retained.
- Author: Malek Takieddine
- Tags: #writing, Artificial Intelligence, Energy, Infrastructure, Law & Regulation

In recent market developments, Alphabet has [acquired an energy developer](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm?ref=malektakieddine.com) and Microsoft has [contracted for electricity](https://www.constellationenergy.com/news/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html?ref=malektakieddine.com) from a nuclear restart. At face value they both fit broadly in the same story about artificial intelligence (AI) and energy, but there is merit in distinguishing between those arrangements. This becomes more relevant when forecasting AI industry practices in securing energy sources.

These are electricity arrangements associated with data-centre expansion in an increasingly AI-driven market, although not every arrangement is dedicated exclusively to AI workloads. For companies expanding AI infrastructure, the question is when greater control over electricity is worth the capital and risk it requires. The legal arrangements determine what is actually being bought: a degree of control over how electricity supply is secured, coupled with obligations if plans change.

That inquiry extends beyond direct electricity transactions. OpenAI and Anthropic show that an AI developer may instead buy computing capacity, while cloud providers, data-centre developers and utilities perform other parts of the infrastructure and electricity-supply chain. Public announcements do not always disclose which entity contracts for the power or how the resulting obligations are allocated.

## When ownership earns its cost

Alphabet’s acquisition of Intersect closed on 10 March 2026\. The deal [excluded some existing assets](https://abc.xyz/investor/news/news-details/2025/Alphabet-Announces-Agreement-to-Acquire-Intersect-to-Advance-U-S--Energy-Innovation-2025-DVIuVDM9wW/default.aspx?ref=malektakieddine.com), but brought a development business capable of coordinating energy investment with computing plans. Ownership makes most sense where a company expects a substantial pipeline of projects and repeated coordination problems make reliance on separate suppliers costly. It also commits capital and exposes the group to development risks that a purchase contract may leave with a specialist.

Industrial self-supply has a long history. However, control of a developer is not permission to generate or deliver electricity. Project permits, connection rights and network-service conditions still matter. An acquisition can bring decisions within the same group without bringing every decision within that group’s discretion.

Microsoft’s 20-year power purchase agreement supports Constellation’s planned restart of the Crane Clean Energy Center in Pennsylvania. Constellation remains the owner and nuclear licensee; the restart was still [under regulatory review](https://www.nrc.gov/info-finder/reactors/ccec?ref=malektakieddine.com) as at 14 September 2026\. This model can provide predictable revenues for the developer while preserving the buyer’s capital for computing infrastructure. Its attraction depends on delivery dates, pricing and remedies for delay being commercially workable. A right to purchase output must also work alongside the network arrangements needed to serve the data centre.

Constellation describes the agreement as helping Microsoft match the electricity used by its data centres in the PJM region with carbon-free energy. It does not identify the contracted output as exclusively serving AI workloads. The transaction remains useful here as a model of long-term energy procurement in the data-centre market, rather than evidence of a dedicated AI power supply.

A long contract can nevertheless restrict future choices through minimum payments, guarantees or exit costs, where these are agreed. A company can avoid owning a power station and still bind its future quite tightly.

A utility programme offers a third route: the customer helps justify investment while the utility develops infrastructure under regulatory oversight. It can combine specialist delivery with contributions to shared network costs, but depends on approved service terms and credible customer commitments. Where a customer’s plans shape a shared network, regulators and other users also have an interest in understanding the resulting commitments and remaining risks.

## When the AI company buys computing capacity

OpenAI illustrates a hybrid arrangement. In January 2026, OpenAI and SoftBank each announced a $500 million investment in SB Energy. OpenAI also signed a [1.2 GW data-centre lease](https://openai.com/index/stargate-sb-energy-partnership/?ref=malektakieddine.com) and selected SB Energy to build and operate its Milam County site, with new generation planned to support the site’s energy needs. OpenAI is therefore investing in the development partner and contracting for capacity, while a specialist develops and operates the underlying campus and associated energy infrastructure.

The arrangements vary by site. OpenAI says that every Stargate campus will have a locally tailored plan under which the project pays for the energy infrastructure it requires. Its [published examples](https://openai.com/index/stargate-community/?ref=malektakieddine.com) identify Oracle and Related Digital working with DTE Energy in Michigan, and Oracle and Vantage working with WEC Energy Group in Wisconsin. Those announcements do not establish that OpenAI itself signs every underlying electricity contract.

Anthropic combines purchased computing capacity with purpose-built facilities. It announced a [$50 billion investment in American computing infrastructure](https://www.anthropic.com/news/anthropic-invests-50-billion-in-american-ai-infrastructure?ref=malektakieddine.com), including custom data centres being built with Fluidstack in Texas and New York. Separately, its [agreement with Amazon](https://www.anthropic.com/news/anthropic-amazon-compute?ref=malektakieddine.com) commits more than $100 billion over ten years to AWS technologies and secures up to 5 GW of computing capacity. It has also contracted for [multiple gigawatts of Google and Broadcom capacity](https://www.anthropic.com/news/google-broadcom-partnership-compute?ref=malektakieddine.com). These are computing commitments, not disclosed power purchase agreements between Anthropic and generators.

Anthropic itself distinguishes between data centres developed with partners for its workloads and capacity leased from existing facilities. It applies its electricity-cost commitments directly to the former; for the latter, it says it is [exploring further ways](https://www.anthropic.com/news/covering-electricity-price-increases?ref=malektakieddine.com) to address the effect of its workloads on prices. In other words, buying computing capacity can move the electricity relationship one contract further away without necessarily removing exposure to delay, scarcity or cost. The allocation must be traced through the computing agreement as well as the electricity arrangements beneath it.

## Capital pressure is not capital exhaustion

The latest accounts do not support a general claim that these groups have exhausted their capital. In the quarter ended June 2026, Alphabet’s [capital expenditure exceeded operating cash flow](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm?ref=malektakieddine.com) by $5.9 billion, yet it reported $242.5 billion in cash, cash equivalents and marketable securities at quarter end. That total included [$80 billion of SpaceX shares subject to short-term selling restrictions](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm?ref=malektakieddine.com), so it should not be treated as entirely available for immediate deployment. Substantial debt and other commitments also limit what can be allocated to new projects. Microsoft [reported $19.6 billion](https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4?ref=malektakieddine.com) in free cash flow, cash from operations less cash capital expenditure, for the same quarter.

These figures describe current capacity, not the motives behind earlier deals. Nor can the position of two diversified groups stand for every AI business. The strategic constraint is what else the money could earn, and what delay might cost. If control over energy development materially accelerates a valuable computing project, ownership may justify its burden. Where a specialist can deliver equivalent certainty under contract, committing additional ownership capital becomes harder to defend.

## An old plant answers a different question

Ownership versus contracting and restarting versus building new are separate choices. A buyer can contract for a new plant’s output without owning it. It can also support a restart through a long-term commitment, as Microsoft has done.

Reusing a suitable site and equipment may bring nuclear capacity back sooner than building a new nuclear station. That advantage depends on the plant’s condition, refurbishment, licensing and network access; Crane requires [substantial equipment work](https://www.constellationenergy.com/news/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html?ref=malektakieddine.com) as well as regulatory approval. A retired plant is not spare capacity waiting for someone to turn a key. Nor does a potential advantage over new nuclear construction establish that a restart beats every new generation option. Its usefulness also depends on finding a suitable retired plant where, and when, the customer needs supply.

Policy also changes the calculation. The US Department of Energy [closed a $1 billion loan](https://www.energy.gov/edf/crane-restart?ref=malektakieddine.com) to Constellation for Crane in November 2025\. [Constellation says the loan lowers financing costs](https://www.constellationenergy.com/news/2025/11/us-government-backs-constellations-plan-to-launch-crane-clean-energy-center-adding-835-mws-of-new-baseload-power-to-the-grid.html?ref=malektakieddine.com). The loan is backed by Constellation’s credit and balance sheet, but public financing remains part of the arrangement. Restart economics cannot be assessed solely as a private choice to buy cheaper electricity.

## Which practice is likely to spread?

We could expect a mixture of developer ownership, direct power purchases, utility programmes and computing-capacity arrangements, with ownership concentrated where control solves a specific delivery problem. Google itself signed a [25-year agreement with NextEra](https://www.investor.nexteraenergy.com/news-and-events/news-releases/2025/10-27-2025-203948689?ref=malektakieddine.com) for electricity from the proposed Duane Arnold nuclear restart in Iowa. That restart also involves public financing: on 8 September 2026, the US Department of Energy [closed a loan of up to $1.9 billion](https://www.energy.gov/articles/energy-department-closes-19-billion-loan-restart-duane-arnold-nuclear-plant?ref=malektakieddine.com) to NextEra to help finance it. The Alphabet group already illustrates why acquisition and contracting can complement each other.

For a company without a recurring development pipeline or the capacity to manage it, an electricity purchase, utility arrangement or contract for computing capacity is the more defensible starting point. That view must remain conditional: these selected cases establish neither the prevalence of each model across the AI industry nor that technology companies have collectively become public utilities.

We should also keep in mind that the regulatory framework will influence that balance. Clear connection rules and predictable service charges could make contracting easier; prolonged uncertainty may encourage companies to bring development closer to their computing plans. Ownership cannot remove a network constraint merely by changing who owns the developer.

In December 2025, the US Federal Energy Regulatory Commission (FERC) directed grid operator PJM to [clarify service rules](https://www.ferc.gov/news-events/news/fact-sheet-ferc-directs-nations-largest-grid-operator-create-new-rules-embrace?ref=malektakieddine.com) for customers located alongside generators. FERC [revised parts of its approach](https://www.ferc.gov/media/e-2-el25-49-002?ref=malektakieddine.com) in June 2026 and required further filings. PJM [submitted a further compliance filing](https://elibrary.ferc.gov/eLibrary/docketsheet?docket%5Fnumber=ER26-1479&ref=malektakieddine.com) on 17 August 2026; FERC's consideration remained pending as at 14 September 2026\. Rules permitting customers to accept interruption could widen their connection options, while charges for reliable network support affect the economics of locating beside a plant. That option has value only if the computing operation can tolerate the interruption.

Customer protection can also support investment by making its cost allocation more defensible. Entergy says [Meta must cover its full cost of service](https://www.entergy.com/news/entergy-louisiana-announces-a-new-agreement-with-meta-that-will-deliver-an-additional-2b-in-customer-savings?ref=malektakieddine.com) and projects savings for other customers. Its [published account](https://www.entergy.com/news/5b-in-customer-savings-delivered-by-data-center-agreements-issues-fair-share-plus-pledge?ref=malektakieddine.com) of data-centre agreements identifies prepayments, financial security and early termination provisions; its stated principles also call for parent guarantees.

OpenAI says Stargate projects will fund the incremental generation and grid upgrades their loads require. Anthropic says it will pay all grid-upgrade costs needed to connect data centres developed for its workloads, procure new generation and address demand-driven price effects where that generation is not yet available. These commitments deserve weight, but published statements cannot by themselves show how enforceable obligations run among the AI company, data-centre developer, cloud provider and utility.

If demand falls short, the practical question is whether enforceable customer commitments cover the costs already incurred and assets left behind. The scope and duration of a guarantee matter as much as its existence. Effective protection strengthens the case for investment; public announcements alone cannot establish its adequacy. Where rules require the customer to pay even if its plans change, avoiding ownership may offer less financial flexibility than the label suggests.

The ability to shape an electricity system around one company’s growth deserves public scrutiny, even when the company pays its bills. If private plans help determine what gets built, the people who may live with the consequences should be able to see where private responsibility ends.