Big Tech Just Bought 10+ GW of Nuclear Power for AI. What That Means for Cloud, Latency, and Your Infra Roadmap.

The AI boom stopped being just a GPU story the moment big tech quietly started buying entire power plants.

Over the last year, Meta, Microsoft, and Google have signed enough nuclear power deals to rival the electricity use of small countries. Meta alone has lined up up to 6.6 gigawatts of nuclear capacity. Microsoft is effectively restarting Three Mile Island through a 20-year contract. Google has committed to a fleet of small modular reactors (SMRs) with Kairos Power.

This is not just energy news. It is a preview of where your future cloud regions will live, how reliable they will be under stress, and what "green AI" will mean in practice.

---

1. The Big Nuclear Deals, in Plain English

Meta: 6.6 GW to Feed Prometheus

Meta has assembled the largest nuclear portfolio so far:

• Three major deals with TerraPower, Oklo, and Vistra announced in January 2026.
• TerraPower:

- Two Natrium reactors (around 690 MW total), targeted for the early 2030s.

- Options for up to six more units, taking capacity to roughly 2.1 GW.

• Vistra:

- More than 2.1 GW from existing nuclear plants in Ohio and Pennsylvania.

• Oklo:

- A 1.2 GW advanced-reactor campus in Pike County, Ohio.

• A previous 20-year deal with Constellation for roughly 1.1 GW from the Clinton, Illinois plant.

Combined, Meta could be drawing up to 6.6 GW of nuclear power by the mid-2030s to feed its Prometheus AI supercluster and broader data centre footprint.

Microsoft: Three Mile Island, Rebooted for AI

Microsoft signed a 20-year, $16 billion power purchase agreement with Constellation Energy to:

• Restart the Three Mile Island Unit 1 reactor in Pennsylvania.
• Take essentially 100% of its ~837 MW output to power Microsoft data centres via the Crane Clean Energy Center.
• Target a restart around 2028, aligning with its next major wave of Azure AI expansion.

Three Mile Island is symbolically heavy, but practically it is a proven site with existing grid connections, cooling infrastructure, and community oversight.

Google: Distributed SMRs with Kairos Power

Google and Kairos Power struck the first large US corporate deal for a fleet of SMRs:

• Up to 500 MW across roughly six or seven reactors.
• First unit targeted around 2030, with the rest staged through the mid-2030s.
• Sited near key data centre clusters to provide firm, carbon-free baseload.

Instead of relying on a few massive plants, Google is betting on a more distributed nuclear mesh that can sit closer to edge and metro regions.

---

2. Why Nuclear, and Why Now?

Two hard constraints pushed hyperscalers in this direction:

• AI workloads are insanely power hungry: global data centre demand is expected to nearly triple between 2024 and 2035.
• You cannot run 24/7 GPU farms on solar curves alone without enormous overbuild and storage.

Nuclear offers:

• High-capacity-factor, always-on output.
• Very low direct emissions, crucial for net-zero pledges.
• Long planning horizons and stable pricing through multi-decade contracts.

For cloud providers, these deals:

• Lock in predictable energy costs for AI expansion.
• Strengthen their story to regulators and investors about "green AI".
• Give them leverage when negotiating new transmission lines and data centre permits.

---

3. How This Will Reshape Cloud Regions

Today, when you pick a region you mostly think about:

• Latency to your users.
• Data residency and legal requirements.
• Service availability and pricing.

In the 2030s, you will also be thinking about energy topology:

• Regions backed by dedicated nuclear PPAs will likely have more stable pricing during heatwaves and grid stress.
• Hyperscalers may market certain zones as "nuclear-backed" with stronger sustainability and uptime guarantees.
• Entire new region clusters will grow up near:

- Revived sites like Three Mile Island.

- New SMR campuses.

- Existing nuclear corridors in North America and Europe.

For SREs and architects, that means:

• Multi-region strategies will need to consider correlated grid and climate risks, not just distance on a map.
• You may prefer to anchor critical workloads in regions with firmer baseload instead of ones heavily exposed to intermittent supply.

---

4. Reliability, Risk, and Security Trade-offs

Nuclear-backed regions are not automatically safer. They move you into a different risk regime:

Pros:

• Plants are hardened, highly regulated assets with strict safety and cyber requirements.
• Operators run constant drills and maintain detailed incident response plans.

Cons:

• In some threat models, nuclear sites are high-value targets in physical or cyber conflicts.
• Grid interconnections between nuclear plants and data centres become part of your critical path.

From a cloud-customer perspective, the key is to:

• Understand which of your regions ride on which parts of the grid.
• Avoid single-region dependencies for workloads where downtime is existential.
• Factor in geopolitical, physical, and cyber risk alongside latency and price.

The AWS UAE drone strike incidents were a reminder that data centres can be pulled into kinetic conflicts. Nuclear-backed regions will be designed with that in mind, but the stakes are higher.

---

5. What This Means for "Green AI"

Marketing will increasingly lean on phrases like "nuclear-powered AI" and "carbon-free compute". The reality is nuanced:

• Lifecycle emissions from nuclear (including construction and fuel) are low and comparable to renewables per MWh.
• Waste, safety, and siting remain politically fraught and will shape where plants can actually be built.

As a cloud customer, the practical questions are:

• How does my provider account for nuclear in its carbon reporting?
• Can I choose regions that align with my organisation's ESG policies?
• Does "carbon-free" in marketing correspond to independently verifiable numbers?

If you are building tools around AI infra costs and sustainability — like region pickers or workload routers — this is fertile ground. Combining pricing data, carbon intensity, and latency into something humans can reason about is a real product opportunity.

---

6. How Developers and Infra Teams Should Respond

You cannot control what power plants hyperscalers buy. You can control how you design systems on top.

Practical steps:

• Make region choice an explicit design decision, not a default.
• Model energy and resilience risks when choosing where long-lived stateful services live.
• Build reliable multi-region failover for anything that truly cannot go down.
• Keep watching how your preferred providers talk about nuclear in their roadmaps and sustainability reports.

If you are in infra, SRE, or platform teams, adding "energy-aware architecture" to your skill set is a good way to future-proof your role. The stack now runs from physical reactors up through Kubernetes and LLMs; people who can see across those layers will be rare and valuable.

---

7. The Takeaway

Nuclear power for AI is no longer a speculative blog post; it is a signed set of contracts measured in gigawatts.

For most developers, this will show up gradually — as new regions, new marketing labels, and more discussions with clients about where their data actually lives. For infra and SREs, it will shape where you place your most critical workloads and how you think about risk.

The AI era is now deeply entangled with the physics and politics of the grid. Understanding that entanglement, even at a high level, will make you a better architect than anyone who still thinks cloud is "just someone else’s computer".