Google just committed to buying nuclear power it won't receive until 2028 from reactors that don't yet have the capacity to produce it, and the markets treated this as unambiguously good news. Constellation Energy's stock jumped 12.25%, closing at $300.40 on October 6, after the two companies announced a 20-year partnership to deliver 890 megawatts of new nuclear capacity to the grid. This is not a typical renewable energy credit deal. Google is literally funding the physical upgrade of 11 existing nuclear reactors to power its AI infrastructure expansion.
What Actually Happened
On October 6, 2026, Google and Constellation Energy announced a 20-year power purchase agreement that will enable Constellation to invest more than $4.3 billion in modernizing 11 nuclear reactors across 6 sites in Illinois, Pennsylvania, and New Jersey. The upgrade process, known as an uprate, involves installing new digital control systems, advanced turbines, and updated steam generators to extract an additional 890 megawatts from existing plant infrastructure. The electricity will start arriving on the PJM Interconnection grid as early as 2028, making this the fastest path to new nuclear capacity outside of restarting a mothballed plant.
The deal does not stand alone. SiliconAngle reported that the two companies struck a parallel 15-year supply agreement for an additional 2,700 megawatts from Constellation's existing operational plants, bringing the total contracted capacity to roughly 3,590 megawatts. That is more than three times the output of the Hoover Dam. Constellation will also deploy Google Cloud and Gemini Enterprise to manage site operations under an expanded five-year technology alliance, creating an AI-for-energy feedback loop in which AI demand directly pays for the nuclear infrastructure that powers AI workloads.
The employment numbers attached to the deal carry political weight. Yahoo Finance noted the contract will sustain 4,400 existing high-skilled operational and union positions while creating approximately 7,200 new construction jobs during the upgrade phase. The PJM grid serves 67 million people across 13 states; adding 890 megawatts of firm, dispatchable, carbon-free power is not just a Google balance sheet entry but a grid reliability improvement at regional scale that benefits every other grid participant in the eastern interconnection.
Why This Matters More Than People Think
Nuclear power operates above a 92% capacity factor, compared to roughly 34% for wind and 24% for solar under real-world conditions. For AI data centers running at full load continuously, the math is unambiguous: nuclear is the only scalable clean energy source that matches the flat, always-on power curve of GPU clusters. Google's $4.3 billion bet is not an environmental gesture; it is a technical decision driven by the physics of inference workloads. Every watt of nuclear capacity contracted now is a watt that does not need to come from natural gas during grid stress events, reducing both cost volatility and carbon exposure for Google's infrastructure simultaneously.
The deal also validates the uprate strategy as a faster path to new nuclear capacity than greenfield construction. A new nuclear plant in the United States takes roughly 10 to 15 years to build and license; a reactor uprate can deliver additional capacity in 3 to 5 years using existing licensed infrastructure. The 2028 delivery date on Google's 890 megawatts is achievable precisely because Constellation is not building new plants. For AI companies racing to expand compute capacity before the next model generation cycle, the distinction is critical: uprates can be contracted today and deliver power on a timeline that aligns with hardware roadmaps.
The technology partnership embedded in the deal is equally underappreciated. Constellation selecting Google Cloud and Gemini Enterprise as its AI platform partner means Google's models will be used to optimize the dispatch, safety monitoring, and capacity planning of nuclear reactors. That creates a reference case for AI-managed nuclear operations that every utility in the PJM footprint can observe. If Constellation's AI-optimized plants outperform peers on uptime and efficiency metrics, the case for broader industry adoption of AI in nuclear operations becomes empirical rather than theoretical, and Google's tools become the default recommendation.
The Competitive Landscape
Google is not the only hyperscaler committing to nuclear this week. Amazon signed a 20-year nuclear power deal with Constellation the day before on October 5, 2026. Microsoft restarted Three Mile Island Unit 1 under a long-term deal with Constellation signed in 2024, with that plant now delivering power to the PJM grid. Across 13 nuclear deals tracked by industry monitors through mid-2026, the four major US hyperscalers have committed over 9.8 gigawatts of nuclear capacity. The race is no longer whether AI companies will buy nuclear power but which company controls the most firm clean capacity before the grid becomes fully contracted and prices reset permanently higher.
The historical parallel is the 2010s cloud real estate race, in which Amazon, Google, and Microsoft each locked up regional data center locations and fiber routes years before demand made those positions obviously valuable. Companies that moved first secured locations their competitors could not replicate; companies that waited paid premiums or accepted inferior geography. The nuclear capacity race follows the same logic: Constellation's 11 sites in the PJM footprint represent irreplaceable, permitted, grid-connected infrastructure. Locking up that capacity now is equivalent to claiming the best data center real estate on the continent a decade ago, before the value of that position became obvious to everyone.
The risk case, however, is real. Skeptics point out that nuclear uprate timelines routinely slip. The 2028 delivery estimate depends on permitting approvals, supply chain reliability for specialized turbine components, and labor availability for nuclear-grade construction work, all three of which have historically caused multi-year delays in comparable projects. If the electricity does not arrive until 2030 or later, Google will have paid a premium for capacity it could not use during the critical build-out phase of its AI infrastructure. The bear case is that Google is betting $4.3 billion on regulatory and construction execution in an industry that is famous for neither.
Hidden Insight: The Grid as a Competitive Moat
The real prize in Google's nuclear deal is not the 890 megawatts of new capacity. It is the 3,590 megawatts of total contracted power in the PJM region, the largest and most interconnected electricity grid in North America. PJM serves the US East Coast from Illinois to New Jersey, which is precisely where the largest concentration of legacy data center infrastructure, financial services firms, government agencies, and research universities is located. A company that controls the cheapest, most reliable power in this region controls the economics of AI inference for the entire Eastern seaboard, effectively setting a floor on what any competitor must pay to operate in the same geography.
The Gemini Enterprise technology partnership embedded in the deal is a Trojan horse that most coverage has glossed over. Constellation will use Google's AI tools to manage nuclear plant operations, which means Google's models will be trained on nuclear operational data: actual reactor dispatch curves, maintenance schedules, failure mode patterns, and grid optimization decisions. That is a proprietary dataset that no competitor can access or replicate. Models trained on real nuclear operational data will be better at energy management, grid optimization, and power trading than models that are not. The AI tools get better because of the nuclear data, which makes the nuclear assets more valuable, which funds more AI investment. Google has quietly structured a compounding advantage loop.
The timing of the deal is worth reading against the backdrop of Nvidia's data center roadmap. The Vera Rubin NVL72 system, currently the highest-density GPU cluster available, draws approximately 200 kilowatts per rack. Nvidia's next-generation Rubin Ultra NVL576 system is projected to reach 600 kilowatts per rack by 2027. The power density of AI compute is tripling within 18 months. Companies that have not locked in firm power contracts now will face a grid that simply cannot supply what they need at any price. Google's 3,590-megawatt position looks like overcautious procurement today; by 2028 it may prove to be just barely enough given the GPU roadmap acceleration.
Finally, the political dimension should not be underestimated. Nuclear power in the United States is experiencing a bipartisan revival driven by AI demand, energy security concerns, and the economic case for carbon-free baseload power. A Google-Constellation deal that preserves 4,400 union jobs and adds 7,200 construction jobs in Pennsylvania, Illinois, and New Jersey, three perennial swing states in federal elections, creates a political constituency for nuclear that the industry has not had in decades. Utilities in those states will now find it easier to secure regulatory approvals for future uprates because Google's committed capital makes the business case obvious to state regulators. The deal is reshaping the politics of nuclear licensing at the same time it is reshaping the economics of AI power procurement.
The insurance logic embedded in Google's position is worth unpacking separately from the energy economics. Across Google's total AI infrastructure spend, the $4.3 billion committed to nuclear capacity represents less than 5% of the company's annual capital expenditure budget, which has been running above $100 billion in recent quarters. For that relatively modest allocation, Google has secured firm price protection on 3,590 megawatts of power for up to 20 years. In a world where natural gas prices spike during grid stress events, or where carbon pricing mechanisms eventually penalize fossil-fuel-based power at the federal level, that hedged position has compounding value well beyond its face cost. The deal is simultaneously a practical power procurement agreement and a financial hedge against energy price volatility, structured over a long enough horizon that the nuclear uprate payback period aligns with Google's data center asset depreciation schedules. The CFO case for this deal is as compelling as the engineering case, which is why markets responded immediately and emphatically when the announcement landed on October 6.
What to Watch Next
The immediate 30-day signal is whether Microsoft responds with a counter-announcement. Microsoft's existing Three Mile Island deal covers the Crane Clean Energy Center in Pennsylvania at roughly 835 megawatts, but expires within a decade. If Microsoft announces a new or expanded nuclear deal within the next month, it will confirm that the hyperscaler nuclear race has entered a sprint phase. Watch for filings with the Nuclear Regulatory Commission for uprate approval requests at the six Constellation sites; those filings are public records and will reveal the realistic timeline for the 2028 delivery promise embedded in Google's press release.
At the 90-day mark, watch Constellation's construction contracting activity. Uprating 11 reactors requires specialized turbine manufacturers, nuclear-grade welding crews, and digital control system integrators. Constellation will need to lock in those contractors quickly to hit the 2028 target. If the company announces major equipment procurement agreements or specialized workforce expansion in Q4 2026, that signals the project timeline is credible. If procurement is slow through December, the 2028 delivery date faces real execution risk regardless of what the contract terms require.
The 180-day horizon is the one that matters most for the broader AI industry. Other hyperscalers, including Meta, Apple, and the leading Chinese AI labs, will need to make their own nuclear commitments or risk being structurally locked out of firm clean power in the PJM region. The total contracted nuclear capacity in PJM is approaching the region's available surplus. A company that has not signed a long-term deal by mid-2027 may find there is nothing left to sign at any price. The window to lock in nuclear power for the AI era is closing, and Google just claimed the largest single position currently available in the eastern grid. Watch also whether smaller advanced nuclear developers, including NuScale, Kairos Power, and TerraPower, announce supply agreements with hyperscalers in the coming months because Google's deal validated the commercial model. If SMR developers begin signing AI company contracts before mid-2027, the Google-Constellation deal will have catalyzed a second procurement wave that extends nuclear's AI energy role well beyond existing reactor uprates.
Google is not buying electricity. It is buying the right to exist as an AI company in the decade when power becomes the binding constraint on compute.
Key Takeaways
- 890 MW of new nuclear capacity: created by uprating 11 existing reactors at 6 Constellation sites in the PJM grid, with electricity arriving as early as 2028
- $4.3 billion committed by Constellation: to fund reactor modernizations that would not otherwise be economically viable without a long-term contracted buyer
- 3,590 MW total deal scope: the uprate PPA is paired with a 15-year, 2,700 MW supply agreement from Constellation's existing operational fleet
- 12.25% Constellation stock surge: CEG closed at $300.40 on October 6, reflecting market confidence that nuclear power has found durable AI-driven demand
- 7,200 new construction jobs: plus 4,400 sustained union positions, creating political support for nuclear licensing in three key US swing states
Questions Worth Asking
- If nuclear uprate timelines slip past 2030, what happens to Google's AI expansion plans given the 600 kW per rack power density of next-generation GPU clusters arriving in 2027?
- The Gemini Enterprise partnership gives Google proprietary nuclear operational data. How long before that advantage shows up in AI-optimized energy management products competing against industrial software incumbents?
- With Google, Amazon, and Microsoft controlling the majority of contracted PJM nuclear capacity, what does that mean for the energy costs of smaller AI labs and startups that cannot afford multi-decade power commitments?
