The previous article explained why existing nuclear power capacity was suddenly repriced by the market. But talking only about nuclear power is actually incomplete, because AI’s real energy problem is not “will there be enough power ten years from now,” but “where will the electricity come from during this three- to five-year window?”
New nuclear power projects start with a cycle of ten years or more, and even SMRs will not have their first commercial unit until around 2030 at the earliest. But hyperscalers are announcing new GW-scale campuses every quarter right now, while GPU delivery cycles are only a matter of months. The time gap in between has to be filled by other power sources.
That is the three-path transition-energy story I want to talk about in this episode: gas turbines, SMRs, and energy storage.
Let’s start with gas turbines — this is the most real, most solid, and also the most underestimated main thread right now.
Natural gas power plants take two to three years to build, are flexible to start and stop, have large unit capacity, and their unit construction cost is an order of magnitude lower than nuclear power. For hyperscalers, this is the most realistic option for the transition period.
The market’s tension is already written into GE Vernova’s order book. This energy equipment company spun out of GE has heavy-duty gas turbine orders already booked through 2031; if you place an order now, you have to wait six years to take delivery. The company’s total gas turbine backlog exceeds 116GW, overall backlog has surged to $176 billion, and it is expected to reach $200 billion by the end of 2027. The buyers are familiar names like Amazon, Google, Microsoft, and Oracle.
This is not “traditional energy companies riding the AI narrative”; it is hyperscalers genuinely pouring orders into the transition period.
Now look at SMRs.
SMR is a new form of nuclear power, shrinking the reactor and making it modular, like building houses in a factory and then transporting them to the site for assembly. In theory, approvals are faster, costs are lower, and siting is more flexible, which makes it highly compatible with data centers.
Google, Amazon, and Microsoft have all signed SMR letters of intent, and stocks like X-energy, NuScale, and Oklo have all run up for a stretch. But the real delivery timeline for SMRs needs to be viewed calmly.
NuScale’s latest statement is that its first commercial unit may not be able to connect to the grid until as early as 2030, and the IEA’s view is that the world will only see the first SMR projects come online around 2030.
That means SMRs are a story for after 2030, and they cannot help with the power gap over the next three to five years. They are a candidate for the “endgame winner,” not the “transition-period winner.”
The third path is energy storage plus solar.
The combination of solar power and storage can independently support a substantial share of a data center’s load without connecting to the main grid. The construction cycle is short, and the cost curve is still declining.
Starting this year, the market has seen a very obvious shift: LFP battery BESS (battery energy storage systems) are replacing diesel generators and becoming the main emergency power source for AI server rooms. Diesel generators take more than ten seconds to start, and a GPU training job worth millions of dollars cannot wait; BESS can switch in milliseconds, and it can also work with solar power to shave peaks, fill valleys, and participate in grid frequency regulation.
Energy storage does not solve the baseload problem, but it smooths out the “sharpest” part of the entire power consumption curve and lowers the barrier to grid access.
These three paths, stacked together, are the complete picture of the transition period. Existing nuclear capacity is the ceiling, new supply relies on gas turbines to hold the line for three to five years, SMRs take over as the endgame after 2030, and energy storage acts as the buffer layer throughout the whole system.
When looking at AI energy investment, don’t just focus on nuclear power alone; gas turbines and energy storage are also major themes being repriced. The fact that GE Vernova’s orders are already booked through 2031 is, by itself, one of the hardest thermometers for AI infrastructure.