Has DeepSeek Popped the “Mini Nuke” Bubble for AI Power?

Originally published at Forbes.com

Artificial intelligence is all the rage, and the money is flowing. The White House recently announced a potential $500 billion investment in AI infrastructure from a new partnership between OpenAI, Oracle, and SoftBank.

The goal? Build out data centers and ramp up the electricity generation needed to make AI work for hundreds of millions of global users. The demand for power is massive: According to Goldman Sachs, data center power demand will grow 160 percent by 2030, with a single ChatGPT query using about 10 times more power than a simple Google search.

To meet demand, Santee Cooper—the largest power provider in South Carolina—now vows to restart construction on a pair of nuclear reactors, while Indiana lawmakers hope to boost nuclear power in the Midwest. Proposed legislation out of the Hoosier State, where I serve as Poling Chair at Indiana University’s Kelley School of Business, would provide a tax incentive for businesses to manufacture small modular reactors (SMRs) for nuclear power.

Is this the future? So-called “mini nukes” powering the future of AI?

I’ve spent more than 25 years as a director of now connected major energy utilities, including Nicor Gas, which merged with Atlanta Gas Light and then got rolled up into the Southern Company Gas, so people often ask me about the potential for SMRs to meet today’s energy demand. Talking to people in the industry, they are generally more bullish on the future of a less power-intensive AI that produces less heat and consumes less energy than they are that mini nukes are the solution. 

More broadly, the emergence of DeepSeek this week is a reminder that energy efficiency is a better bet than one of the largest energy production ramp-ups in human history. DeepSeek is the opening act in the final solution, which is delivering the same AI capabilities at a fraction of the cost. DeepSeek's AI models seem to be faster, smaller, and a whole lot cheaper, necessitating less energy than U.S. rivals.

This is what I suspected, and it matches what I hear when talking to the top players: The SMR approach feels like a crapshoot. In short, the AI power bubble may end up being a borderline con—not unlike the SPAC bubble and bust from the COVID-19 pandemic era. While ideas for nuclear development are a dime a dozen, “mini nukes” often look more like “paper nukes,” meaning that they have no product, no actual design, no technology that has been tested or vetted, and no progress on regulatory approval. More likely than not, these paper nukes won’t even present an actionable product until 2040 or 2050.

To understand the likeliest course of action in energy production, it is important to remember the big picture: The companies that need this new quantum of power are the world’s largest technology firms, with market capitalization values of $1 trillion or more. The largest nuclear operating utilities are “only” in the $100 billion market cap range, and they are all licensed and regulated by the states in which they operate. The largest tech companies will spend what they need on AI-related investment, primarily on natural gas generators that are owned privately, can be brought online relatively quickly, and remain dedicated to single customers or geographies.

Who will partner with Big Tech? It will be The largest oil companies, which possess the fuel supply, know how to put up natural gas generators quickly, and are not subject to utility regulations. These companies can locate generators near data centers without worrying about the public grid—the likes of Exxon Mobil and Chevron are already active on this front. The best current estimate is that natural gas will provide at least 60 percent of the power demand growth from AI.

There is a place for “mini nukes,” but my guess is that it will be for utilities to add increments of carbon-free capacity with more manageable increments and pre-approved designs. One company, General Electric (namely, GE Hitachi Nuclear Energy), is closer to having the first working SMR, although GE’s definition of “small” is 500 megawatts. This is about half the size of a typical nuclear utility’s Generation 3 unit of 1,000 megawatts—currently being upgraded to reach 1,150 megawatts—so GE’s SMR won’t be fast enough for the AI industry, and the Canada-based project is not even finished yet.

The best guess is that the magnificent seven technology companies pay up to piece together a variety of power sources—natural gas, nuclear, and others—to maintain their leadership positions during this next phase of AI development, with only three or four survivors dominating the AI market in the end.

Perhaps DeepSeek will have a lasting seat at the table, beating out U.S.-based competitors. The Chinese company is already telling us that, beyond the particulars of natural gas versus nuclear, the most likely solution to the power problem is reengineering how the data training is done and how the AI chips are designed, given that the power demands of current technology extrapolated out is beyond most scenarios of what today’s energy technologies can provide in a realistic time frame. There is no obvious near-term solution to the extrapolation of power demand—because the demand is unprecedented.

How can you use less energy in the first place? The first shot in the AI war has been fired. Now, let’s see how the future of AI efficiency unfolds.