Nuclear energy could power the AI boom—but only if proliferation risks are minimized

On May 10, Oklo Inc., a nuclear energy startup, began trading on the New York Stock Exchange. The chairman of the company is none other than Sam Altman, the CEO of the artificial intelligence leader, OpenAI, that launched the generative AI revolution with the release of ChatGPT late in 2022. The staggering language, image, and video processing capabilities of ChatGPT and other similar chatbots depend on large language models (LLMs) trained through computations of data at unprecedented scale. Computational power has thus become, in Sam Altman’s words, “the currency of the future” and “the most precious commodity in the world.”

The training and use of LLMs require huge amounts of electrical power and cascades of advanced microchips. Altman’s nuclear investment reflects his belief that Oklo’s microreactors can satisfy the future power requirements of AI models. In some ways, the compatibility is intuitive. Large data centers, especially those set up in remote areas with greater land availability, require power sources that avoid both the intermittency of renewables and the fuel delivery requirements of traditional thermal power plants. Such is Oklo’s narrative as it touts recent power purchase agreements with data center operators such as Wyoming Hyperscale.

But the intuitive compatibility between AI and nuclear power does not exempt the latter from traditional concerns about economics and safety, despite efforts by Oklo and other vendors of “advanced” reactors to downplay those concerns. The escalating costs of NuScale’s first VOYGR plant have cast into doubt the economics of small modular reactors, and the US nuclear regulator has yet to endorse the safety of Oklo’s Aurora microreactor, having denied its combined license application in early 2022.

Arguably the most problematic aspect of Oklo’s microreactor concept is the proliferation implications of its fuel cycle. Simply put, Oklo’s concept could increase the availability of fissile materials needed for nuclear weapons and, therefore, the likelihood that bad actors would acquire those materials.

Oklo’s Aurora microreactor uses uranium fuel enriched so that 19.75 percent of it consists of the fissile uranium 235 isotope, compared to the 5 percent low-enriched uranium fuel currently utilized by industry. The higher enrichment level allows the reactor to have smaller cores and longer refueling cycles, but it also means that the fuel is far closer to the level of enrichment needed for nuclear weapons than what is used in traditional power reactors. That means a malicious actor with access to the higher enrichment fuel would be able to produce weapons-grade uranium using less diverted material, in less time, and with lower detectability.

Even more problematic: The Aurora microreactor’s sodium-cooled fast-neutron configuration is ideal for the breeding of plutonium 239, which could be employed directly in nuclear weapons once separated from spent fuel. Indeed, the CFR-600 reactor that the US government claims China is using to produce plutonium for an expanded nuclear arsenal is of the same sodium-cooled fast breeder design. Oklo’s vision is not to make a nuclear weapon but rather to extract energy from the plutonium after it is reprocessed into fresh fuel. But the dual-use potential of Oklo’s technology, the dangerous policy changes it implies, and the ways in which it heightens the risk of diversion throughout the fuel cycle are obvious.

Oklo is aware of the proliferation hazards of its fuel cycle but has forged ahead regardless, resorting to such euphemisms as “recycling” to downplay the risks associated with reprocessing. The company has also asserted that its method of plutonium separation, commonly known as pyroprocessing, is resistant to proliferation, as it does not produce any pure plutonium streams.

Such claims are technically dubious at best. A US national laboratory study of alternative separation methods found that pyroprocessing provided “only a modest improvement in reducing proliferation risk over existing PUREX technologies,” particularly when it came to state actors. By removing highly radioactive fission products from spent fuel, pyroprocessing nullifies the high radiation levels that provide self-protection against diversion of fissile material, even if the plutonium remains unseparated from the fuel. To be fair, Oklo is partnering with Argonne National Laboratory to develop “advanced sensor technologies” integrated with machine learning algorithms to enhance the detection of diversion within its prospective reprocessing facilities. But augmented material accounting alone will not eliminate proliferation risks. Nor would it mitigate the dangers of a shift in US policy on reprocessing.

Breaking the US no-reprocessing norm. The United States renounced commercial reprocessing in the 1970s on account of its proliferation potential and poor economics. The voluntary move has helped Washington compel its allies and partners likewise to forgo the reprocessing option to help limit the spread of technology that allowed countries like India to develop nuclear weapons arsenals. In this context, Oklo and others’ bid to revive commercial reprocessing with the support of the national labs could convince other countries that reprocessing is necessary for the advanced nuclear energy needed to satisfy the power requirements of critical AI technologies.

The risks to regional and international security are substantial. As an example, successive US administrations have for decades kept ally South Korea’s pyroprocessing ambitions in check. To date, Washington’s example and a bilateral agreement between Seoul and its security guarantor have blocked efforts by South Korea’s nuclear lobby to engage in pyroprocessing and match growing back-end fuel cycle activities in China and long-delayed plans in Japan. A decade ago, when Seoul sought to restructure a bilateral nuclear cooperation agreement so it could use pyroprocessing technology similar to what Oklo proposes, Washington compromised only as far as agreeing to joint studies of the technical, economic, and nonproliferation feasibility of the technology.

The Joint Fuel Cycle Study produced a final report in 2021 that did little to resolve the policy controversy surrounding the pyroprocessing question. Regional stability is now in flux amid North Korea and China’s rapid nuclear weapons buildups and questions about the degree to which a potential Trump administration might undermine US alliances. Changing US reprocessing policy at this time would create additional risks.

AI and nuclear energy: evaluating risks. As the modern economy increasingly integrates artificial intelligence technologies, their scale and demand for energy will grow. To avoid exacerbating climate change, it makes sense for AI executives like Altman to look at nuclear energy as a potential power source. They should, however, predicate their investment on an honest, deliberate, and transparent evaluation of risks. That means no more euphemisms to cover up the proliferation dangers of reprocessing, plus an objective assessment of whether the technology’s proclaimed benefits truly justify its liabilities.

Silicon Valley’s innovation-driven investment mentality may reward startups like Oklo as the self-proclaimed pursuers of “what’s next in nuclear.” But technological advancement takes place far more slowly and deliberately in the nuclear domain than in other technologies, given the slowness of industrial supply chains and the need for safety, security, and nonproliferation regulations to keep pace. A nuclear vendor cannot afford to “move fast and break things” like a Silicon Valley IT venture, for a nuclear energy company that fails to address the proliferation implications of its technology risks breaking critical guardrails against the existential threat of nuclear war.

To be sure, there are ways to advance nuclear innovation without increasing the risks of proliferation. Many advanced reactor concepts do not entail breeding plutonium or reprocessing spent fuel. Bill Gates’s TerraPower is developing a sodium-cooled fast reactor just like Oklo but has addressed the hazards of its technology by declaring its reactors “will not require reprocessing and will run on a once-through fuel cycle that limits the risk of weapons proliferation.” Additionally, both Bill Gates and Amazon founder Jeff Bezos have invested in the technology of nuclear fusion.

Nonproliferation awareness does not entail a Luddite approach to nuclear innovation. It simply requires that industry exercise its due diligence to innovate responsibly, and that governments create the incentives for doing so. Washington has supported domestic nuclear innovation to restore US technological leadership against Russian and Chinese competition, both as a national security priority and as a means of setting strong global nonproliferation standards. The United States should prioritize support to industry players that have minimized the proliferation risks of their prospective technology. If Washington is serious about competing with Russia and China on its nonproliferation bona fides, it should put its money where its mouth is when it comes to the next generation of nuclear reactors.