They Built AI. Now They're Building Bunkers.

Three recent and distinct news reports may be more related than they appear. We learned last week from NBC News that the Pentagon is devising a startling new nuclear strategy in case of a regional war with China or Russia. While not yet adopted as policy, the plan would place far greater reliance on short-range tactical nuclear weapons than ever envisioned. Also last week, the Financial Times published a surreal report on mega-wealthy tech bros preparing to survive an apocalypse while safely ensconced in remote enclaves. This comes on the heels of continuing confirmation of earlier findings that in war game simulations AI chose a nuclear option about 95% of the time. I’m not a big believer in coincidences.

The revised nuclear planning, led by Elbridge Colby, deputy secretary of defense for policy – who was key to ending US aid to Ukraine, dramatically pivots from established protocol of using long-range missiles to dismantle an adversary’s nuclear capacity. Colby figures that our enemies already know the US is prepared to use tactical nukes, so why not bake it into formal policy. NBC News sources offered a seemingly counterintuitive explanation asserting that use of tactical nukes could prevent a conventional war from expanding into a widespread, catastrophic nuclear exchange. Maybe, but perhaps it’s also because we’ve foolishly depleted our stock of conventional munitions in the war with Iran.

Seasoned military experts are alarmed by the naivete and short-sightedness of this reasoning. Those voices say that reliance on smaller tactical nukes could diminish an adversary’s wariness of our big stick nuclear threat and make it more likely that nuclear exchanges occur. Those critics argue that a greater role for regional size nuclear weapons could drop the bar for nuclear deployment and enhance the risk of a limited conflict spinning into a global catastrophe.

We certainly don’t need to increase the odds of nuclear warfare – artificial intelligence has already done that for us. Battlefield AI is no longer a speculative sci-fi scenario. It’s reared its head in Iran, Ukraine, and even in the U.S. incursion into Venezuela. In Iran, an AI targeting flub is believed to be the cause of 175 people, mostly children, dying in a school. That fatal error pales in comparison to the grave glimpse of carnage revealed in recent AI war game exercises. In research at King’s College London, AI models used for a simulated battle expanded conflicts by threatening nuclear weapon strikes in 95% of scenarios.

In the study, three of the top AI products, GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash were deployed in 21 nuclear crisis scenarios. In each of the games, AI signaled nuclear intentions or “signaling” for at least one side of the battle, and mutual nuclear signaling in 95% of the battles. “None of the AI models ever chose accommodation or surrender.” The leader of the study noted that “Claude and Gemini especially treated nuclear weapons as legitimate strategic options, not moral thresholds.”

Those are some of the same brand AI models that Secretary of Defense Pete Hegseth is actively incorporating into the Pentagon’s AI decision-making initiative. The U.S. military is aggressively integrating artificial intelligence into battlefield planning, command and control, and autonomous weapon systems. Driven by what it calls an “AI-first” and “wartime footing” strategy, the Department of Defense is “shifting away from slow bureaucratic procurement to rapid Silicon Valley tech adoption, leveraging generative AI for mission planning, building massive data infrastructure, and deploying real-time operational intelligence platforms.” Hegseth insists humans will stay at the center of final strike decisions, but will AI let them?

While the Pentagon shifts away from slow traditional methods toward rapid Silicon Valley technology, some of the humans responsible for that innovation, who know best what AI is capable of, are racing away to remote outposts and bunkers. An eye-opening Financial Times piece took us into the mindset and plans of some of the wealthiest technology powerbrokers associated with the Pentagon’s initiatives. The key takeaway – these guys don’t plan to stick around when their technology hits the proverbial fan. The idea of rich people spending their money to secure their existence during a major crisis isn’t new. For example, for several years now you could plunk down $3 million for a luxury unit in an underground condo complex in Kansas, replete with top end appliances, a pool, and giant murals that let you pretend you’re in a sunny above-ground environment. But the tech broligarchs aren’t in Kansas anymore.

For some, the foothills of the Andes in Argentina are the place to hunker down when Armageddon arrives. The FT article featured tech founder Martin Varsavsky, who started at least ten companies including AI, biotechnology, autonomous vehicles, renewable energy and telecommunications firms. He’s currently CEO of Certuma, an autonomous clinical AI company. His net worth is around $700 million. Varsavsky’s a big fan of the Defense Department’s plan to put AI data centers on military bases. Over the past two years, small prefab homes, an airstrip, and geodesic domes have emerged on remote ranch not far from Patagonia on desolate land larger than the size of Manhattan. Varsavsky explained that “the whole of Argentina and the Southern Cone region” could support “complex societies” in the event of nuclear war, due to their distance from missile targets and radiation clouds, and abundant food and energy production. “My idea isn’t just to have little groups locked up in bunkers.”