암호화폐 채굴자들, 추론 수요 노려 AI 제공자로 전환
Bitcoin miners have spent years securing prime real estate: land, electrical substations, and grid connections with reliable, always-on power—the essential foundation for facilities consuming electricity around the clock. Now, as cryptocurrency prices remain volatile and mining rewards diminish with periodic Bitcoin halvings, that same infrastructure is proving far more valuable for artificial intelligence.
CoreWeave exemplifies the pivot. The company began in 2017 as Atlantic Crypto, using graphics processors to mine Ethereum. Its prospectus reveals that before 2022, most revenue came from cryptomining operations it has since abandoned. Today, CoreWeave operates a purpose-built AI cloud for model training and inference—and has become central to the generative AI revolution. Other former miners—IREN, HIVE Digital Technologies, Hut 8, and Bit Digital—have followed a similar trajectory into AI computing or data center leasing. Core Scientific went further, agreeing to supply AMD with 530 megawatts of AI-ready capacity from 2027, a deal that could eventually expand to 2.5 gigawatts.
The transition reflects a critical market shift. AI companies face severe shortages of powered data center space, while crypto miners possess something increasingly scarce: reliable access to vast amounts of electricity and strategically located infrastructure. As one HIVE executive put it: "AI is the new industrial base and compute is the factory floor."
Converting these facilities demands substantial investment. The application-specific integrated circuits that mine Bitcoin are poorly suited to AI workloads. Inference and training require expensive GPUs, fast networking, reliable fiber connections, redundant power systems, and far more sophisticated cooling infrastructure. IREN has installed direct-to-chip liquid cooling at its Childress site in Texas and is replacing some Bitcoin mining equipment with GPUs to serve customers running Nvidia processors for training and inference workloads.
Inference—delivering responses from existing models rather than training new ones from scratch—represents the most immediate opportunity. Unlike model training, which is intensive but intermittent, inference runs continuously as users interact with AI applications, requiring the low-latency, high-reliability characteristics that existing mining infrastructure can support. CoreWeave's prospectus projects that inference infrastructure alone could represent a $49 billion market by 2028. Bit Digital's cloud-services revenue rose 84% year on year in early 2025 even as mining revenue fell 64%; its AI operation has since spun into the separately listed WhiteFiber. HIVE is converting a former mining facility in Sweden to AI standards, recently signing a three-year GPU contract worth approximately $220 million and planning a Toronto campus designed specifically for low-latency inference.
The transformation carries risks. Many mining operations lie near cheap electricity but remote from customers, fiber routes, and skilled technical workers. Yet as AI companies scramble for capacity, power has become an asset worth more to data center operators than the cryptocurrencies the facilities were originally built to produce.