Dan Loeb's Third Point LLC has disclosed investment positions in former Bitcoin mining operations now pivoting hardware capacity toward AI compute infrastructure. The hedge fund, managing approximately $5.7 billion in assets, did not disclose position sizes or specific portfolio companies in the filing.
The pivot capitalizes on stranded GPU infrastructure originally deployed for proof-of-work validation. Multiple Bitcoin miners accumulated tens of thousands of graphics processing units during the 2020-2021 crypto expansion, then faced margin compression as Ethereum moved to proof-of-stake in September 2022 and Bitcoin's hashrate difficulty climbed 47% between Q4 2022 and Q4 2023. Several operators—CoreWeave precursor operations, Hive Blockchain, Hut 8—began repurposing facilities for machine learning inference workloads starting in mid-2023. The shift required minimal capital expenditure: existing cooling infrastructure, power contracts, and rack density already exceeded most colocation standards.
Third Point's entry signals institutional validation for a thesis emerging across distressed asset desks. AI model training demand from OpenAI, Anthropic, and Google DeepMind has pushed H100 cluster lease rates to $2.20-$2.80 per GPU-hour in Q4 2024, compared to $0.80-$1.10 in Q4 2023. Former mining facilities offer 15-20% lower operational costs than purpose-built AI data centers due to pre-negotiated power purchase agreements, many locked at 3.2-4.8 cents per kilowatt-hour through 2026. The arbitrage window exists because hyperscalers cannot build fast enough: Microsoft and Meta have $38 billion in combined data center capital expenditure budgets for 2025, but new facilities require 18-24 months from groundbreaking to production.
Loeb's timing aligns with tightening supply in the inference market. NVIDIA's GB200 shipments face yield issues at TSMC's CoWoS-L packaging line, pushing initial delivery windows from Q1 2025 to Q2 2025 for non-priority customers. That creates a four-to-six-quarter window where converted mining facilities can command premium rates before purpose-built infrastructure comes online at scale. The strategic question is whether these operators can establish sticky customer relationships—long-term inference contracts with mid-tier AI labs or enterprise customers—before the window closes. Early data suggests mixed results: Hive Blockchain reported 68% utilization rates in Q3 2024, while smaller operators struggle to reach 40% due to compliance and latency requirements from Tier 1 customers.
The investment carries concentration risk specific to the transition period. Power grid constraints in Texas and Wyoming—two states with heavy mining infrastructure—may limit scaling. ERCOT has flagged 12 GW of incremental data center demand through 2030 but approved only 6 GW of new generation capacity. Operators face potential curtailment during peak demand periods, which undermines the 99.95% uptime guarantees required for inference workloads. Additionally, the hardware refresh cycle compresses: H100 clusters remain competitive for 24-30 months, compared to 48-60 months for previous-generation mining ASICs. That accelerates depreciation and raises the bar for cash flow sustainability.
Watch for three developments in the next six-to-nine months. First, public filings from Hut 8 and Core Scientific in March 2025 will clarify whether converted facilities achieve 60%+ utilization rates, the threshold where unit economics surpass traditional colocation. Second, NVIDIA's GB200 production ramp in Q2 2025 will signal whether the supply bottleneck persists or breaks. Third, power contract renegotiations in Texas through mid-2025 will test whether operators retain their cost advantage as utilities reprice industrial rates. Third Point's entry suggests at least one sophisticated allocator believes the window stays open long enough to generate alpha, but the thesis depends on execution that most mining operators have never attempted at this level of operational rigor.