Author: Ariel Deschapell, Drew Armstrong, Bitcoin Magazine; Translated by: Baishui, Golden Finance
The story of human progress can be simplified to the story of increasing energy efficiency. We use energy to create order, both biologically and socially. Excess energy can create wealth in various forms, which in turn leads to new technologies to efficiently use more energy. This fact has inspired famous concepts such as the Kardashev Scale, which measures civilizations by their ability to use their energy resources for useful purposes.
Computing is a natural extension of this endeavor. Modern digital technologies convert increasing amounts of electricity into advanced value-creating processes. The recent surge in demand for computing comes primarily from two technologies: Bitcoin mining and, more recently, high-performance computing (“HPC”), specifically graphics processing units (“GPUs”) for artificial intelligence. The dramatic rise in energy consumption resulting from these technologies raises many questions: What impact will these power-hungry technologies have on our energy system? What kind of interactions will occur between them given that they consume large amounts of energy from each other? What do these developments mean for humanity?
We explore the fundamental characteristics of each of these technologies and how they can provide alternative markets for excess electricity, thereby actually improving the efficiency of energy systems. Based on this exploration, we also argue that Bitcoin mining and HPC are complementary rather than competitive. As we will see, their respective trade-offs provide a symbiotic ability to maximize the value created by energy resources, thereby benefiting society as a whole.
In short, we advocate maximizing computation.
Energy efficient
Modern technology relies on converting a variety of energy sources into electricity, which comes with several challenges and trade-offs. Chief among them is limited portability.
This is due to a few simple realities. Electricity requires an electrical grid, essentially a series of giant circuits that transfer energy in real time. The grid must be balanced, meaning that the amount of electricity generated must roughly equal the amount of demand at any point in time.
This is difficult for two reasons:
First, energy resources are not always conveniently distributed, development cycles are long, and dispatch capabilities vary.
Second, both transmission and storage are expensive, with similarly long delivery times and inefficiencies. It is estimated that 8-15% of electricity is lost in transmission and distribution by the time it reaches local consumers, and even higher losses in long-term battery storage.
The upshot is that it is always cheaper and more efficient to use electricity immediately at the source than to transport it across time or space. The most effective solution, therefore, is not to transport electricity more widely and less efficiently to where it can be used, but to shift the use case to where it can be used. Computing is an ideal use case for this excess electricity because it is power-dense, highly portable, and scalable; we have yet to discover limits to demand for computing. Meanwhile, “physical space” constraints are a strong limiting factor for traditional forms of energy consumption, such as aluminum smelting and manufacturing.
Bitcoin mining has become an ideal use case for local surplus electricity, providing dispatchable and revenue-generating load to balance the grid. More recently, the demand for high-performance computing (especially GPUs) has also had a non-negligible impact on energy utilization. Many expected the two technologies to compete for the same energy resources, but as we explore the characteristics of each technology, the potential symbiotic relationship will become self-evident.
Bitcoin Mining
Bitcoin mining can be thought of as a permissionless consumption of energy. Bitcoin's proof-of-work consensus mechanism amounts to a proof of energy-intensive computation. Miners must perform this energy-intensive computation to create new blocks of transactions, thereby receiving Bitcoin as a reward. It is this proof-of-work that provides global settlement guarantees in a decentralized and permissionless manner.
In practice, this looks like millions of computers (now called application-specific integrated circuits, or "ASICs") running in basic data centers around the world. One of the great things about Bitcoin mining is its permissionless nature; anyone, anywhere in the world can plug in an ASIC. In effect, Bitcoin allows miners around the world to participate in the global energy market; whoever has the lowest electricity costs will have the highest profits.
This global decentralized network is one of the reasons why Bitcoin continues to grow steadily in popularity as people seek a new monetary and financial system that operates 24/7, has no single point of failure, and avoids the perverse incentives of politically controlled central bank monopolies.
Compared with GPU/HPC infrastructure, Bitcoin mining has the following characteristics:
No customers
No customer acquisition
No support
High interruption rate
Low operational complexity
Low connection requirements (less than 100MB/s)
Low profit (usually)
HPC
Data center GPUs are the newest form of HPC, and demand for them has exploded over the past two years due to rapidly increasing interest in the AI/ML breakthroughs that rely on them. These technologies unlock entirely new classes of digital operations and capabilities that were previously impossible, and the resulting use cases are only beginning to be explored. The sudden surge in interest in these technologies has quickly made NVIDIA, the leading manufacturer of the underlying GPUs, the most valuable company in the world.
Initially, this sudden surge in demand created a severe bottleneck in the underproduction of GPUs themselves. However, this was only temporary, and as production increased over time, the problem was alleviated, and the focus quickly shifted to a new bottleneck: data center rack space and cheap electricity. As a result, construction of new data centers exploded wherever there was a large and stable supply of electricity. This put GPU infrastructure in competition with Bitcoin mining in many areas with excess electricity.
Compared with Bitcoin mining, GPU/HPC has the following characteristics:
client
Customer Acquisition
Customer Support
Low interruption rate
High operational complexity
High connection requirements (10 - 100GB)
High profit (usually)
Reduce operating costs by using excess electricity
Over the past decade, the demand for Bitcoin and AI/ML technologies has continued to grow, proving their usefulness to society. This demand has led to a surge in their respective computing resources.
To reduce operating costs, both markets are looking to use excess power, which is often cheaper. This naturally addresses some of the grid inefficiencies discussed above, but it also means that data center builders and operators will find themselves asking which form of computing should be supported and invested in given the same amount of available power.
Both forms of computing are energy intensive and relatively location agnostic (barring legal or jurisdictional considerations which are beyond the scope of this article), which makes them appear to be competing, but in reality they can be highly complementary tools to maximize and profitably exploit this excess or stranded power.
GPU workloads have higher operational complexity and lower disruption tolerance, as well as higher upfront capital investment. This makes it a poor choice for exploiting brief periods of excess power, such as peak windows of electricity generated by solar panels. Unlike Bitcoin mining, customers of GPUs are generally sensitive to issues such as uptime and availability. There are exceptions, such as spot instances and frameworks that can failover from such instances, but in general, the interruption tolerance of GPU infrastructure will never match that of Bitcoin mining due to the presence of customers. Coupled with higher capital costs and complexity, under these circumstances, we can expect Bitcoin mining to continue to grow and become a highly flexible, dispatchable load on the grid.
On the other hand, sustained power surpluses, such as a largely fixed difference between base generation from a hydro or nuclear power plant and its peripheral consumption, are ideal opportunities for GPU infrastructure to close the gap and establish a new baseline consumption and balance. These situations favor low-disruption GPU infrastructure and justify increased expenditures and operational complexity to ensure significantly higher revenues. As long as supporting bandwidth is available to facilitate GPU workloads (at least 10GB/s, ideally 100GB/s), these sites will always offer more profitable opportunities than dedicated Bitcoin mining.
Hybrid Data Center Strategy
There are also strategies that leverage both technologies to maximize revenue and return on investment.
First, Bitcoin mining can serve as an initial load on energy resources before the site is suitable for HPC. Examples include: (1) using a semi-portable modular Bitcoin mining data center to monetize the power while building out the rest of the HPC data center infrastructure (redundant power/internet lines, buildings, backup energy systems, etc.); or (2) using Bitcoin mining to tap into idle energy resources, some of which may eventually be used for HPC. In fact, Core Scientific’s recently announced deal with CoreWeave can be seen as an example of this happening in the wild, as Bitcoin mining led to the development of large-scale substations and data center enclosures that will eventually be used for HPC.
The second, more advanced strategy combines HPC and Bitcoin mining workloads together, using Bitcoin mining as a counterweight to the HPC workload’s power consumption fluctuations. While HPC workloads require reliable power, “inference workloads” that host production AI/ML models can fluctuate based on real-time user usage levels, resulting in typical periods of high activity and power consumption, and low activity and low power consumption. To date, the value of this HPC has greatly outweighed any inefficiencies introduced by fluctuating power usage, but the highly flexible and interruptible nature of Bitcoin mining can be used to provide stable power consumption, thereby reducing the effective price of electricity, in addition to providing additional revenue for the data center as a whole. Some have described this strategy as a “mullet data center,” with AI in the front and Bitcoin in the back. While it’s still early, this approach holds promise for leveraging the best of both HPC and Bitcoin mining to deliver the highest value data center deployments possible with current technology.
Industry Impact
Until recently, the data center industry has been dominated by colocation providers. These providers build facilities for hosting industrial servers and rent out space, power, connectivity, and sometimes even the servers themselves to tenants. Traditionally, most of these tenants have been large enterprises and hyperscale cloud providers. In many cases, these hyperscale and enterprise tenants have also built their own data centers to support their own growth.
Since around 2017, Bitcoin mining has really reached an industrial level, with entire data center complexes being built specifically to support Bitcoin mining, and the electricity production and consumption in these regions varies greatly. Now, in 2023 and 2024, we see even more significant and disruptive changes in the market. As demand for GPU infrastructure surges, many data centers that were previously focused on colocation have begun purchasing and hosting this GPU infrastructure themselves. Hyperscalers, meanwhile, are moving behind the meter, co-located with large baseload power plants, seeking cheap, reliable power for a new surge in HPC demand. This is particularly noteworthy because intermittent renewables have been the most popular form of electricity generation in recent years, largely thanks to government subsidies.
we estimate:
1. Energy demands for both forms of computing continue to grow.
2. New data center construction will become the next bottleneck for expanding the HPC footprint, and a large number of Bitcoin mining facilities will be repurposed for higher-profit use cases.
3. Mining hardware will migrate to the edge, seeking out remote locations and places with variable inefficiencies where HPC workloads are not suitable for monetization.
4. Hybridizing Bitcoin mining and HPC in a “mullet data center” would leverage the high revenue potential of HPC and the flexibility of Bitcoin mining, effectively balancing power consumption and the local grid while outperforming traditional data center strategies.
in conclusion
When new energy-hungry technologies emerge, there is often concern about their energy utilization and their externalities. Bitcoin mining and HPC are no exception, with politicians and armchair technologists alike calling for them to be mitigated or controlled. But such energy-hungry technologies represent a natural trend in human progress. Beyond the self-evident utility provided by the Bitcoin settlement network and AI/ML workloads, we can show that they can be deployed in an efficient manner that maximizes the use of new and existing energy resources for useful economic purposes.
