The most contrarian trade in AI right now isn't a bet on a new model architecture or a dark-horse startup. It's a bet against the companies everyone assumes will win.
Leopold Aschenbrenner, the former OpenAI researcher who left the company amid controversy over his aggressive timelines for artificial general intelligence, is putting his money where his manifestos are. His fund—now swollen to $13.6 billion—is shorting Nvidia and AMD while going long on bitcoin mining companies. The thesis is elegantly simple and potentially devastating to conventional wisdom: the constraint on AI progress is shifting from compute to power, and the entities best positioned to deliver power at scale aren't chipmakers. They're the firms that already own the electricity contracts and purpose-built data centers that crypto mining demanded.
The infrastructure inversion
For the past three years, the AI investment playbook has been almost comically straightforward: buy Nvidia, buy anything Nvidia touches, buy the companies buying from Nvidia. The logic was sound—training frontier models requires staggering amounts of GPU compute, and Nvidia held something close to a monopoly on the hardware that mattered.
But Aschenbrenner's bet reflects a growing belief among AI infrastructure watchers that we're approaching an inflection point. As models scale and inference demand explodes, the binding constraint is increasingly the mundane stuff: megawatts, cooling capacity, grid interconnection agreements. Bitcoin miners spent a decade solving exactly these problems, often in remote locations with cheap hydroelectric or stranded natural gas. They built relationships with utilities, secured long-term power purchase agreements, and constructed facilities designed for nothing but converting electricity into computation.
Why the timing matters now
The trade isn't purely theoretical. Several bitcoin mining companies have already begun pivoting toward AI hosting, and the economics are compelling. A facility optimized for SHA-256 hashing isn't identical to one optimized for transformer inference, but the core competencies—power procurement, thermal management, site selection—transfer remarkably well. Meanwhile, hyperscalers are struggling to build new data center capacity fast enough to meet demand, creating an opening for anyone with existing infrastructure and available power.
Aschenbrenner's fund appears to be betting that this transition will accelerate faster than the market expects, and that the resulting demand destruction for mining-specific ASICs will be more than offset by the premium AI workloads command.
Our take
This is either a genuinely prescient infrastructure play or an elaborate cope from someone who needs AI progress to accelerate on his predicted timeline to justify his entire public thesis. Probably both. What's undeniable is that Aschenbrenner has identified a real tension in the AI buildout: the chip supply chain has received billions in investment and attention, while the power and real estate layers remain surprisingly fragmented. If he's right that electricity becomes the new chokepoint, the bitcoin miners sitting on gigawatts of capacity could find themselves accidentally positioned at the center of the most important technology transition of the decade. Stranger things have happened in markets—though not many.
