Data center electricity demand increased by 17% in a single year, according to the International Energy Agency. In absolute terms, this represents approximately 70 TWh in additional supply added to the global grid in twelve months — equivalent to the annual consumption of a country like Belgium or Sweden, and not Spain whose total consumption is nearly three times higher. The grid cannot keep up. So the cloud giants found another solution: buy the electricity before it reaches the grid. Or buy outright the reactor that produces it.

This movement is not marginal. The global pipeline of agreements between data centers and small modular reactors (SMRs) has grown from 25 gigawatts at the end of 2024 to 45 gigawatts in early 2026. Microsoft, Amazon, Google and their competitors are no longer queuing to buy electricity on the market. They are signing direct supply contracts, reserving slices of reactors under construction, or financing the manufacturers themselves to ensure priority. This is a bifurcation in the history of American electrical infrastructure, and its effects are already being measured on household bills.

The Essentials

  • Global data center consumption was ~415 TWh in 2024 according to the IEA. The 17% increase recorded in 2025 represents approximately 70–75 TWh in additional supply in one year.
  • The SMR pipeline contracted with hyperscalers grew from 25 to 45 GW between late 2024 and early 2026, nearly doubling in eighteen months.
  • Microsoft concluded an agreement with Constellation Energy to reopen Three Mile Island Unit 1; Amazon invested in X-energy for SMR reactors; Google placed an order with Kairos Power for 500 MW of reactors.
  • Studies from the Brookings Institution document residential rate increases in counties where data centers install without connection costs being shared.
  • The regulatory question remains open: who pays for the transmission infrastructure needed to supply these massive installations?

45 Gigawatts: What This Figure Conceals

Forty-five gigawatts is approximately 70% of France’s installed nuclear capacity, which has 57 reactors for a total of some 62 to 63 GW. Seeing a figure of this magnitude appear in a pipeline of private contracts between technology companies and reactor developers shows the scale of change underway. Two years ago, SMRs were still presented as a promising technology for the following decade. In eighteen months, they have become the object of an unprecedented rush of private investment in the nuclear sector since the 1970s.

It is not technology alone that has changed. It is demand. The IEA documents data center consumption growth that has accelerated well beyond 2023 projections: generative artificial intelligence models have multiplied consumption per computation by three to five times compared to traditional cloud workloads. Training a large-scale language model can consume as much electricity as a mid-sized city over several weeks. The capabilities AI is deploying today require physical infrastructure whose growth network operators had not anticipated.

The American grid, managed as a mosaic by dozens of regional operators and state regulators, is not designed to absorb such localized and sudden demand spikes. Interconnection lead times to high-voltage transmission networks exceed seven years in several states. Hyperscalers do not have seven years to wait. They have service commitments, investors and models to train.

Microsoft, Amazon, Google: Three Strategies for the Same Workaround

Microsoft struck first, with the most symbolic move. On September 20, 2024, the company signed an agreement with Constellation Energy to reopen Unit 1 of Three Mile Island, the Pennsylvania reactor whose name is inseparable from the 1979 accident. The plant, renamed Crane Clean Energy Center, is to supply 835 megawatts of electricity almost exclusively to Microsoft for twenty years. This is dedicated capacity purchasing, effectively removing this tranche from the wholesale market.

Amazon follows a different logic, more focused upstream. The group invested in X-energy, developer of Xe-100 high-temperature reactors, to guarantee itself priority access to this technology when it becomes operational. Google, for its part, placed an order with Kairos Power for 500 MW of molten salt reactors, with deployment scheduled between 2030 and 2035.

These strategies converge toward the same result: an increasing fraction of nuclear capacity under development in America is contracted directly between private parties with three or four buyers before even being built. The electricity thus produced will never pass through the public grid under normal market conditions. It feeds directly into computing campuses, sometimes with a dedicated physical connection that bypasses shared transmission lines.

An Infrastructure Fracture Taking Root in Household Bills

Until now, one might read this story as good news: private companies are financing decarbonized nuclear capacity without mobilizing public funds, which accelerates the transition and keeps the grid in working order. The argument is real. It does not cover everything.

The Brookings Institution has documented the local effects of the massive arrival of data centers in several states. In counties in Northern Virginia, Georgia and Ohio where hyperscalers concentrate, residential rates have risen faster than the national average. The mechanism is precise: data centers require extensions to local distribution network capacity, whose costs are shared among all subscribers according to state regulator pricing rules. Companies pay for their connection, but not for the full network reinforcement their presence makes necessary. The difference lands on neighbors’ bills.

This phenomenon is still limited in absolute value. The documented increases remain significant in the affected zones, but they illustrate an underlying dynamic: America’s public electrical infrastructure does not generate enough revenue to modernize at the pace the data center demand imposes, and current cost-sharing rules are not calibrated for this situation. Regulators in several states have opened tariff review procedures, but administrative timeframes are measured in years.

There is an even deeper asymmetry. Hyperscalers investing in dedicated SMRs are building partial electrical autonomy over thirty to forty years. They will no longer depend on public grid quality for their critical operations. Small manufacturing firms, hospitals, carriers, households: all will continue to depend on a grid whose modernization will be partly starved of the most solvent and most predictable demand base.

What SMR Manufacturers Can Do, and What They Cannot

The hyperscaler rush has an undeniably positive effect on the nuclear industry: it finances scaling up. SMR developers like NuScale, X-energy, Kairos Power or TerraPower have operated for years with ambitious roadmaps and fragile order books. Contracts from cloud giants have changed the financial situation.

NuScale, the first company to obtain Nuclear Regulatory Commission design certification for an SMR, went through a difficult period after the cancellation of its Utah UAMPS project in 2023 due to cost overruns. The dynamic of hyperscaler contracts has made it possible to restart negotiations with industrial buyers capable of absorbing higher prices than regulated utilities. Kairos Power, whose FHR molten salt reactor is still in demonstration phase, was able to accelerate its timeline thanks to the visibility provided by the Google agreement.

The risk is building a private nuclear industry optimized for a single type of customer. SMRs financed by hyperscalers are designed to meet continuous baseload needs, in medium-size configurations adapted to computing campuses. They are not necessarily optimized for residential distribution networks, which need different flexibility. The IEA signals a growing trend toward direct connections and “behind the meter” arrangements, meaning that a portion of capacity thus built will not automatically strengthen public electrical system resilience — even if agreements like Microsoft’s with Constellation do pass through transmission networks.

The Infrastructure Divide: A Question of Decades, Not Years

This is where the long arc of this subject becomes uncomfortable. The reactors that Microsoft and Amazon finance today will have an operational lifespan of forty to sixty years. The exclusive or priority supply contracts that accompany them run for twenty to thirty years. The assets thus built are generational assets: they structure a private electrical capacity that the next generation of engineers, entrepreneurs and industrialists will not be able to easily redirect toward the public grid.

At the same time, the American electrical grid needs, according to the IEA and the Department of Energy, several trillion dollars in investments by 2050 to ensure the transition and meet the electrification of transport, heating and industry. These investments depend on grid operators’ ability to generate sufficient revenue. If the most solvent demand, the most predictable and the most geographically concentrated is progressively absorbed by private self-generation systems, the revenue base of public grid operators is affected.

This is not immediate collapse. It is slow and cumulative pressure. The analogy with telecommunications networks of the 1990s is suggestive: large companies migrated to dedicated fiber private networks while the public telephone network aged from lack of sufficient investment. The digital divide that widened in the 2000s owes something to this infrastructure bifurcation. One can wonder whether the electrical divide of the 2030s is not being constructed today on the same pattern. This risk is at the heart of what Acemoglu and Johnson theorized in their analysis of technological capture dynamics: technology does not benefit everyone, except when forced to do so.

Regulators Seeking Their Answer

The regulatory response is forming, but it starts from far back. The Federal Energy Regulatory Commission (FERC) opened in 2024 a proceeding on dedicated interconnection agreements for data centers, which allow these installations to bypass normal queues for connection. Several states have begun revising their rate rules to better assign network extension costs to direct beneficiaries.

The U.S. Congress is examining several proposals to impose a data center contribution to network infrastructure costs, modeled on what some state regulators already require of large industries. The principle argument is simple: an installation consuming 500 megawatts and connecting directly to a dedicated reactor nonetheless avoids using shared transmission lines to import or export electricity, and its existence alters grid flows in ways that generate costs for other users.

The difficulty is as much regulatory as political. The nuclear sector needs massive private investment and hyperscalers are providing it. Imposing too heavy constraints on these investors risks slowing them down or pushing them toward less decarbonized solutions, like combined-cycle natural gas, which several major players have already begun mobilizing in parallel. Microsoft announced in early 2025 agreements to build or finance several gas plants for its peak power needs, illustrating the tension between availability, cost and climate ambition.

The question regulators must settle is not whether hyperscalers can build private reactors. It is defining the conditions under which this private capacity contributes to common infrastructure rather than circumvents it. Mechanisms exist: obligations to sell on the wholesale market during periods of low data center demand, financial contributions to grid modernization funds, capacity-sharing obligations in crisis situations. None are yet generalized.

The Bifurcation Still Avoidable

The movement of hyperscalers toward private nuclear is real, it is massive and it is largely irreversible in the short term. Signed contracts run for twenty years. Reactors under construction or in advanced development will be commissioned. Forty-five gigawatts of pipeline does not evaporate.

But the infrastructure bifurcation it heralds is not yet consummated. American regulators still have the capacity to establish cost-sharing and public grid contribution rules before the first installations are fully operational. The European Union, which is watching this dynamic with attention from Brussels, has a more centralized regulatory framework that would allow it to impose different conditions on actors seeking to replicate this model on the continent.

The most open question is not technical. It is political: will the nuclear capacity that tech private money will build over the next ten years be integrated into a shared electrical system, or will it constitute the beginning of a parallel network reserved for actors able to finance their own infrastructure? The regulatory responses of coming years will determine whether Silicon Valley’s nuclear becomes a national asset or an industrial enclave.

Sources

  1. International Energy Agency — Data Centre Electricity Use Surged in 2025: https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
  2. IEA — Energy and AI report (April 2025): global data center consumption ~415 TWh in 2024: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
  3. Brookings Institution — analyses on the local rate impact of data centers, 2024-2025: https://www.brookings.edu/articles/confronting-and-addressing-rising-energy-bills-linked-to-data-centers/
  4. Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report, 2024 edition
  5. Nuclear Regulatory Commission — NuScale Power Module design certification
  6. Federal Energy Regulatory Commission — proceeding on dedicated interconnection agreements, file opened in 2024
  7. Constellation Energy SEC Form 8-K: Microsoft/TMI agreement signed September 20, 2024: https://www.sec.gov/Archives/edgar/data/0001868275/000186827524000058/ceg-202409208kexh991.htm
  8. Google Official Blog: Google-Kairos Power agreement (500 MW SMR, October 2024): https://blog.google/outreach-initiatives/sustainability/google-kairos-power-nuclear-energy-agreement/
  9. Red Eléctrica (REE): Spain electricity consumption 2024 ≈ 249–256 TWh: https://www.ree.es/en/press-office/news/press-release/2025/03/electricity-generation-from-renewable-energies-in-spain-grows-by-10-3-in-2024-reaching-record-levels
  10. EDF: France nuclear capacity = 62.9 GW (57 reactors): https://www.edf.fr/en/the-edf-group/producing-a-climate-friendly-energy/nuclear-energy/the-nuclear-fleet-in-france
  11. CNBC / TechCrunch: Microsoft-Chevron agreement for Texas gas plant (2025-2026): https://techcrunch.com/2026/06/22/microsoft-and-chevron-plan-one-of-the-largest-gas-powered-data-center-projects-in-us/