Leopold Aschenbrenner, the former OpenAI safety researcher turned AI infrastructure entrepreneur, is making a bold repositioning bet on bitcoin miners as the backbone of next-generation AI compute. His $13.6 billion venture, Redwood Research, is shorting semiconductor giants Nvidia and AMD while aggressively accumulating stakes in bitcoin mining operations that control critical electricity and data center assets.
The strategy reflects a contrarian thesis. Aschenbrenner argues that as AI models demand exponentially more compute power, ownership of physical infrastructure beats ownership of chips. Bitcoin miners have already spent years building out redundant power grids, cooling systems, and real estate. That sunk capital becomes a moat when scaled for AI workloads.
Nvidia's stock has priced in massive AI expansion. AMD trails further behind. Both depend on global supply chains vulnerable to geopolitical pressure and cyclical chip cycles. Bitcoin miners, by contrast, have already amortized their infrastructure costs and operate on wafer-thin margins that improve as their assets get repurposed for AI.
The bet hinges on infrastructure scarcity. Energy capacity and low-latency data center proximity matter more than raw chip availability as AI labs compete for compute resources. Aschenbrenner's position essentially bets that the holders of electrons and steel beat the sellers of silicon.
This move carries execution risk. Bitcoin miners face regulatory headwinds in key markets, energy price volatility, and unproven capability to pivot toward AI workloads at scale. His short positions on Nvidia and AMD also expose Redwood to significant drawdowns if semiconductor demand continues its bull run.
Still, the thesis taps into a real infrastructure transition underway. Major AI labs already negotiate directly with power providers. Data center operators command premium valuations. Aschenbrenner's $13.6 billion wager signals that serious capital sees the next phase of AI scaling as fundamentally infrastructure-constrained rather than chip-constrained. Whether bitcoin miners can capture that value depends on execution and regulatory clarity.