Google signs 20 years of nuclear with Constellation: 890 MW for AI on the PJM grid

Google signs 20 years of nuclear power with Constellation: 890 MW for AI on the PJM grid

🔎 AI is no longer just a chip problem, it's a megawatt problem

Google has just signed one of the most massive energy supply contracts in the history of the digital era. Twenty years. 890 MW of new nuclear capacity injected into the PJM grid. More than $4.3 billion in direct investment to modernize Constellation Energy's fleet. The announcement, made official on October 6, 2026, marks a definitive turning point: hyperscalers are no longer just looking to buy green electricity — they are now financing the construction of critical power infrastructure to guarantee the sovereignty of their AI models.

Key takeaways

• 20-year framework agreement between Google and Constellation Energy for 890 MW of additional nuclear capacity on the PJM grid (Pennsylvania-New Jersey-Maryland Interconnection), equivalent to a large-scale reactor.
• Investment of more than $4.3 billion for "uprates" (power output increases) across 11 units in Illinois, Pennsylvania, and New Jersey; first operational uprate targeted for 2028.
• 5-year "AI for Energy" technology alliance: deployment of Gemini Enterprise on Google Cloud to optimize plant operations, OT cybersecurity, and reduce time-to-power.
• Strong market signal: Constellation stock (CEG) jumps +13.86% to $304.72 on the day of the announcement; a direct response to PJM's "Bring Your Own Power" initiative.

Recommended Tools

The Google–Constellation deal: anatomy of a landmark agreement

890 MW of new capacity without building a reactor

The technical and financial feat lies in the term "uprate". Constellation is not building new reactors (EPR or SMR), which would take 10 to 15 years. The operator is upgrading 11 existing units — turbines, steam generators, digital control systems — to extract additional output certified by the NRC. That's 890 MW of firm, dispatchable, low-carbon capacity, added to the PJM mix starting in 2028 for the first tranche.

The Constellation Energy press release specifies that this is "entirely new capacity" (new nuclear capacity) in the regulatory sense, distinct from existing output[^1]. MarketChameleon confirms the envelope: >$4.3B over the period, with a phased rollout beginning in 2028 across six sites[^2].

A 20-year PPA backed by the federal UPRISE program

The Power Purchase Agreement (PPA) falls under the UPRISE program (U.S. Program for Reactor Innovation and Sustained Energy), led by the Department of Energy (DOE). This federal framework aims to accelerate the qualification of uprates and to share the regulatory risk. For Google, it's an insurance policy: the capacity will be delivered and certified, and the electricity price is locked in for two decades — a rarity in a PJM market where capacity prices have exploded (+800% in the 2025/2026 auction).

Utility Dive notes that Constellation signed two 20-year nuclear PPAs in a single week: Google (890 MW) and Amazon (190 MW at Calvert Cliffs), totaling ~1.1 GW of contracted nuclear expansion[^3]. The race for firm capacity is on.

The "AI for Energy" Alliance: Gemini Enterprise Enters the Control Room

Beyond Buying Electrons: AI-Driven Optimization

Google isn't content with merely being a customer. The agreement includes a 5-year technology alliance dubbed "AI for Energy". The goal: deploy Gemini Enterprise on Google Cloud for three critical use cases identified by Constellation:

1. Operations optimization: predictive models on core/primary loop thermohydraulics to push uprates as close as possible to NRC safety margins.
2. OT (Operational Technology) cybersecurity: anomaly detection on industrial control systems (ICS/SCADA) via time-series log analysis — a major challenge since the 2023 TSA directive on nuclear resilience.
3. Reducing time-to-power: digital twins to simulate post-shutdown restart sequences, shortening maintenance outages from weeks to days.

ESG News summarizes it as: "deployment of Google Cloud + Gemini Enterprise for optimization, OT cybersecurity, time-to-power reduction"[^4]. This is the first time that a frontier-class LLM (Gemini 3.1 Pro / Deep Think) has been formally integrated into the American nuclear operations loop at this scale.

Why Gemini Enterprise and Not an Open-Source Model?

The answer comes down to two words: certification and SLA. Nuclear operators require full traceability of training data, reproducibility of inferences, and contractual support at a "safety-related" level. Gemini Enterprise provides:
- Data isolation (VPC Service Controls, CMEK).
- Compliance with NIST 800-53 Rev. 5, FedRAMP High.
- A guarantee of no retraining on customer data.

To compare the reasoning capabilities of current models on this type of complex task (agentic, code, log analysis), see our benchmark: Google Gemini vs ChatGPT vs Claude: which one for which use case?. The "Deep Think" version of Gemini 3 Pro ranks at the top of the Agentic leaderboard (score 95.4) just behind GPT-5.5[^5].

PJM, a grid under strain: why this deal is a game changer

67 million people, a 250 GW queue

PJM Interconnection operates the largest wholesale electricity market in the world: 67 million consumers, 180 GW of peak demand, but an interconnection queue exceeding 250 GW (mostly solar/batteries). The existing nuclear fleet (≈ 30 GW) is aging. Firm capacity (available 24/7) is becoming the scarcest asset.

The "Bring Your Own Power" (BYOP) initiative launched by PJM in 2025 allows large consumers to bring their own capacity in order to bypass the queue. Google answered the call. EnergyChoiceMatters notes that the October 6, 2026 announcement "adds new capacity on top of PJM's total generation"[^6] — a crucial legal nuance: this is not merely a reshuffling of existing electrons, it's a net creation of capacity.

Leverage on capacity prices

The PJM 2025/2026 capacity auction cleared at $269.92/MW-day (RTO zone), versus $34 the previous year. By securing 890 MW through nuclear uprates — a technology eligible for the most lucrative "Capacity Performance" product — Google hedges against this volatility. Conversely, Constellation locks in a predictable revenue stream over 20 years, making it easier to finance the $4.3 billion in CapEx (fixed-rate senior debt, investment-grade rating).

The hyperscalers' nuclear race: Google, Amazon, Microsoft, Meta

Amazon: 190 MW at Calvert Cliffs, but also SMRs

A week before Google, Amazon announced a 20-year PPA for 190 MW at the Calvert Cliffs plant (Maryland), once again with Constellation[^3]. In parallel, AWS is investing in X-energy (Xe-100 SMR) and TerraPower (Natrium) for post-2030 deliveries. A two-pronged strategy: uprates in the short term, SMRs in the medium term.

Microsoft: Three Mile Island and the rebirth of the TMI-1 site

In September 2024, Microsoft signed an agreement with Constellation to restart TMI-1 (835 MW) at Three Mile Island, shut down since 2019. Estimated cost: $1.6 billion, with the restart targeted for 2028. Same logic: uprate/restart of existing assets, long-term PPA, immediate firm capacity.

Meta: advanced geothermal and "advanced" nuclear

Meta is betting on next-generation geothermal (Fervo Energy, 400 MW contract in 2026) and exploring SMR partnerships (Kairos Power, Oklo). Less immediate light water nuclear, more technology bets.

Comparative summary

The trend is clear: hyperscalers are becoming de facto power infrastructure developers. They bring the capital (CapEx), long-term visibility (20-year PPAs), and now operational AI (Gemini, Bedrock, Azure AI).

Geopolitical, regulatory, and implications for the French nuclear industry

UPRISE and the American nuclear renaissance

The UPRISE program (DOE) funds feasibility studies, accelerated safety reviews, and risk-sharing on uprates. The Google–Constellation agreement is the first UPRISE "major project" to come to fruition. This validates the model: the State de-risks the regulatory side, while the private sector brings the money and the demand. Expect a cascade of similar announcements (Duke, Southern, TVA) by the end of 2027.

Cybersecurity: the TSA directive and AI's contribution

Since 2023, the TSA has required nuclear operators to implement OT cybersecurity plans validated by the NRC. The integration of Gemini Enterprise for OT anomaly detection directly addresses this requirement. This sets a precedent: generative AI is entering the nuclear "safety/security" perimeter. France (ASN, IRSN) is closely following this matter for its own future post-Fukushima safety reviews / 4th ten-year outages.

And where does France fit in all this?

EDF is carrying out its Grand Carénage program (€50bn, 2014–2025) and is preparing 7% uprates across its 900 MW/1300 MW fleet (targeted for 2030+). But the economic model differs: no 20-year PPA with a single hyperscaler (ARENH, the French capacity market). The Google–Constellation deal proves that a single, solvent, technically sophisticated buyer can unlock billions in nuclear CapEx without direct public funding. A lesson for the next multi-year energy programming plan (PPE 2028).

Common Mistakes

Mistake 1: Confusing "uprate" with "new reactor"

What's wrong: Headlining "Google funds a new nuclear reactor".
The reality: These are power uprates on existing reactors (turbines, steam generators, I&C). No new concrete for the reactor building. Timeline: 3–5 years vs 10–15 years. Cost: ~$4.8M/MW vs ~$10–15M/MW for an EPR/SMR.

Mistake 2: Believing Google physically "consumes" those 890 MW

What's wrong: Saying "Google's data centers will be powered by this plant".
The reality: The electricity is injected into the PJM grid (wholesale market). Google buys certificates (RECs) and the associated capacity. Electrons follow Kirchhoff's laws, not the contract. Additionality is accounting and contractual, not physical.

Mistake 3: Underestimating technical execution risk

What's wrong: Treating the 890 MW as a done deal by 2028.
The reality: Each uprate requires an NRC license amendment (LAR — License Amendment Request), a safety review, and power-ascension testing. A delay at a single site (e.g., steam generator corrosion, forging lead times) pushes back the phasing. The contract provides for penalties, but there's no magic.

Frequently Asked Questions

Why doesn't Google build its own SMRs?

Because the timeline is incompatible with the AI inference demand curve (2025–2030). SMRs (Xe-100, Natrium, BWRX-300) target first concrete in 2026–2028 and commercial operation in 2030–2033. Uprates deliver firm capacity as early as 2028 on a proven supply chain and safety basis. Google is doing both: short-term uprates + long-term SMR investments (Kairos, etc.).

What is the MWh price in this PPA?

Not disclosed. Long-term nuclear PPAs in the US typically trade at $60–90/MWh (full cost + margin level), well below PJM spot prices in summer 2026 (peaks > $500/MWh). The value for Google lies in capacity certainty (avoided cost of capacity) more than in the energy price alone.

Is Gemini Enterprise "nuclear certified"?

Not in the "safety-class" (Class 1E) sense. It operates in decision support (optimization, predictive maintenance, OT cybersecurity), not in direct control of reactor protection systems. The NRC requires software qualification (IEEE 7-4.3.2, RG 1.152) for any software in the safety loop — which is not the case here. The contribution is real, but the scope is clear.

Is this deal replicable in Europe / France?

Structurally yes, contractually no. In France, ARENH (€42/MWh, capped at 100 TWh/year) and the future nuclear "contract for difference" (CfD) mechanism change the game. A hyperscaler cannot sign a direct 20-year PPA with EDF for uprates without going through the French regulatory framework. That said, the "private CapEx in exchange for long-term visibility" model is inspiring the European electricity market reform (Regulation 2024/1747).

Conclusion

Google has just laid down the clearest marker to date: generative AI is financing the American nuclear renaissance. 890 MW, 20 years, $4.3 billion, and an AI (Gemini Enterprise) to run the plants safer, longer, faster. Hyperscalers are no longer mere customers of the grid — they are becoming its financial and technological architects. Next step: seeing whether the UPRISE + uprates + operational AI model will spread to Duke, Southern, TVA… and whether France will manage to replicate this CapEx/AI/Regulation alchemy before 2030.

[^1]: Constellation Energy, Google and Constellation Announce Landmark Agreement to Bring 890 MW of New Nuclear Capacity to PJM Grid, October 6, 2026. https://www.constellationenergy.com/news/2026/10/google-and-constellation-announce-landmark-agreement-to-bring-890-mw-of-new-nuclear-capacity-to-pjm-grid.html
[^2]: MarketChameleon, CEG Google 20-Year 890 MW Nuclear Uprate PJM Power Deal, October 6, 2026. https://marketchameleon.com/articles/b/2026/10/6/ceg-google-20-year-890-mw-nuclear-uprate-pjm-power-deal
[^3]: Utility Dive, Constellation Google deal will bring 890 MW of new nuclear to PJM, October 7, 2026. https://www.utilitydive.com/news/constellation-google-deal-will-bring-890-mw-of-new-nuclear-to-pjm/832223/
[^4]: ESG News, Google Constellation sign 20-year deal to add 890 MW of nuclear power, October 7, 2026. https://esgnews.com/google-constellation-sign-20-year-deal-to-add-890-mw-of-nuclear-power/
[^5]: AI-Master.dev, Google Gemini vs ChatGPT vs Claude: which one for which use? (June 2025 model benchmark). https://ai-master.dev/google-gemini-vs-chatgpt-vs-claude-lequel-pour-quel-usage
[^6]: EnergyChoiceMatters, Google Constellation enter strategic framework to explore new generation announc, October 6, 2026. https://www.energychoicematters.com/news/google-constellation-enter-strategic-framework-to-explore-new-generation-announc

Note: the title in footnote [^5] was originally in French and has been translated; the URL slug was left untouched as requested.