Different from the past hot products in the Internet era, software services related to artificial intelligence have a distinctive feature: in the process of development and continuous provision of services, such services will generate endless computing power requirements - and also power requirements. According to the latest news on Tuesday,As a leader in the global AI industry, Microsoft is viewing nuclear energy as a solution to long-term power problems, and is also using AI to solve the long-standing problem of nuclear energy approval.

Is the nuclear energy approval process long and expensive? Leave it to AI to solve it

Nuclear energy is gaining traction as the world moves towards carbon neutrality. Unlike photovoltaic and wind energy that investors are familiar with, the advantage of nuclear energy is that it can provide stable energy around the clock. But this energy source also comes with a very significant obstacle: the daunting and costly regulatory process for nuclear energy.

For example, the nuclear energy industry is currently accelerating the commercialization of "small modular reactors" (SMR).

Currently, only NuScale Energy's SMR design has received regulatory approval in the entire United States. It is reported that NuScaleThe entire approval process cost more than $500 million, and a total of12,000 pages of application forms, with more than 12,000 supporting documents2 million pages.

Against this background, although nuclear energy looks promising, the situation on Earth is that nuclear power generation continues to decline. The "Report on the Status of the World Nuclear Industry" released last week mentioned that,Nuclear power accounted for 9.2% of global power generation last year, the lowest level in nearly 40 years. As the construction costs of new energy projects gradually decline, the price of traditional fossil fuel power generation has also fallen after skyrocketing, making nuclear power a "less economical" option.

But for AI giants eager to obtain 24-hour stable and large amounts of electricity, nuclear energy is still an item that is difficult to give up in their long-term development goals.

It is reported that since half a year ago, Microsoft has deployed a team internally to train AI using documents on U.S. nuclear energy supervision. The goal is to train a special large model that can quickly reduce the time and cost of nuclear power plant construction approval.

Specifically, Microsoft is working with a nonprofit called TerraPraxis, which aims to convert old coal plants into small nuclear power plants. Microsoft provided the code, and TerraPraxis provided expertise in nuclear regulation.

TerraPraxis co-CEO Eric Ingersoll explained,AI can reduce the human writing time required to obtain approvals for new nuclear power plants by 90%. At the same time, for other new energy developers, AI may also reduce the time required for processes such as environmental impact assessments. To train AI using regulatory documents, Microsoft and the TerraPraxis team are breaking these documents into smaller parts.

"What we do is train (a large language model) on a very specific, highly structured document to generate another highly structured document that is almost identical to the previous document," Ingersoll said. Those "AI hallucination" problems will not arise in this process.

Kirsty Gogan, another co-CEO of TerraPraxis, further explained that the idea is to automate routine tasks so that professional regulators and developers can deal with "problems that require human beings rather than machines."

Michelle Patron, senior director of sustainability policy at Microsoft, concluded, “We are very excited about the potential of artificial intelligence to change the rules of the game in this area.

Microsoft leads the way and other “big electricity consumers” are also waiting and watching

As we all know, Microsoft founder Bill Gates invested in the small nuclear power plant start-up TerraPower as early as 2008. Since then, he has been a staunch supporter of nuclear energy. In his 2021 book How to Avoid Climate Disasters, he mentioned that “No other clean energy source is more reliable than nuclear energy, and no reliable energy source is cleaner than nuclear energy."

In the "Gates Notes" in May this year, he also mentioned that TerraPower is building a new nuclear power plant in Wyoming, which is expected to be put into operation in 2030. The former world's richest man revealed that the new nuclear power plant not only uses liquid sodium (instead of water) for heat dissipation, but also uses extensive use of supercomputers during the design process to simulate possible accidents to ensure that the nuclear power plant can remain safe under various circumstances.


In addition to looking ahead to nuclear power plants a decade from now, Microsoft's data centers now use nuclear power.

United Energy, which owns the largest number of nuclear power plants in the United States, reached an agreement with Microsoft in June this year to provide power supply "accurate to the second" to Microsoft's Virginia data center when wind power and photovoltaic supplies are insufficient. Joseph Dominguez, CEO of United Energy, praised that data centers and nuclear energy are a "match made in heaven."

United Energy also expects that with the outbreak of AI,Data center power needs will be 5 to 6 times greater than electric vehicle charging needs.

As a competitor of Microsoft, Google has not started to purchase nuclear power, but it has expressed interest in small nuclear power plants.

Maud Texier, global director of clean energy and decarbonization development at Google, said that nuclear energy, especially advanced nuclear energy, is making great progress. Texier also compared the cost of nuclear energy projects to wind power and photovoltaics 15 years ago, emphasizing thatCost reduction will become the driving force for deployment.

By the way, due to the current high cost of nuclear energy development, I am afraid that only "big financiers" like Microsoft can afford it. Although NuScale Energy, the company mentioned above, received government approval, the company's first commercialization project officially failed last month. Originally, NuScale's Utah project had the opportunity to become the first SMR in the United States, but due to rising development costs, many villages and towns chose to withdraw from the project.

The reason is that NuScale predicted in January this year that the target price of this nuclear power plant would be US$89/MWh, an increase of nearly 50% from the previously expected US$53/MWh. The local power company was worried that not enough customers would pay for it.