IMF criticizes the environmental impact of Bitcoin and artificial intelligence

In a new report by the International Monetary Fund (IMF), authors Shafik Hebous and Nate Vernon-Lin look at the environmental impact of cryptocurrencies and artificial intelligence (AI) and possible government measures to solve this problem.

Right at the beginning of the report, the authors note that the commonality between AI and cryptocurrencies is the large demand for electricity.

Crypto mining and data centers are now responsible for 2 percent of global electricity consumption and almost 1 percent of global emissions, and their footprint continues to grow.
_{Excerpt from the report}

Based on another IMF report and an estimate by the International Energy Agency, the authors put the two sectors' global shares of electricity consumption at 3.5 percent and emissions at 1.2 percent in 2027.

Apart from the fact that the electricity consumption of the Bitcoin network was overestimated in the reporting due to the flawed Cambridge study, the IMF report suggests to the reader that the footprint of both technologies continues to grow, which is also partially supported by the figures. The fact that this is a falsely conveyed impression will be explained in more detail in a moment.

The authors of the IMF report also criticize the government incentives and the lack of taxation for companies in these sectors. In order to reduce their emissions, the authors propose transnational measures.

For example, they call for CO₂ credits, certificates for sustainable energy sources and a global price on carbon. This should prevent the migration of companies and encourage industries to use more energy-efficient devices and clean energy sources, while reducing the use of fossil fuels and CO₂ emissions. In addition, policy incentives could create AI applications that lead to more efficient energy use and thus a reduction in electricity demand, the authors note.

Finally, the authors propose the introduction of a special tax per kilowatt hour (kWh) used, also taking into account the effects of air pollution. This would generate enormous additional revenue for the state, but would increase the price of electricity for miners - and therefore the majority of their total costs - by 85%.

However, the authors set a lower price for the taxation of artificial intelligence than for Bitcoin mining, as AI data centers allegedly use "greener energy sources" than the mining facilities. The authors suggest a price of USD 0.052 per kilowatt hour used for AI applications and USD 0.089 for Bitcoin mining. The basis for the price of the Bitcoin mining tax was another IMF report by the same authors entitled "Cryptocarbon - How much is the corrective?". In this report, however, the authors refer to sources such as Alex de Vries or Mora et al. - which have already been refuted several times. The figures used in the Cambridge study are also demonstrably out of date. Furthermore, this report deals exclusively with cryptocurrencies. The basis for the better price of the AI tax is not mentioned. Ultimately, the current IMF report once again gives the impression that AI data centers have less of an impact on the environment than Bitcoin mining facilities.

Although the chart used in the report shows that only the shares of AI will increase by 2027, while the shares of cryptocurrencies will decrease, there is no correct classification of AI applications and Bitcoin mining.

The authors of the IMF report ignore many environmentally friendly aspects of Bitcoin mining, which have been documented in various studies and for the most part also taken up by the mainstream media. Instead, they rely on outdated data sets and discredited research findings, thereby stripping their "research" of any credibility.

Instead of correctly classifying Bitcoin mining and AI applications, the authors equate the two technologies in terms of environmental impact. However, this is either a misconception due to poor research work or a deliberate attempt to cast Bitcoin mining in a worse light. The latter seems more likely due to the IMF's conflict of interest regarding Bitcoin. For this reason, policy makers should avoid IMF reports.

In order to better understand the facts of the case, it is necessary to look at other academic papers, which the authors of the IMF report have unfortunately chosen not to do.

It is true that Bitcoin mining and artificial intelligence have things in common. They consume a lot of electricity and are therefore also competitors when it comes to the purchase of electricity - especially in the USA, where the expansion of electricity capacities has not been pushed forward since 2007. In addition, both technologies can be installed and operated in mobile and modular data centers.

However, they differ in terms of energy consumption patterns, flexibility and scalability, cost sensitivity and profitability, independence of location and time, and ultimately environmental impact.

In a study by the Bitcoin Policy Institute, authors Margot Paez and Troy Cross use the metaphor of locusts and dung beetles to illustrate the difference between AI data centers and Bitcoin mining facilities.

According to the report, Bitcoin mining plants are "flexible, scalable, transportable, location-independent and price-sensitive power consumers" that are tireless in their search for surplus energy (waste) - for example, when combined with intermittent solar or wind power, when utilizing unused hydropower or geothermal energy or for methane reduction in oil fields or landfills. They utilize energy in an often helpful way.

Miners can only be profitable with the cheapest energy (usually less than 0.05 US dollars per kWh), which otherwise finds no buyers. Such prices are mainly possible with sustainable energy sources, which means that the profit margins of plants that use fossil energy are predominantly smaller. Due to halving events, regular modernization of equipment is also necessary to remain profitable.

The wasted energy can be absorbed by miners anywhere in the world. High latency in the inter-connection is not a problem, as the use of communication resources is minimal. In addition, the miners are able to adapt to power shortages and surpluses and ramp up and down operations within a few seconds. On the one hand, this happens when electricity is scarce or the price of electricity is too expensive. On the other hand, the miners also voluntarily reduce operations if the grid operator requires them to do so, as is the case with the demand response programs of the Texan electricity grid operator ERCOT, for example. Through this demand management of Bitcoin miners, an electricity grid operator can prevent grid outages, improve the reliability of the grid and thus stabilize the grid and electricity prices.

So while Bitcoin mining consumes large amounts of electricity, just like the common dung beetle, it mainly uses what others leave behind.
_{Excerpt from the study}

In contrast, the authors of the study associate locusts with conventional data centers, for example for email servers or managing video streams, as well as the new types of data centers for AI or HPC (high-performance computing) with significantly higher power density. The locusts pounce on every energy source and consume it on their terms. They consume as much energy as the power grid can produce, which drives up prices and emissions.

In particular, AI training and inference, where the AI learns and draws conclusions based on user queries, are responsible for the high energy consumption. In contrast to Bitcoin mining systems, AI data centers also require additional power generation for peak demand.

According to the study, the profit margins for AI applications of 3 to 5 US dollars per kWh for Nvidia graphics processing units (GPUs) are currently up to 25 times higher than the profit margin for Bitcoin mining, where the most efficient machines generate 0.17 to 0.20 US dollars per kWh. This means that the locusts are less sensitive to electricity costs. They can be profitable even when the price of electricity is high, so the price of electricity need not be a determining factor for location and the locusts can pay more than the dung beetles or other electricity users. However, large-scale AI models require multiple GPUs. Many AI companies aim for the minimum size of 10 megawatts for a plant, while mining companies can generate profits from the first ASIC miner in a location.

Although location is irrelevant in terms of the price of electricity for AI applications, locusts are not location-independent. In particular, the inferences are dependent on location and demand. They require low latency times in order to be able to process requests from urban areas quickly. The locusts are therefore forced to locate their operations close to densely populated areas. In addition, the value of AI calculations can be time-dependent, meaning they are not always worth the same, unlike Bitcoin mining, whose rewards are usually worth almost the same at any time of day.

A major difference between Bitcoin mining and AI models lies in the energy consumption pattern. While the flexibility of dung beetles can perfectly compensate for the fluctuating energy supply from renewable energies, the power consumption of locusts is largely inflexible. AI data centers must be available around the clock. The contractual uptime of the machines is 99.99 percent, so without special software to manage the workloads, the machines cannot be switched on or off spontaneously and are therefore not as adaptable to the intermittent nature of renewable energy.

In addition, AI applications require special cooling, which, according to Steven Barbour (Upstreamdata), is also very expensive and water-intensive and requires better air quality than in Bitcoin mining facilities.

However, the authors also note that there can be overlaps between dung beetles and locusts. Accordingly, there are also short periods of time in which Bitcoin mining companies act like locusts and can mine profitably even when electricity prices are high - especially when the Bitcoin price is high. However, the incentive structure attracts more miners, who reduce the chance of finding a block and the profit margins for all participants. Finally, favorable electricity prices become more important for profitability.

According to the authors' assumptions, AI applications are increasingly developing into dung beetles, which could also be more flexible in the future, at least during training, in order to take advantage of the most favorable electricity prices. The corresponding software, which regulates the flexible power consumption of Bitcoin miners, already supports AI applications. Low-cost energy sources will therefore also become more important for AI data centers, especially if model training is distributed.

As in the IMF report, the authors of the study attempted to estimate the future electricity consumption of the two sectors. However, the authors point out several times in the study that this is a difficult undertaking. Forecasts of future energy consumption are even more difficult than estimates of current or past consumption. As there are many factors that could influence the results, caution is advised when dealing with such estimates - there are also some inaccuracies in this study. There are no current industrial or academic studies on this topic.

The authors estimate the electricity consumption of Bitcoin mining facilities in the US at around 48 terawatt hours (TWh) in 2023 and forecast around 160 TWh for 2027. Estimates for the electricity consumption of US AI servers in 2023 vary widely from 20-125 TWh. This means that AI applications in the USA could already consume more than twice the electricity of mining facilities.

The growth rate of AI is also expected to be much higher - more capital is currently flowing into the AI market. The International Energy Agency predicts a 10-fold increase in electricity consumption for AI data centers by 2026. Further forecasts for 2027 put the electricity consumption of US AI data centers at between 70 and 240 TWh. According to this study, energy consumption is therefore guaranteed to increase and continue to make headlines.

However, the authors would like political decision-makers to take a different approach to increasing energy consumption. To this end, the study uses data from ten different mining companies in the USA and Canada to demonstrate for the first time the emission savings that result from the flexible nature of Bitcoin mining facilities. The mining facilities studied were switched off between 5 and 31 percent of the time when the price of electricity was too expensive or as part of demand response programs. This flexible energy consumption of the dung beetles caused far fewer emissions than the inflexible data centers of the locusts: In three months, the systems avoided more than 13,500 tons of carbon dioxide compared to the data centers in continuous operation. This shows that the flexible nature of the dung beetles is not only important for stabilizing the power grids, but also for avoiding additional emissions.
The use of waste heat for heating processes of various kinds also leads to further emission savings.

Flexible data centers such as Bitcoin mining facilities [help] grid operators to decarbonize the grids, while inflexible data centers lead to grid operators having to feed more fossil base load into the grid to meet their round-the-clock requirements.
_{Daniel Batten} , _{The Bitcoin ESG Forecast #14}

Other experts also underline the ecological advantages of Bitcoin mining compared to AI applications. According to Steve Barbour, for example, the cost of installing an AI data center per megawatt is 10-50 times higher than that of a Bitcoin mining facility. The resource intensity and the associated environmental impact are correspondingly higher. The necessary base load for AI data centers also requires redundant, reliable power sources, which are primarily achieved with coal and natural gas. In the event of power outages or to increase flexibility, additional diesel generators are required to provide backup. These aspects all have an impact on emissions.

In addition, according to Barbour, not only is the cooling infrastructure in an AI data center more extensive, but also the operating time and skills of the workforce are significantly higher than in Bitcoin mining facilities.

Barbour also estimates that AI data centers also require more space, as the power density of AI rigs is only half that of mining rigs.

Daniel Batten also sees a better environmental balance for Bitcoin miners in terms of electronic waste. While many computer parts and GPUs contain environmentally harmful substances and are therefore not so easy to recycle, according to Batten, ASIC miners should not contain any harmful substances such as heavy metals and should be 100 percent recyclable. Previous research has repeatedly been based on incorrect data from Alex de Vries, for example regarding the useful life of ASICs. The devices are used for more than five years, which is almost three times higher than de Vries' estimates.

Ultimately, ASICs discarded by the industry often find new uses in other regions of the world, for example as a catalyst for the expansion of power grids and renewable energy sources.

With the results of the study, the authors of the Bitcoin Policy Institute study are addressing political decision-makers. They should focus less on the absolute power consumption of data centers and more on how data centers consume power. Ultimately, policymakers should create the appropriate structures and incentives for flexibility and the integration of computing power and energy generation in order to promote the flexibility of data centers, the reliability of power grids and the reduction of emissions.

The Bitcoin protocol has created a compulsion to seek low-cost power and thus flexibility. Policymakers should not view the flexible load of bitcoin mining as a fixed load and should ensure that "data centers of all types that offer demand-side flexibility coordinate with grid operators". Flexibility is also possible for all other types of data centers. Each data center could optimize its own backup mode to be more flexible or, for example, generate its own electricity to relieve the burden on utilities. While this mode of operation could increase emissions and capital costs for the grid, demand response programs could offset these costs.

Companies like Verrus are already planning flexible data centers with battery microgrids to be completed in 2026/2027. In addition, more and more Bitcoin mining companies are also integrating artificial intelligence into their systems. In addition, big data companies such as Microsoft, Meta, Google and Amazon are investing in energy generation projects and are thus also becoming energy companies.

The flexible data centers ultimately accelerate the expansion of sustainable energy sources, as more flexibility can absorb more renewable energy and thus enable additional revenue for electricity producers. This ultimately also counteracts the negative consequences of a "plague of locusts".