With the rapid development of generative AI, hyper-sized models, and cloud-based inference, global data center electricity consumption has exploded in just a few years, from 460 TWh in 2022 to 1,000 TWh by 2026, which is almost equal to the electricity consumption of Japan in an entire year. Driven by this trend, the energy supply and demand game has escalated, and the decision of New York State to build at least 1 GW of small modular reactor (SMR) nuclear power plants announced on June 23, 2025 has become a hot focus in the industry and policy circles.
In this article, China Exportsemi will systematically interpret this event from the five dimensions of data, power market, technology, industry and policy, put forward professional insights, and discuss its far-reaching significance to the semiconductor and AI industry.
Ⅰ Electricity Consumption in the AI Era: The "Power Monster" of Data Centers
Global and U.S. pressures
According to the International Energy Agency (IEA), data centers used about 460 TWh of electricity in 2022, accounting for about 1–1.3% of global electricity demand that year. According to this trend, consumption is expected to reach 1,000 TWh (+117%) by 2026, almost the same as Japan's total electricity consumption. Data center electricity consumption in the U.S. is expected to grow from 200 TWh to nearly 260 TWh by 2026, accounting for 6% of the country's electricity demand, and AI forecasting services will explode within this.
In addition, the energy consumption of a single request in the AI inference process is nearly 10 times higher than that of a typical web search. For example, ChatGPT consumes a few watt-hours to tens of watt-hours in a conversation, which also makes AI an important factor driving the soaring power consumption.
Profiles of the semiconductor industry
From the perspective of semiconductors, AI inference requires higher performance, lower latency, and consistent power supply for chips. NVIDIA's latest server card racks can consume up to 240 kW. To this end, semiconductor giants, AI cloud fabs, and infrastructure distributors are working together to redefine the standard for data center power supply, upgrading from 250 kW/rack to 1 MW/rack.
Pictured: Vogtle Nuclear Power Plant in Georgia, USA
Ⅱ Why did SMR become a "panacea" for New York?
SMR definition and technical advantages
Small modular reactors (SMRs) usually refer to small reactors with an output of 100 megawatts, which have multiple advantages such as modular factory prefabrication, rapid deployment, baseload + peak shaving adaptability, and advanced safety design. It uses free cooling and passive safety systems to effectively avoid the risk of accidents in traditional large-scale reactors.
Cost & Build Responsive
In the past, large-scale nuclear power projects, such as Vogtle 3/4 in Georgia, took more than 15 years and invested more than $17 billion, far exceeding the budget and seriously lagging behind. In contrast, SMR design refinement and modular production are expected to greatly reduce the cost and construction period per megawatt.
In addition, New York has announced that NYPA will move forward with the site selection, evaluation and permitting process over the next 12 months to actively connect with private sector resources, including technology vendors and capital.
Ⅲ Policy and industry promotion: Halo continues to go strong
State federal policy linkage
Based on the urgent demand for electricity in the AI and semiconductor industries, Governor Kathy Hochul proposed an "energy sovereignty" strategy, striving to combine energy and electricity, reduce dependence on fossil fuels, and ensure controllable electricity bills and reliable supply.
At the federal level, the Biden administration has promoted the Small Reactor + Nuclear Capacity Triple Program through the Inflation Reduction Act (IRA) and the President's Executive Order to provide tax, subsidy, and licensing support for SMRs. If the bill falls through in the future or the budget is reduced, it will cause disruption to the progress of the project.
Tech giants form alliances
Microsoft, Google, Amazon, Meta and other tech giants join the SMR track:
* Microsoft partners with Constreschen to restart the Three Mile Island nuclear station, aiming to provide 100% power supply for 20 years;
* Google signs an upfront power supply agreement with Kairos Power;
* Amazon invested in X-Energy and Meta and reached various expansion agreements.
This "supply and demand integration" model will greatly enhance the financing capacity and stable demand support of SMR projects.
Ⅳ Opportunities and challenges in the semiconductor and AI industries
Charging opportunities: Stable power drives high growth
The semiconductor industry regards electricity as its core infrastructure. From chip manufacturing to high-performance computing, every watt of power can support billions of dollars in investment. This SMR construction is expected to achieve more wafer fabs, AI training centers and high-performance cloud computing bases, which is a visible "infrastructure dividend" in the industrial chain.
Investment considerations: cost, time and regulatory thresholds
Despite the potential of SMR, it is still in the demonstration and licensing stage, and it will take 5-7 years for it to be commercialized and scaled. In the interim, cost, regulation and public attention will be uncertain variables.
Paths to success include: rapid demonstration, strict control of engineering accuracy, transparent establishment of trust mechanisms, and response to public concerns (scrap, nuclear safety, etc.).
Ⅴ Sustainable strategy: Build a "nuclear + wind and solar + storage" energy mix
The wave of electricity in the AI era is not the pursuit of a single energy source, but the reconstruction of the energy structure. Although photovoltaic, wind power, and energy storage can alleviate some of the fluctuating loads, they lack all-weather baseload capacity. SMR makes up for this gap and is more suitable for the "high-intensity and continuous" power demand of semiconductor fabs/AI centers.
A more realistic framework is the "ternary power supply" strategy: baseload SMR + renewable + long-term energy storage (battery/green hydrogen). This is essential to ensure power load resilience and system reliability.
Ⅵ Risk Warning and Regulatory Advice
1. Risk of demonstration failure: If the on-site SMR technology is less effective than expected, it is easy to cause a double collapse of public opinion and investment confidence;
2. Public trust challenges: issues such as nuclear waste and emergency response need to be transparently communicated and supervised by third parties.
3. Policy continuity risk: If the federal policy is adjusted after 2026, it will affect the state-level advancement;
4. Possibility of market mismatch: Rising electricity prices or declining subsidies may lead to a decline in the ROI of AI data centers.
Suggestion: The project should give priority to promoting the "pilot module" and expand the capacity in stages; At the same time, combined with industry-university-research cooperation and high-reliability supply chain construction, we will strengthen project information disclosure and community communication.
Ⅶ Conclusion
* Data & Trends: AI is driving data center power consumption to 1,000 TWh, and the semiconductor industry urgently needs long-term reliable power supply;
* Technology choice: SMR is a rational choice under realistic needs that takes into account both safety and controllability and scheduling flexibility.
* Policy-driven: U.S. federal linkage accelerates the introduction of capital and technology synergies for SMRs;
* Industry impact: Nuclear power + semiconductors + AI to build a new industrial convention, adding to the assertion that "electricity is the new oil";
* Risks and challenges: Regulation, transparency, community engagement and interdisciplinary collaboration need to be strengthened.
Overall, the SMR nuclear power plant project in New York is a primary energy support test for the AI and semiconductor industries to enter a period of high growth, and also represents a landmark case of the U.S. energy policy leap from traditional to technology-driven energy system. The next 12-24 months of licensing progress, demonstration project construction, capital structure implementation and other nodes will determine whether the model can be replicated in other states and even the global semiconductor and AI highlands.
