Former OpenAI researcher bets billions on AI's true constraint: power
Original: This is WILD!
Deep summary
Who: Posted by @MilkRoadAI (crypto/finance newsletter turned AI investing content), quoting @StockMKTNewz (financial news aggregator). The underlying subject is Leopold Aschenbrenner, a former OpenAI safety researcher who published the Situational Awareness document series in 2024.
What's new: Aschenbrenner's Q1 2026 13F filing for his Situational Awareness Fund reveals a concentrated infrastructure and power thesis translated directly into public equity positions. The largest single holding is Bloom Energy at $878.7 million, followed by SanDisk at $724.4 million, CoreWeave at $556.1 million, and a cascade of power-adjacent compute plays. The portfolio is almost entirely devoid of model companies or application-layer software.
How it works: The thesis, laid out in the 165-page Situational Awareness document, argues that the binding constraint on development is not algorithmic progress or chip supply but grid-scale electrical power. AI compute demand has been scaling at roughly half an order of magnitude per year, which means training cluster power requirements compound dramatically: from roughly 10 megawatts for the GPT-4 training run in 2022, to 100 megawatts by 2024, to an estimated 1 gigawatt today, to a projected 10 gigawatts by 2028 and 100 gigawatts by 2030. US grid capacity has grown only about 5% over the past decade, making behind-the-meter generation a structural necessity rather than an option. Bloom Energy sells solid-oxide that can be deployed at data center sites without waiting years for grid interconnection approval, which is the specific bottleneck they sidestep.
The numbers: Bloom Energy reported Q1 2026 revenue up 130% year over year and announced an Oracle deal to deploy 2.8 gigawatts of fuel cell capacity across AI data centers. The stock is up approximately 1,400% over the past year. Aschenbrenner reportedly began accumulating Bloom Energy shares in the mid-teens price range during 2025, built to an $875 million position, which appreciated to roughly $2.73 billion by end of Q1 2026. Other major positions include CoreWeave at $556 million, Iren Limited at $401 million, Core Scientific at $389 million, and Applied Digital at $320 million, all of which are either power-intensive compute hosts or Bitcoin miners with convertible data center infrastructure.
Caveats: This post is promotional content from a newsletter explicitly pitching its own paid tier, so the framing is selective. The 100-gigawatt-by-2030 projection implies a single training cluster consuming more than 20% of current total US generation, a number that has no historical precedent and depends on sustained half-order-of-magnitude annual scaling continuing uninterrupted, which has never held indefinitely in any prior compute era. The SanDisk position at $724 million is conspicuous and does not fit the stated power thesis cleanly, suggesting either a storage-for-checkpointing angle or a separate thesis entirely. Bitcoin miner positions like Riot, CleanSpark, and Bitfarms make structural sense as stranded-power and data center conversion plays, but the thesis conflates several distinct infrastructure bets under a single narrative umbrella.