On October 6, 2026, Google and Constellation Energy completed a 3,590-megawatt long-term power supply agreement covering the service area of PJM, the largest U.S. grid, with the contract triggering more than $4.3 billion in new investment.
Agreement Breakdown: A Dual-Track Structure of Nuclear and Clean Energy
According to CNBC, the total 3,590 MW is divided into two parts: 890 MW comes from a 20-year nuclear power purchase agreement signed with Constellation Energy, involving uprates and capacity expansions at 11 operating nuclear reactors in Illinois, Pennsylvania, and New Jersey; the other 2,700 MW comes from a long-term supply agreement with existing Constellation Energy plants operating within PJM’s footprint, not tied to a specific generation source.
The first electricity from the nuclear portion is expected to be connected to the grid in 2028.
Amanda Peterson Corio, Google’s global head of energy and power, said in a statement: “We are committed to meeting growing demand responsibly by actively investing in clean, reliable power and bringing new capacity to the nation’s grid.”
The Policy Trigger for the “Bring Your Own Power” Era
To understand the context of this contract, one must first understand the structural changes taking place at PJM. PJM is the largest regional grid operator covering 13 U.S. states, and in recent years it has faced explosive growth in data center interconnection demand. PJM management has proposed a mandatory policy: data center customers connecting to the grid must either provide their own power source or accept remote power cuts during peak grid demand periods.
This “Bring Your Own Power” policy fundamentally changes the commercial logic of AI infrastructure. In the past, data centers only needed to pay electricity bills and connect to the grid. In the future, under the PJM framework, large data centers without their own power supply will bear the risk of supply interruption.
Both Google and Constellation Energy explicitly confirmed that this contract is a direct response to PJM’s “Bring Your Own Power” proposal.
Why Nuclear: Uprating Existing Plants Is Faster Than Building New Ones
The technical path for the 890 MW nuclear portion: Google did not choose to build new nuclear plants, but instead to increase generation by uprating 11 existing reactors. This choice was not accidental.
In the United States, the approval and construction cycle for building a new nuclear plant from scratch usually exceeds 10 years; for power uprates at existing units, only regulatory review and engineering modifications are needed, making the cycle significantly shorter. The timetable for first electricity to be connected to the grid in 2028 depends precisely on this shortcut. Constellation Energy is the largest nuclear operator in the United States, with a complete existing fleet.
This is not Google’s first time adopting a nuclear strategy. In October 2025, Google signed a 25-year agreement with NextEra Energy to support the restart of the shuttered Duane Arnold nuclear plant in Iowa (615 MW), expected to resume power supply in 2029. Earlier, Microsoft locked in the restart of Three Mile Island (Crane Clean Energy Center) in Pennsylvania with a 20-year agreement, with capacity of about 835 MW.
Side-by-side comparison: Microsoft’s Three Mile Island agreement is a single unit, 835 MW; Google’s nuclear portion this time is 890 MW from aggregated uprates across 11 units. The former is a big bet on one landmark, while the latter is a diversified bet on a fleet—completely different risk structures.
Sense of Scale: How Much This Contract Accounts For
According to industry data, electricity demand from AI data centers within the PJM grid is expected to add about 30 gigawatts (GW) by 2030. Google’s 3,590 MW contract accounts for about 12% of that increment.
Globally, the current picture is clearer: in 2026, U.S. data centers account for about 45% of global data center electricity consumption, while China accounts for about 25%. Total U.S. data center load has reached about 41 GW, while China’s is about 8.5 GW. China’s AI compute electricity scale is currently less than one-fifth of that of the United States, but a Wood Mackenzie forecast shows that by 2030 China’s data center electricity consumption will jump from about 170 terawatt-hours in 2025 to 774–800 terawatt-hours, at which point it will account for about 6% of China’s total electricity consumption.
Tech Giants Are Becoming Power Companies
The deeper logic revealed by this deal goes beyond simple energy procurement. What Google, Microsoft, and Amazon are doing is an infrastructure positioning battle using decades-long contracts as weapons. A 20-year nuclear power purchase agreement means Google effectively assumes long-term asset risk similar to that of a power company—it is no longer just a power user, but a substantive financier of nuclear uprates. The $4.3 billion in new investment is driven by Google’s demand commitment.
This model is rewriting the boundaries between the energy and technology industries. Traditionally, nuclear investment was borne by utility companies, with risk shared by regulators and end users; now, part of the risk has been shifted onto the balance sheets of large tech companies in exchange for long-term certainty of power supply.
For grid operators, this is a relief: PJM does not have to bear alone the infrastructure pressure of serving AI data center load growth. For Constellation Energy, this is stable long-term cash flow. For Google, this is insurance for operational continuity, at the cost of deeply binding itself to nuclear infrastructure in the eastern United States.
2028, a Tight Time Node
The phrase in the agreement that “the first electricity will be connected to the grid in 2028” needs to be treated cautiously. Nuclear power uprates involve the approval process of the U.S. Nuclear Regulatory Commission (NRC), which has historically been delayed many times. Which of the 11 units completes its upgrade first, and whether actual grid connection can be delivered on schedule, is the biggest execution variable in this agreement.
At the same time, Google’s AI compute expansion cannot wait until 2028 to begin—meaning that before nuclear power comes online, Google will rely on the other 2,700 MW of existing clean energy in the agreement to fill demand. This supply is not tied to a specific generation source, equivalent to transitional protection while waiting for nuclear delivery.
Conclusion
Google’s 3,590 MW contract is essentially a judgment: the bottleneck in AI compute competition has shifted from chips to electricity. Chips can wait for tape-out cycles, but data centers cannot wait for blackouts. When PJM announced “bring your own power or be cut off,” Google’s answer was to fortify its power supply moat with $4.3 billion in nuclear investment.
This is not a tech company’s posture performance on energy transition, but a direct confrontation with a real engineering constraint: without solving the electricity problem, the ceiling on AI expansion is not algorithms, not data, but watts.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接