On September 8, 2026, OpenAI announced it had solved the Navier-Stokes problem, one of seven Millennium Prize Problems worth $1 million each, using a swarm of 10,000 coordinating AI agents. The agents completed their work on September 5, about 88 hours after being launched on September 1. The 90-year-old mathematical puzzle about how fluids move has stumped the world's brightest minds since the 1930s. OpenAI's system exchanged 2.7 million messages and produced 130 billion output tokens to crack it. Earlier in May, another OpenAI model solved an 80-year-old geometry problem called the unit distance conjecture, posed by legendary mathematician Paul Erdos in 1946. These aren't incremental advances. They're existential shifts in what machines can do without human guidance. But there's a darker implication lurking beneath the celebration. If AI can independently solve problems that have defeated human mathematicians for nearly a century, what else can it figure out with enough computing power thrown at it? The knowledge embedded in an unfiltered AI model trained on humanity's accumulated scientific literature is staggering. Physics, chemistry, nuclear engineering, weapons design, enrichment techniques, all of it absorbed and ready to be queried by anyone with access. A 2026 industry safety report found that several frontier AI labs recently added restrictions to their systems because they could not rule out that their models might assist novices in developing chemical or biological weapons. The Manhattan Project required some of history's greatest minds working for years. Today's aspiring weapons developer doesn't need genius. They need a powerful enough AI model, the right prompts, and enough hatred in their soul to use what it tells them. The timing couldn't be more volatile. On February 28, 2026, the United States and Israel launched coordinated strikes on Iran, assassinating Supreme Leader Ali Khamenei and targeting the country's nuclear facilities at Fordow, Natanz, and Isfahan. President Trump stated the strikes aimed to prevent Iran from ever acquiring a nuclear weapon. Iran's nuclear program had been advancing for years after Trump withdrew from the Joint Comprehensive Plan of Action (JCPOA) during his first term. By June 2025, Iran had accumulated uranium enriched to 60 percent, a short technical step from weapons-grade material at 90 percent. The International Atomic Energy Agency (IAEA) withdrew inspectors after the strikes, leaving the status of Iran's program unclear. Some 440 kilograms of 60 percent enriched uranium remain unaccounted for. Here's the paradox Trump won't acknowledge: bombing Iran to prevent nuclear proliferation while AI systems capable of assisting weapons development proliferate globally makes no strategic sense. Iran announced plans in May 2026 to build its first GPU-based data center to host a national AI operating system. The country's deputy of science, technology, and knowledge-based economy, Hossein Afshin, said the facility would be operational by the Iranian calendar year 1404 (2025-2026) and that Iran is establishing local chip production through the Sahand National Project. Even under heavy sanctions that restrict access to advanced chips like Nvidia's A100 and H100, Iran has a long history of smuggling IT equipment. If data centers and compute power are the new nuclear enrichment facilities, then proliferation is already happening at digital speed. The 2026 Iran war has turned data centers into legitimate military targets. Iranian forces struck Amazon Web Services facilities in the United Arab Emirates and Bahrain, Oracle's Dubai data center, and other tech infrastructure across the Gulf. These attacks caused widespread disruption to banking systems, payments, and enterprise services. The United States retaliated by striking a data center in Tehran operated by Iran's state-run Bank Sepah on March 11. Iran's Islamic Revolutionary Guard Corps (IRGC) published a list of 29 tech targets and specifically threatened Stargate, a $500 billion AI data center initiative backed by OpenAI, SoftBank, and Oracle. Data centers, which were once behind-the-scenes infrastructure, are now front-line assets in modern warfare because they power AI systems that militaries use for intelligence analysis, operational support, and targeting. The uncomfortable truth is that making a nuclear bomb still requires access to fissile material like highly enriched uranium or plutonium. You can't build one without it. No amount of compute power changes that physical reality. But what AI does is eliminate every other bottleneck. It can optimize enrichment processes to extract more weapons-grade material faster. It can troubleshoot engineering challenges that stumped previous generations of weapons designers. It can simulate dozens of weapon designs and identify which will work without the need for physical testing. It can compress what took the Manhattan Project four years into months or weeks for anyone with the right materials and enough computing power. Research warns that AI may eventually help potential proliferators overcome the remaining bottlenecks to building the bomb. The speculative fervor around AI offers perfect cover for countries to fast-track nuclear programs under the guise of competing in the AI revolution, building up enrichment and reprocessing infrastructure necessary for a usable weapon. We are watching two technological arms races collide. One is the race to build more powerful AI systems capable of solving problems humans cannot. The other is the race to prevent those same systems from falling into the wrong hands or being used to design weapons of mass destruction. The Biden administration's UN General Assembly adopted a resolution in December 2025 on risks from AI in nuclear command, control, and communications. At the April 2026 Asilomar Conference, over 100 experts developed seven principles for governing AI applications in nuclear and biological security. But these governance efforts are moving at bureaucratic speed while AI capabilities advance exponentially. Private developers may be the first to identify dangerous model capabilities, but they cannot evaluate those threats on their own. Governments are always playing catch-up.