High Performance Computing (HPC) infrastructure for artificial intelligence (AI) presents a promising opportunity for Bitcoin miners. However, questions remain about the sustainability and available capacity for such a transition, according to a recent report by Needham & Company.
Converting existing Bitcoin mining sites to HPC infrastructure requires significant capital expenditure. While some components can be repurposed, most of the infrastructure needs to be rebuilt to support HPC operations. Currently, HPC data centers demand an investment of $8-10 million per megawatt (MW) in capital expenditure (capex), excluding GPUs. In contrast, Bitcoin mining sites typically cost between $300,000 and $800,000 per MW, excluding ASICs.
Most Bitcoin miners use ASICs (Application-Specific Integrated Circuits) for Bitcoin mining rather than GPUs (Graphics Processing Units), which were commonly used for Ethereum mining. This means they lack the excess GPU capacity necessary for HPC tasks.
To generate AI revenue more quickly and with less capital expenditure, Bitcoin miners venturing into AI are predominantly acquiring GPUs, particularly H100s, and co-locating them in third-party sites. However, Needham suggests that it is more advantageous for miners to own their data centers, especially if they have access to inexpensive power, rather than relying on third-party sites for GPU co-location.
Several large, publicly traded Bitcoin miners plan to more than double their power capacity over the next 12-24 months, encompassing both their mining and AI/HPC expansion plans. Needham highlights the importance of location, favoring northern, less dusty climates over regions like West Texas due to temperature and dust/wind conditions.
Applied Digital is noted for having the most ambitious AI/HPC site plans among the covered companies. Originally a Bitcoin miner, Applied Digital rebranded after selling a ~200 MW Bitcoin mining hosting site to Marathon Digital (NASDAQ).
Bit Digital has become the first cloud service platform in Asia to offer NVIDIA DGX SuperPod H100s. “We do not include any revenue from AI/HPC in our model but will be monitoring for updates as 2024 progresses,” Needham stated.
HUT 8 Mining has launched a third-party HPC service, despite not currently generating revenue from it. The company has ordered 1,000 Nvidia (NASDAQ) H100s and secured a $20 million annual run-rate contract with a venture capital-backed AI large language model customer, with revenues expected to begin in the second half of 2024.
Bit Brother has secured a three-year contract to provide GPU resources to a customer developing its own proprietary large-language model (LLM), expecting $50 million in annual revenue following a 2023 funding round. Additionally, Core Scientific’s multi-year contract with CoreWeave is projected to generate more than $100 million in total potential revenue.
HIVE Blockchain had the majority of its H100s delivered in March 2024 and reports growing positive HPC operating income, with $25 million in run-rate annual HPC revenue guidance as of Q1 2024.
Related Topics:
