OpenAI calls for an IAEA for superintelligence

Imagine an AI system that produces more economic output than a Fortune 500 company and outperforms human experts at nearly everything. Now imagine trying to regulate it after it exists.

OpenAI just published a framework for governing what it calls superintelligence—AI systems far more capable than even AGI—which the lab believes could arrive within the next decade. Instead of waiting for things to spiral, OpenAI wants world leaders and labs to start laying the regulatory tracks right now.

An IAEA for supercomputers

OpenAI’s headline proposal is the creation of an international oversight body modeled directly after the International Atomic Energy Agency (IAEA).

Under this setup, any AI project exceeding a certain capability or compute threshold would face mandatory safety audits, compliance inspections, and strict deployment limits. How would anyone actually enforce this? OpenAI suggests tracking compute clusters and energy consumption, which are much easier to monitor than raw code.

Beyond a global agency, the post suggests leading frontier developers could collectively agree to cap the annual growth rate of AI capabilities, or team up on government-backed joint initiatives.

Crucially, OpenAI drew a sharp line between frontier mega-models and everything else. The lab explicitly argued that open-source projects and models below the capability threshold should be exempt from heavy burdens like licensing and audits. Today's systems, they noted, carry risks roughly on par with the regular internet.

Why not just stop?

If the existential threat is real, why build superintelligence at all?

OpenAI’s answer is simple: stopping it is virtually impossible and counterproductively dangerous. With compute costs plummeting every year and the number of global actors rising, enforcing a blanket ban would require a draconian global surveillance regime that probably wouldn't even work anyway.

Since the train is leaving the station, OpenAI argues the only sensible move is steering it safely, while letting the public democratically set the guardrails for how these models behave by default.

Why it matters

We think this proposal is a textbook example of frontier labs attempting to write their own regulatory playbook before politicians do it for them.

Tying oversight to compute and power thresholds is clever. Unlike software weights, advanced GPUs and massive data centers leave a massive physical footprint. If you want to police AI without micromanaging code, auditing electricity and silicon is the only realistic lever.

OpenAI’s explicit defense of open source is also a strategic relief valve. Critics have spent months accusing top labs of using existential dread to erect regulatory moats that squash open-source competition. By advocating a high threshold that shields smaller models from licenses, OpenAI is trying to neutralize that backlash.

That said, getting rival superpowers to agree on capability caps or submit their sovereign data centers to international inspections feels wildly optimistic in today's geopolitical climate. For builders right now, the takeaway is clear: the heavyweight regulatory sledgehammer is aimed at the trillion-parameter frontier, leaving the day-to-day open-source ecosystem room to breathe.