Cambridge Puts Ethereum's Power Draw at 0.90 Megawatts

The Cambridge Centre for Alternative Finance has put a measured figure on how much electricity Ethereum uses, estimating the network's annual consumption at 7.87 gigawatt-hours and its continuous power demand at 0.90 megawatts. The same work put the network's annual emissions at 2.37 kilotonnes of carbon dioxide equivalent. Set against the roughly 2.4 gigawatts of continuous demand the network drew before the Merge, its switch from proof-of-work mining to proof-of-stake validation, that is a reduction of more than 99.9%, according to the figures reported on Sunday by crypto.news and Coinpedia. The estimate is a measurement of the network as it runs today, built from node hardware rather than from the mining economics that governed the earlier design.

The Change: From 2.4 Gigawatts to Nine-Tenths of One Megawatt

Proof of work secures a chain by having specialised machines race to solve a computational puzzle, with a difficulty setting that adjusts to keep block production on schedule. Electricity is what makes that race costly, so the power draw of such a network scales with how much hardware is competing. Proof of stake replaces the race: the right to propose and attest to blocks is allocated according to capital committed to the protocol, and the machines involved are ordinary servers and desktops running the client software. The energy question therefore changes shape entirely. Instead of measuring an open-ended competition, the exercise becomes counting how many machines run the network and how much each one draws.

The Method: 8,522 Discoverable Nodes at About 105 Watts Each

Cambridge built its estimate from 8,522 discoverable full nodes — the machines that hold the chain's history and validate what other participants publish. Measured hardware ranged from 18 watts for a residential setup to 152 watts for a workstation-class machine, with a network-weighted average of roughly 105 watts per node. Those two figures produce the headline result directly: some 8,522 machines at about 105 watts apiece is close to 0.90 megawatts of continuous demand, and running that continuously for a year yields the 7.87 gigawatt-hour total. Converting electricity into emissions required a further assumption about where that power comes from, and the researchers used a grid mix of 39.4% renewable, 17% nuclear and 43.6% fossil generation.

The hardware range is the part of the method that connects to protocol design rather than to accounting. Modest machines can validate the chain today, and keeping it that way is the direction the protocol's own researchers have argued for in public, most recently in the case for an extremely lean base layer. Requirements that push node operators toward heavier hardware would raise the per-node figure; requirements that lower them would cut it.

Against Other Networks: Solana at 13.48 Gigawatt-Hours

The centre published comparable figures for other chains. Solana came in at 13.48 gigawatt-hours a year, BNB Chain at under 1 gigawatt-hour, and NEAR, Tron and TON in a band between 3.6 and 5.1 gigawatt-hours. Because raw totals reflect how many machines a given network runs rather than how much it settles, the researchers also published an intensity measure: energy used per $1 million of market value. On that basis Ethereum registered 33 kilowatt-hours and Solana 283 kilowatt-hours, a difference of about 8.5 times. Both measures are sensitive to the same inputs — node counts and per-node draw — so they move together when either assumption is revised.

What Is Not Settled: A Node Count and an Assumed Grid

The figure rests on nodes that can be found from the outside. Machines that do not advertise themselves are not in the 8,522, and the grid mix behind the emissions number is an assumption applied to the whole network rather than a measurement of where individual operators actually draw power. The eightfold spread between an 18-watt setup and a 152-watt workstation means the weighted average is doing considerable work. None of that undermines the order of magnitude, which is what the comparison with the pre-Merge network turns on. It does mean the precision implied by two decimal places belongs to the model rather than to the meter.