Samsung C&T is backing Kairos Power with up to $100 million so the company can build a nuclear reactor for Google, and the deal is a clean look at what AI data center energy costs actually involve once you get past the price of a GPU.
The investment splits into $70 million of equity and about $30 million of engineering services, aimed at two reactors rising in Oak Ridge, Tennessee.
Hermes 1 is a low-power demonstration unit. The commercial-scale version, Hermes 2, is cooled by fluoride salt and fueled with ceramic-coated uranium designed to resist meltdown.
Google signed its side of the deal in fall 2024, calling for about half a gigawatt of nuclear power within roughly a decade. Hermes is set to supply the first slice of that commitment.
Samsung C&T brings something else to the table too: it has already built about a dozen reactors elsewhere. So the real risk in this project is schedule, not unproven engineering.
AI Data Center Energy Costs Are Bigger Than the Chips
Global data center electricity use hit about 415 terawatt-hours in 2024, or roughly 1.5% of everything the world consumes. That share has been climbing far faster than overall electricity demand for years now.
The Brookings Institution puts U.S. data centers at 45% of that global total, with the International Energy Agency's base case projecting American demand to grow 130% by the end of the decade.
Numbers like that explain why Google isn't the only company signing power deals instead of just building bigger data halls. Amazon bought a Pennsylvania campus next to the Susquehanna nuclear station specifically so it could draw power straight off it.
Microsoft struck a deal to bring Three Mile Island back online. A French operator, Data4, locked in a slice of nuclear output from EDF under a long-term contract. Reactors are turning into a procurement line item, the same way GPUs already are.
How the Big Nuclear Deals Compare
Lined up next to each other, the deals share a shape: buy or build access to a specific reactor rather than wait for the grid to catch up.
| Company | Nuclear deal | What it buys |
|---|---|---|
| Up to $100M into Kairos Power via Samsung C&T | 50MW now from Hermes, aiming for 500MW by 2035 | |
| Amazon | Campus next to the Susquehanna nuclear plant | Direct power draw from an existing reactor |
| Microsoft | Three Mile Island restart agreement | Reviving idle nuclear capacity for its own use |
| Data4 | 12-year contract with EDF | 40MW of nuclear power over the contract term |
What Rising Power Costs Mean for Businesses Buying AI
None of this shows up on an invoice yet. Cloud providers are still absorbing the capital cost of reactors, chips and cooling rather than passing a line-item power charge to customers.
Microsoft's own numbers show the scale of what's being absorbed. It reported $30.9 billion in property and equipment additions for a single fiscal quarter, and $80.1 billion for the first nine months of its fiscal year, according to an earnings report cited by Virtualization Review.
Spending at that scale doesn't stay invisible forever. It surfaces eventually in token pricing, in which models get first claim on cheap compute, and in which regions get new data center capacity at all.
Planning an AI budget for a fintech company or easing a real estate business into AI already means treating compute access as a real constraint, not a given. Our AI and automation work with clients starts from that same premise.
The Hermes reactors won't come online before 2030. Even so, the pricing pressure behind them is already here, and it's worth budgeting for AI as though the infrastructure underneath it is scarce, because increasingly, it is.
Cover photo by Mikhail Nilov on Pexels
