24-year-old fund manager: annual returns are 24 times higher! His AI investment portfolio targets the “most scarce resource”

Former OpenAI researcher Leopold Aschenbrenner, through his fund, doubled $225 million to $5.5 billion. He pointed out that the bottleneck in AI development lies in electricity, not chips or models.

Former OpenAI safety researcher, 24-year-old German Leopold Aschenbrenner, took less than a year to turn his managed fund, Situational Awareness, from $225 million into $5.5 billion. While Wall Street funds flooded into AI models and chip companies, he identified a overlooked blind spot: power. By precisely betting on infrastructure solutions to address AI’s electricity consumption issues, he achieved astonishing investment returns.

Leaving OpenAI to enter the AI investment market

After leaving OpenAI, Aschenbrenner wrote a 165-page report asserting that artificial general intelligence (AGI) will arrive sooner than everyone thinks, and that the ultimate winners won’t be the companies with the strongest AI models, but those that “control electricity.” To this end, he founded the hedge fund “Situational Awareness LP” and invested $875 million in fuel cell company Bloom Energy.

This week, Bloom Energy announced a $2.8 billion fuel cell order with Oracle, causing its stock to surge 15% after hours. The paper value of Aschenbrenner’s holdings also skyrocketed to nearly $2 billion.

Portfolio reveal: Long infrastructure, short traditional IT

The news indicates his investments follow the “power first” logic:

• Bloom Energy (BE): Invested $875 million in this fuel cell company. Its technology allows data centers to generate power on-site, eliminating reliance on aging power grids. Benefiting from the 2.8 GW order with Oracle, its stock value has surged, reaching nearly $2 billion.
• CoreWeave (CRWV): Invested $700 million in this top AI cloud computing provider, focusing on its scarce infrastructure resources.
• Infosys (INFY): Heavily shorted this Indian IT outsourcing giant, expecting AI coding agents to thoroughly destroy traditional IT outsourcing businesses.
• Intel (INTC): Used Intel call options for leverage, earning multiple times during the stock’s strong 53% rebound.
• Core Scientific (CORZ): Holds a 10% stake. This former Bitcoin mining company is transforming its existing power facilities into AI data center hosting sites.

The power-hungry monster behind computing: electricity consumption doubles annually

Aschenbrenner states that, looking back at 2022, training GPT-4’s compute cluster consumed about 10 MW of power, costing around $500 million. However, AI compute demand is expanding at roughly half a magnitude per year, meaning the largest training clusters’ power needs double every 12 to 18 months.

By 2024, the largest compute clusters will consume up to 100 MW—equivalent to running 100k high-end GPUs simultaneously. By 2026, leading training clusters will require up to 1 GW of continuous power, comparable to a large nuclear reactor’s output.

Who controls electricity, controls AI’s future

He estimates that by 2028, AI training will consume 10 GW of power—more than many U.S. states’ entire electricity generation; by 2030, it will reach 100 GW, consuming about 20% of the current total U.S. electricity production. This only accounts for “training” models; when including the inference power used for public deployment, the electricity consumption becomes even harder to imagine.

However, U.S. total power generation has only grown marginally by 5% over the past decade. Reports of transformer shortages and stalled data center construction prove that the grid cannot support such growth. This is why he dares to heavily back Bloom Energy: the real bottleneck in AI development isn’t chips or software, but whether humans can produce enough electricity.

• This article is reprinted with permission from 《Chain News》
• Original title: “24-Year-Old Fund Manager’s Annual Return of 24 Times! AI Portfolio Targets the ‘Scarcest Resource’”
• Original author: Co2