AI’s power consumption has grown to the point where even the grid is starting to gasp for breath. And that has put an unexpected protagonist in the headlines: nuclear power. Tech giants like Amazon, Google, and Microsoft have all gone off in recent years to invest in or sign nuclear deals, and the hottest keyword among them is the SMR.
This piece spells out the SMR. First what it is and how it differs from conventional nuclear, then why AI needs it, how far the giants have gotten, and one pragmatic reminder: it’s really an option for after 2030. This is the energy deep-dive of the data center and power gate.
What Is an SMR?
SMR stands for Small Modular Reactor. The International Atomic Energy Agency generally defines it as a reactor whose single module has an electrical output of no more than 300 MWe.
Set it against a conventional large nuclear plant and it gets clear. A conventional plant is often 1,000 MW or more per unit, requiring long on-site, custom construction that takes well over a decade to build. The SMR goes the other way, down the “small and modular” road: standardized modules are prefabricated in a factory, shipped to the site for assembly, and can even be added in phases like building blocks.
Here’s an analogy: conventional nuclear is like constructing a high-rise on site from the ground up; the SMR is like shipping prefabricated container units to the site and snapping them together — in theory faster and more standardized. But note that “modular means cheaper” is still an engineering assumption for now; it only counts once the first batch of plants actually delivers.
Core-Data Snapshot
A few numbers below to help you grasp the scale of SMRs and AI power. Most are institutional projections or target values and will shift as things progress.
| Topic | Data | Timing / Nature |
|---|---|---|
| Global data-center electricity use | About 415 TWh in 2024 (around 1.5% of the global total), estimated at about 945 TWh in 2030 | IEA projection |
| First batch of SMRs in commercial operation | Most estimates put it from around 2030 | IEA assessment / developers’ targets |
| U.S. tech sector’s planned SMR | More than 20 GW already planned/financed | IEA tally of announced projects |
| NuScale design | 77 MW per module, 462 MW total across 6 modules; design approved by the NRC in 2025 | U.S. Nuclear Regulatory Commission |
| Taiwan’s earliest SMR power | Estimated around 2035 (authorization and review required first) | National Atomic Research Institute estimate |
Why AI Data Centers Are Starting to Snap Up Nuclear Power
The key is that AI’s power consumption is growing too fast — and is too picky.
Look at the volume first. The International Energy Agency estimates that global data-center electricity use will grow from about 415 TWh in 2024 to about 945 TWh in 2030, and in the U.S. roughly half of the added power demand over the next few years comes from data centers alone. At that pace, grids in many places have started falling behind.
Now look at “picky.” AI server rooms don’t just want lots of power — they want it 24 hours a day without interruption, low-carbon, and lockable under a long-term contract. Wind and solar live at the mercy of the weather, natural gas carries emissions, and nuclear happens to fill that slot: stable, low-carbon, and contractible long-term. That’s why the tech giants have started turning their minds to nuclear.
SMR Developers and the Giants’ Nuclear Plays
Lay out the main players and you find one thing in common: everyone is on the road, but no one is truly supplying power at large scale yet.
SMR developers: NuScale’s design has been approved by the U.S. Nuclear Regulatory Commission (77 MW per module); Oklo and X-energy are mostly still in the pre-application and design phases; Kairos Power’s Hermes 2 demonstration plant broke ground in April 2026 and will, through the Tennessee Valley Authority’s grid, supply power for the electricity demand of facilities supporting Google’s data centers; TerraPower has obtained a construction permit, though its Natrium is about 345 MW, strictly speaking already above the common 300 MW SMR threshold.
The tech giants’ collaborations: Amazon is investing in X-energy, targeting more than 5 GW of new nuclear before 2039; Google has a deal with Kairos, targeting deployment of 500 MW by 2035, with the first batch online in 2030; Microsoft signed a 20-year contract with Constellation to restart the Three Mile Island plant (about 835 MW, which is a restart of existing large-scale nuclear, not an SMR). What to see clearly is that most of these are investment or procurement agreements, which is not the same as the power already being in place.
One Pragmatic Reminder: This Is an Option for After 2030
The SMR story is compelling, but you have to get the timeline right.
The real state of things in 2026 is this: a handful of developers have obtained design approvals or construction permits, individual demonstration plants are breaking ground, but most are still at the permitting, design, fuel, and supply-chain stage. The first batch of commercial power mostly lands after 2030. In other words, the power gap for AI server rooms from 2026 to 2030 still has to be filled mainly by grid expansion, natural gas, renewables, the restart and life extension of existing nuclear, plus storage and demand-side dispatch.
So the more pragmatic view is this: the SMR is a “long-term option for the 2030s” for AI power, not a solution you can plug in over the next year or two. On this topic, patience matters more than haste.
Taiwan’s Role
Here in Taiwan, nuclear is a fairly sensitive issue on the policy front. Maanshan Unit 2 was shut down in May 2025 when its 40-year operating license expired.
The National Atomic Research Institute has launched SMR research, but by its own assessment, even if international light-water SMRs commercialize smoothly and Taiwan can obtain foreign authorization, the most optimistic case would still be 2035 before it could generate its first kilowatt-hour of SMR power — and Taiwan currently lacks the system-design capability for the reactor body itself. The more likely role is to enter on heavy-electrical equipment, transmission and distribution, precision machinery, materials, and some nuclear-grade components (transformers, switchgear, cables, control systems, for example).
The market often puts heavy-electrical and grid-upgrade companies into the “nuclear theme stock” discussion, but this is closer to a “grid and supply-chain theme”; being named is not the same as having secured a nuclear order, nor does it constitute investment advice. Just treat it as one piece in understanding the energy transition.
Key Takeaways for This Gate
After looking at the SMR, first remember its positioning: when AI server rooms eat power to the point the grid gasps for breath, nuclear gets put on the table for being stable, low-carbon, and contractible long-term, and the SMR is the hottest option among them.
But keep the timeline clear: in 2026 it’s still in the permitting, demonstration, and construction phase, and the first batch of commercial power mostly has to wait until after 2030. The power gap over these next few years still has to be carried by the grid, natural gas, renewables, and existing nuclear. Taiwan’s nuclear policy and nuclear-safety review are still in progress, and the part more likely to participate in the short term is heavy-electrical equipment and supply-chain components, not the reactor body itself.
To see the overall power and compute ceiling of AI server rooms, go back and read the data center and power gate; to see all eight gates of the chain, head back to the supply-chain overview.