Read our latest report assessing the present and future energy impacts of Bitcoin and AI.
The rapid rise of bitcoin mining, and more recently, the computing required for artificial intelligence (AI), have both raised alarm regarding their consumption of electricity and their emissions. Yet, the energy and emissions impacts of these two forms of computation are vastly different.
In this report, we explain and document those differences, outline the challenge of meeting the coming electrical demand, and make recommendations to policymakers for how best to meet it.
1. We explain bitcoin mining’s basic design parameters as a kind of energy-intensive lottery that secures a distributed network and fairly distributes new bitcoin. Those design parameters cause it to be a flexible, scalable, portable, location-agnostic, and price-sensitive consumer of electricity. That, in turn, dictates its relentless search for waste energy, whether mitigating flare gas in remote oil fields or co-locating with solar and wind installations. Bitcoin miners, like dung beetles, subsist on waste.
2. For contrast, we compare AI computing to locusts, which feed on valuable crops: their power usage is mostly inflexible, location-dependent, scale-dependent, and price-insensitive. AI data centers add to peak demand, requiring additional peak generation from grids.
3. Estimates of AI server electricity usage (20-125 TWh) range widely, but AI could already use more than double the power of the bitcoin network (48 TWh), and has a steeper rate of growth.
4. Caveats: We see AI eventually evolving to be more like bitcoin mining in the future as compute loads are shifted to the cheapest power source. We note, too, that in brief periods of bitcoin price appreciation, bitcoin mining can behave in a locust-like way, when mining is profitable on all available electricity.
Bitcoin miners often tout their ability to turn their machines off whenever electrical grids are stressed, but until now, the extent of miner flexibility has not been independently substantiated. We gathered detailed power usage data from 10 bitcoin mining companies in the US and Canada and found that they curtailed power usage between 5% and 31% of the time.
We used marginal emissions factors from RESurety and WattTime to compute the amount of carbon dioxide emissions avoided when a bitcoin mining facility reduced its power demand, relative to a constant-uptime data center. Over three months, the miners we studied reduced their carbon dioxide emissions by 13.6 kilotons. If our sampled miners are representative of the network, the marginal impact of miner flexibility is equivalent to avoiding 4.4 million tons of carbon dioxide annually, or taking 956,521 cars off the road.
We need informed and differentiated policy that encourages both bitcoin mining and AI to evolve to more flexible patterns of energy use, such as rate structures that reward demand response and reflect scarcity and abundance in pricing. We also need transmission and interconnection reform, which would allow power generation projects to meet the coming demand.
