CRITICAL EVENT UPDATE · AI Infrastructure × Power × Nuclear
Google Signs 3.59 GW of Power Agreements as AI Industrialization Enters the Bring-Your-Own-Power Phase
Critical Event Update | 3.59 GW Supply × 890 MW Nuclear Uprates × Bring Your Own Power | 7 October 2026
Google and Constellation Energy signed long-term arrangements covering 3,590 MW. Of that total, 890 MW is additional output from uprates across 11 operating nuclear units, with first delivery expected in 2028; another 2,700 MW is a PJM supply agreement not tied to one generation source. The deal confirms that AI expansion is writing power supply, financing and grid risk into long-dated contracts. It does not mean all 3.59 GW is new nuclear capacity or online today.
Combined scale of the two supply arrangements
Additional output from nuclear uprates
Expected Constellation investment
Operating nuclear units scheduled for upgrades
3.59 GW contracted | 890 MW new nuclear output | First delivery expected in 2028
01 · RESEARCH BRIEF
The one-minute brief
Google announced a 20-year power-purchase agreement that gives Constellation revenue certainty to uprate 11 units at six operating nuclear plants, adding 890 MW by the end of 2032. Constellation expects to invest more than $4.3 billion.[1] Reuters also confirmed a separate 15-year, 2,700-MW PJM supply agreement not tied to a specific generation source.[2] The structure responds to PJM's bring-your-own-power direction: large data-center loads increasingly need to secure dependable supply alongside their grid connection.
Audio transcript
Google and Constellation signed long-term arrangements covering three thousand five hundred ninety megawatts, including eight hundred ninety megawatts of new output from nuclear uprates. The deal connects AI power, financing and grid access, but the capacity has not yet been delivered and is not the same as usable compute.
Known facts and open questions
- Confirmed
- 20-year PPA supporting 890 MW of nuclear uprates
- Confirmed
- 15-year, 2,700-MW PJM supply agreement
- Confirmed
- First new nuclear output expected in 2028
- Unconfirmed
- Power price, project returns, final cost and data-center load mapping
Demand
Google data-center growth creates contractable long-duration load
Finance
Long-term offtake raises revenue certainty for power assets
Supply
Existing nuclear uprates can arrive faster than new plants
Grid
PJM links large new loads to dependable supply
Industry
AI, power, nuclear engineering and cloud software converge
02 · FACTS → IMPACT → VIEW
Why does this change matter?
AI industrialization is constrained by financeable, deliverable power as well as chips
- What is confirmed
- ACIS has argued that the AI Industrialization Stack must answer who funds it, who powers it and who captures the return, with time to power as a central delivery gate.
- Why it matters
- Google is using 20-year offtake to support 890 MW of nuclear uprates and adding a 15-year, 2,700-MW regional supply arrangement while PJM pushes large data-center customers to bring dependable power.
- ACIS view
- Who powers AI has moved from planning into long-duration contracting and asset finance. Platforms with operating assets, permits, grid positions and credible offtakers gain relative value, but undelivered megawatts are not usable compute.
03 · EVIDENCE & ANALYSIS
Evidence and analysis
01｜What Happened
Google and Constellation signed two arrangements: a 20-year PPA supporting uprates across 11 units at six nuclear plants, adding 890 MW by the end of 2032, and a 15-year agreement covering 2,700 MW of PJM supply. Constellation expects to invest more than $4.3 billion, with first incremental nuclear power expected in 2028.[1][2]
02｜Why It Matters Now
PJM capacity prices have increased more than elevenfold since 2024 as data-center demand contributes to a power shortfall. The agreement responds directly to bring-your-own-power: AI campuses must increasingly secure generation, financing and grid responsibility before the computing capacity can be deployed.
03｜Confirmed Facts vs Uncertainty
Terms, capacities, unit count, investment and target delivery were confirmed by Google and Reuters. The full 3.59 GW is not new nuclear capacity: 890 MW is incremental nuclear output, while 2,700 MW is not tied to one source. Pricing, returns, final cost, approvals, exact data-center loads and full on-time delivery remain undisclosed or unfinished.
04｜Transmission Mechanism
AI data-center load growth → hyperscaler signs long-term offtake → generator gains revenue certainty and funds nuclear uprates → operating assets release capacity faster than new plants → time to power improves → deployable compute expands. Engineering, nuclear fuel, equipment, storage and grid-service demand also rise, while unsigned loads face a higher interconnection hurdle.
05｜Prior ACIS View → New Evidence → Updated View
The prior view was that AI expansion must solve funding, power and cash returns together. The new evidence converts Google's long-term cloud and AI load into financeable power offtake and uses operating nuclear assets to shorten delivery. The AI buildout is moving from GPU procurement to an integrated power-grid-compute-revenue system.
06｜Cross-Asset / Cross-Industry Read-through
Nuclear operators, turbine and controls upgrades, engineering, transmission and storage gain demand support. Power-contract duration and cost become inputs into hyperscaler AI margins and capital efficiency. Whether costs shift to other PJM customers still needs validation. Google will also deploy Gemini Enterprise in Constellation's operations, showing AI both consuming power and helping optimize power assets.
07｜What Does NOT Change
The deal does not mean 3.59 GW is already built, dedicated entirely to Google or all nuclear. Uprates still require engineering and regulatory work; the 2.7 GW is not new nuclear output. Strong AI demand is supported, but project value still depends on delivery, unit power cost, utilization and cash flow.
08｜Risks / Alternative Scenarios
Base: first uprates arrive in 2028 and scale through 2032. Upside: the structure becomes a repeatable model for operating nuclear assets and long-term corporate offtake. Downside: cost overruns, approval delays or equipment constraints. Alternative: inference efficiency, flexible load, gas, storage or other resources reduce dependence on a single nuclear path.
09｜Next Validation
Within 24 hours: Constellation disclosures and PJM rules. In seven days: comparable offtake agreements, equipment orders and pricing implications. In 30 days: permitting, capital spending, the 2028 delivery path and the mapping between Google load and data-center regions.
04 · INVESTMENT IMPLICATIONS
Industry and asset implications
AI infrastructure
Power moves from operating procurement into project-finance architecture.
Nuclear and engineering
Operating-plant uprates gain large-customer offtake and capital support.
PJM grid
Large-load supply responsibility and interconnection hurdles rise.
Cloud and software
Power cost and delivery link more directly to AI margin and returns.
The contracts and investment direction are confirmed. Incremental, on-time grid capacity available for AI loads still needs year-by-year validation.
05 · VALIDATION & RISKS
What to watch next
24 hours
Company disclosures align and PJM path is clear
What would weaken the view: Material differences emerge in capacity or cost allocation
7 days
Comparable offtake or supply-chain orders follow
What would weaken the view: Regulatory or grid opposition becomes material
30 days
Permits, equipment and delivery milestones are traceable
What would weaken the view: Cost, approvals or timing deteriorate materially
What would change our view?
The main analytical error is to label all 3,590 MW as new nuclear capacity or to treat a contract as delivered compute. The 890-MW uprate and 2,700-MW regional supply agreement must remain separate, while cost, delivery and utilization require independent validation.
06 · FAQ
Key questions
Is all 3.59 GW new nuclear power?
No. The 890 MW is incremental output from operating nuclear plants; the other 2,700 MW is long-term PJM supply not tied to one source.
Why do operating-plant uprates matter?
They use existing sites, licensing foundations and grid connections to release dependable capacity faster than a new reactor, although engineering and approvals are still required.
Does this solve the AI power bottleneck?
No. It validates a financing and contracting route, while delivery, cost, grid connection and load growth remain open tests.
07 · TERMS & SOURCES
Terms, sources and related research
Key terms
- Nuclear uprate
- Equipment and control-system upgrades that increase output from an operating reactor.
- PPA
- A long-term power-purchase agreement that provides revenue certainty for generation investment.
- Bring Your Own Power
- A requirement or market direction that large new loads secure dependable supply so they do not undermine grid reliability.
[1] Google｜Why we're backing America's existing nuclear plants｜2026.10.06 ↗
[2] Reuters｜Google enters 3.59-GW power deal with Constellation Energy｜2026.10.06 ↗
This report uses Google's 6 October 2026 announcement and Reuters reporting. Capacity, duration, investment and target delivery are published plans. Pricing, returns, final cost, approvals, data-center mapping and actual grid delivery remain unverified.