|
The ongoing proliferation of generative artificial intelligence (AI) applications in enterprise has set off a global race for hardware assets that are capable of backing the heavy processing and data loads that chips will be subjected to when dealing with large language models (LLMs). The most in-demand product for supporting this type of algorithm has become the graphics processing unit (GPU), but the production of these complex semiconductors is quite limited relative to the rapidly increasing demands of a world that is suddenly in an aggressive AI adoption phase.
AI is largely accessed remotely through the cloud, allowing users to utilize various services and platforms remotely. The providers of these services, in turn, often rely on power-hungry data centers to back up the applications with concentrated loads of computing power. As MRP has noted in several recent intelligence briefings, data centers are already encumbered with unprecedented processing needs that are spurring a boom in new construction of these facilities over the next several years. But those facilities will take time to build while the utilization of AI is spiking now, in real time.
Dedicated data centers are not the only facilities that maintain vast shalves stocked full of GPUs, however, as the world’s increasingly vast number of crypto-mining facilities – a sort of data center themselves – also host these semiconductors. Crypto-mining primarily leverages rigs of GPUs, as well as application-specific integrated circuits (ASICs), in a process known as hashing that allows for “solving” blocks of transactions, which can then be added to a canonical ledger of previous blocks (hence, blockchain). A block is loaded with transactions that are essentially settled by a miner who is then rewarded with newly-minted cryptocurrency and transaction fees associated with that block.
Bitcoin is the primary asset miners are focused on, as it is the largest cryptocurrency by market capitalization at $1.36 trillion. MRP has previously published a trove of Intelligence Briefings that explain and analyze the intricacies of Bitcoin mining in depth, but what is important to understand in the context of this report is that the same machines which undertake this process can also be used as high-powered computing (HPC) in the cloud.
Crypto-miners’ foray into the realm of more traditional data centers processes is not a particularly new development, but agreements between these enterprises and cloud computing providers are becoming more common and formalized. Publicly traded Canadian miner Hut 8 began pivoting some of its operations towards AI and HPC in early 2022. Not only can existing mining capacity be retrofitted to support AI demand, but the operators of these facilities can take their expertise in mining (particularly the acquisition of cheap and abundant energy supplies) and apply it to more traditional data center operations. Two years ago, Hut 8 acquired five data centers to support cloud and colocation services, recognizing that demand for HPC capacity would only be set to increase down the line. In its most recent earnings report, Hut 8 noted that it has recently finalized commercial agreements for a new AI vertical under a GPU-as-a-service model, forecasted to generate revenue at annual rate of approximately $20 million – roughly 18.2% of fiscal 2023 revenue of $110 million (translated from 150.7 million CAD).
While that does not represent a huge share of Hut 8’s revenue according to the prior year’s earnings, it must be remembered that firms in the cryptocurrency industry are subject to volatility in the underlying price of the assets they are producing or managing. For example, while Bitcoin is trading close to $70,000 per unit, it was worth less than half of that amount just a year ago. By creating a baseline level of revenue, which can be generated from a service that is less volatile, miners can de-risk their business model and increase financial stability during large drawdowns in cryptocurrency valuations.
Moreover, the amount of Bitcoin that can be mined per block was just cut in half earlier this year in a process aptly known as the “halving”, wherein the block reward was slashed from 6.25 BTC to 3.125. The latest halving was the fourth iteration of the event, but this programmatic monetary policy will continue to occur once every 210,000 blocks (about four years’ time) until 2140 when the final Bitcoin is mined and the digital currency reaches its supply cap of 21 million. In theory, the reduction in supply should raise the price of Bitcoin significantly over time if demand for the asset continues to grow, offsetting the near-term revenue drop. Obviously, that is not a perfect science, and it can be difficult to project just how long bear markets in crypto can last. In such a scenario, miners should now be able to quickly pivot some of their less efficient processing resources – which would otherwise simply be turned off and generate no revenue at all if the cost of mining outweighed the potential crypto-yield derived from running the machine – to HPC tasks that can still generate revenue.
Last week, major US miner Core Scientific inked a massive $3.5 billion deal with specialized cloud infrastructure provider CoreWeave, the largest private operator of GPUs in North America. This arrangement will see Core Scientific provide CoreWeave with about 200 megawatts (MW) of infrastructure to power HPC services over the next 12 years; a big step up from a previous agreement between the two companies that extended 16MW of infrastructure to Coreweave. Core Scientific claims to have at least 300 MW of additional HPC capacity available, which suggests non-mining processes could to take on a more than 41% share of the 1.2 gigawatts (GW) of operational capacity and contracted power at its disposal. The miner also owns a further 372MW of partially built infrastructure at its two Texas data centers. CoreWeave was so enthused about its Core Scientific partnership that it launched an unsolicited $1 billion bid for the whole company. This was shot down quickly by Core Scientific’s board, but it is unlikely to be the last M&A effort targeting Bitcoin miners.
Per JPMorgan, miners are attractive targets not only for the GPUs and power resources they have access to alone, which is now equivalent to 5GW in use and an additional 2.5GW of availability, but due to their attractive power contracts with utilities. Because mining can take place virtually anywhere, miners tend to flow toward power markets with cheap and abundant energy resources that allow miners to buy power in advance, as well as capitalize on periods of excess generation. Usually, the latter category might go unused and generate losses for the local utilities, constraining investment in the grid, but miners can essentially function as “buyers of last resort” for electricity. Data centers are likely to focus more on power costs in the future as electricity bills will be rising alongside the power needs of GPUs and other advanced AI chips. Axios notes that data center operators will need to spread out more than they’ve been accustomed to for this reason. It could be that large cloud computing providers could simply buy up Bitcoin mining firms and use real estate that they are already in possession of to convert existing campuses into data centers. This could be difficult, however, as data centers are usually larger than mining facilities.
Investors can gain exposure to the crypto mining industry via the Valkyrie Bitcoin Miners ETF (WGMI).
|
Leave a Reply