AI Data Center Power Demand Breaks Critical Point
On July 30, 2026, the US Nuclear Regulatory Commission (NRC) formally approved the first commercial license for a small modular reactor (SMR), marking a new phase in nuclear renaissance. The same day, Microsoft announced a 20-year Power Purchase Agreement (PPA) with SMR operator NuScale Power to supply stable clean electricity for its hyperscale AI data center cluster in Virginia. This series of events reveals a trend: explosive AI computing demand is forcing tech giants to turn to nuclear, a round-the-clock carbon-free power source.
The Power Hungry of AI Training and Inference
According to a recent report by the US Energy Information Administration (EIA), by 2027, US data centers will consume over 8% of total national electricity generation, with AI data centers as the main driver. The NVIDIA H100 GPU peaks at 700W, while the new Blackwell architecture B200 GPU exceeds 1000W. A typical 100,000-GPU AI cluster has a peak power demand near 100 MW, equivalent to a small city's load. Grid capacity warnings have emerged in data center clusters like California and Virginia, with new project interconnection wait times extending to 3-5 years.
The SMR project locked by Microsoft is in southwestern Virginia. Each reactor has a rated power of 77 MW, with the first deployment of 4 units totaling about 308 MW, supporting full load of about 30,000 GPUs. NuScale's VOYGR series SMR uses pressurized water reactor technology, occupying only one-tenth the area of conventional large nuclear plants, and can be modularly expanded as needed, ideal for colocation with data centers.
Wall Street Reprices Nuclear Assets
After the news, US nuclear stocks collectively surged. NuScale Power (NYSE: SMR) rose 14.5% on the day, hitting an all-time high. Traditional nuclear operator Constellation Energy (CEG) rose 6.2%, and Vistra (VST) followed with 4.8%. Analysts note that nuclear stocks have been suppressed by inflation and policy uncertainty over the past two years, but AI data centers' rigid demand is rewriting valuation logic.
Goldman Sachs upgraded SMR sector ratings in its latest report, believing the "marriage of AI and nuclear" will create a hundred-billion-dollar incremental market. According to its model, by 2030, over 15 GW of SMR will be dedicated to AI data centers in the US, corresponding to ~$30 billion in capital expenditure. Additionally, uranium miners like Cameco (CCJ) and uranium ETFs (URA) have also gained traction, as SMR requires 1.5 times more low-enriched uranium than traditional large reactors.
Beyond Nuclear: Synchronous Grid and Cooling Upgrades
Nuclear alone cannot solve all problems. AI data center power also needs supporting grid infrastructure and efficient cooling technology. Benefiting from this logic, grid equipment makers like Quanta Services (PWR) and Vertiv (VRT) also attract capital. Vertiv's latest quarterly report shows its data center thermal management revenue grew 45% YoY, with liquid cooling solutions contributing half the growth.
Notably, Microsoft's PPA requires NuScale to provide "dispatchable" power, meaning real-time output adjustment based on AI workload fluctuations. Traditional nuclear units typically run at constant power, but SMR design is more flexible. NuScale's modules can be adjusted within 20%-100% power range, with ramp rate up to 5% per minute, offering potential to match AI training and inference power price changes.
Investment Insights: Positioning for AI Infrastructure 'Second Curve'
Reviewing US AI computing market, the past two years focused on GPU chips (NVIDIA) and AI servers (Super Micro Computer). But as chip efficiency improvements slow, physical infrastructure like power, cooling, and land become new bottlenecks. From the supply chain perspective, segments like nuclear plants, grid upgrades, liquid cooling, and backup batteries have larger flexibility and relatively reasonable valuations.
The Aspen Institute Energy Policy Director noted: "Data center operators' insatiable appetite for power will drive nuclear regulatory reform and project finance innovation. In the next two years, 5-10 SMR projects may start simultaneously." For US investors, three directions: nuclear operators/developers (NuScale, Constellation, Vistra); grid & distribution equipment (Quanta, Eaton, ABB); data center cooling and power management leaders (Vertiv, Schneider Electric).
Risk Disclaimers
- SMR projects face construction cost overruns; first unit commercialization requires time to verify.
- Nuclear fuel supply chain may be constrained; low-enriched uranium capacity expansion is slow.
- If AI computing demand cyclically declines, power agreements may face renegotiation.
Overall, the combination of AI and nuclear marks the entry of computing infrastructure into a 'heavy asset' era. This trend will not only reshape US investment landscape but may also force a global energy policy shift. Subsequent attention should be paid to whether more tech giants like Google and Amazon will follow suit, and the Department of Energy's regulatory direction on data center carbon emissions.
