Bitcoin’s Other Footprint: Mining Hardware and E-Waste in 2026

The footprint the energy debate keeps forgetting

Almost every argument about Bitcoin and energy is an argument about a single number: the watts flowing into the network. Is it 138 terawatt-hours a year or closer to 190? Is the fuel mix half renewable or mostly gas? Does it rival Thailand or Argentina? Those fights have filled conference panels, regulatory filings, and a long shelf of explainers, including several of our own. They all share one blind spot. They measure the electricity and ignore the machines that burn it.

That omission matters more in 2026 than it did a year ago. Bitcoin trades around $81,000, roughly a third below its October 2025 record, yet the network has never worked harder. Mining difficulty jumped 5.92% on October 16, its third-largest increase of the year, to about 140.6 trillion, the highest reading since March per CoinWarz, and total hashing power is back above one zettahash per second (more than 1,000 exahashes). Every bit of that record output runs on purpose-built silicon that is being manufactured, deployed, and discarded faster than any other class of consumer or industrial computer on earth.

This is Bitcoin’s other footprint: not the power plant feeding the machine, but the machine itself. It is made of advanced chips that took energy and rare inputs to fabricate, it does useful work for a year or two, and then it is pushed off the economic cliff by something faster. The headline energy-mix figures say nothing about it. Neither, as we will see, do the disclosure rules now taking hold in Europe and quietly dissolving in the United States. This piece is an attempt to measure the part of the ledger the watts-in number never reaches, and to be honest about how uncertain, and how recoverable, that part actually is.

Two ledgers: watts in, machines out

Think of mining’s environmental cost as two separate accounts. The first is the operating ledger: the electricity a running fleet consumes, minute by minute, for as long as it is switched on. This is what the Cambridge Centre for Alternative Finance (CCAF) and the Digiconomist Bitcoin Energy Consumption Index try to pin down, and it is where the whole measurement war lives. The 2025 Cambridge study put annual consumption at 138 TWh, about 0.5% of global electricity, with 52.4% of it coming from sustainable sources.

The second account is the hardware ledger: the energy and materials embodied in building the machines, plus the electronic waste they become when they retire. This account barely appears in the public debate, and when it does it is a footnote, usually one line about e-waste tucked at the end of a consumption report. Yet it is a distinct physical cost with its own dynamics. The operating ledger scales with how much power the fleet draws. The hardware ledger scales with how fast the fleet turns over, which is a different variable entirely, governed by chip economics rather than kilowatt prices.

The two accounts can move in opposite directions. A miner that swaps an old rig for a new one that does three times the work per watt improves the operating ledger (more hashes per kilowatt-hour) while worsening the hardware ledger (a functioning machine just became scrap). Efficiency, the hero of every green-mining pitch, is also the engine of the waste problem. That tension is the real subject here, and it is why a serious look at Bitcoin’s energy mix has to count the machines, not just the megawatts.

The 1.3-year machine

The single most-cited number in this debate comes from a 2021 paper by Alex de Vries and Christian Stoll, Bitcoin’s growing e-waste problem, published in the journal Resources, Conservation and Recycling. Their estimate of the average economic lifespan of a mining chip was 1.29 years. De Vries, who also runs Digiconomist, called that span “extremely short compared to any” other electronic device, and the comparison is hard to argue with: people keep phones for three or four years and laptops for five or six.

The word doing the work in that sentence is economic. A mining machine does not stop functioning after sixteen months. It stops being worth the electricity it draws. An ASIC (application-specific integrated circuit) is a single-function device: it computes SHA-256 hashes and nothing else. When a newer model arrives that produces more hashes per watt, every older machine in the fleet earns less relative to its power bill, and at some point the power bill wins. The chip still works perfectly. It is simply no longer the most profitable way to spend a kilowatt-hour, and in a commodity business competed down to the marginal cost of production, that is a death sentence.

Digiconomist’s live Bitcoin Electronic Waste Monitor currently assumes mining gear turns over in roughly 1.5 years and translates that into about 83 grams of e-waste per transaction, which it likens to tossing out half an iPhone. Whether you accept 1.3 years or 1.5, the point is the same: the schedule on which Bitcoin retires hardware is set not by wear but by the pace of a technology race. Understand the race and you understand the waste.

What actually kills a miner: the efficiency arms race

ASIC efficiency, measured in joules per terahash (J/TH, lower is better), has fallen about elevenfold in a decade. The Antminer S9 that defined the 2016 to 2018 era needed roughly 98 J/TH. Today’s flagships are closing in on 9. Bitmain’s Antminer S21 XP does 13.5 J/TH; its hydro-cooled S23 generation reaches 9.5; Bitdeer’s SEALMINER A4 Ultra claimed an industry efficiency record at 9.45 in April; and the Antminer S23 XP Hyd, with first batches due to ship in November, is rated at 8.9 J/TH per the latest mining-hardware roundups, the first sub-9 Antminer. Bitmain’s own specifications for the prior generation are published on its support site.

Each of those steps forward shortens the useful life of everything already on the floor. There is a second, quieter depreciator working alongside it: difficulty. Because Bitcoin targets one block every ten minutes regardless of how much hashing power shows up, every new machine plugged in anywhere on earth dilutes the share of rewards your machine can win. The 5.92% difficulty jump on October 16 means a rig that was marginally profitable in September is closer to the line today, without anything changing inside the box. Hardware obsolescence in mining is therefore partly a story about better chips and partly a story about a network that keeps raising the bar on the chips already running.

How big is the pile?

The honest answer is that nobody knows precisely, and the range of estimates is wide. The de Vries and Stoll paper put annual mining e-waste at about 30.7 metric kilotonnes as of 2021, which they compared to the small-IT-equipment waste of a country like the Netherlands. Digiconomist’s live monitor currently shows a lower figure, around 23 kilotonnes a year, reflecting different assumptions about fleet size and turnover. Greenpeace USA has argued the number could climb toward 64 kilotonnes in a high-price environment, when miners rush to deploy the newest gear and retire the rest. The inflow side is just as striking: Greenpeace’s analysis found that Bitmain alone shipped more than 5.4 million kilograms of new machines, roughly 6,000 US tons, into the United States between the middle of 2023 and early 2024, a single manufacturer’s single-country haul in well under a year.

What makes mining e-waste unusual is not only the volume but the un-recyclability. A retired gaming GPU can be resold to a researcher, a renderer, or an AI startup; the silicon is general-purpose. A SHA-256 ASIC can do exactly one thing, and when it can no longer do that thing profitably it has no second computing use at all, not even for mining a different coin. Greenpeace notes that the machines contain toxic process chemicals that make dismantling hazardous, that only about 17% of global e-waste is formally recycled, and that no major ASIC manufacturer participates in mainstream electronics-sustainability certification. The result is a waste stream that is both concentrated and awkward to process, even where the will exists.

A note on the per-transaction framing those sources use (grams of e-waste, or kilowatt-hours, per transaction). It is intuitive but misleading. Bitcoin’s hardware turnover tracks hashrate and block rewards, not how many payments clear; a day with half the transactions still retires exactly the same number of chips. Layer-2 systems and batching can move huge transaction volumes without adding a single machine. Per-transaction metrics make the footprint look like it scales with usage. It does not. It scales with the economics of the race.

The embodied-energy blind spot

There is a second reason the machines belong in an energy-mix discussion, and it is literally energy. Before an ASIC hashes anything, a fab has already spent a great deal of electricity and water turning sand into leading-edge logic. Advanced chips are among the most energy-intensive manufactured objects in the world. According to a study of chip production’s footprint by the think tank Interface, a single cutting-edge wafer can carry on the order of one to two tonnes of CO2-equivalent before it leaves the plant, dominated by fab electricity and by fluorinated process gases whose warming potential runs into the thousands. TSMC, which fabricates the chips inside most mining rigs, used roughly 25 TWh of electricity in 2023, about 6% of Taiwan’s entire power supply. A single leading-edge fab can draw 50 to 100 megawatt-hours every hour it runs and consumes ultrapure water at the rate of a mid-sized city, so a new chip reaches the mine already carrying an energy and water debt before it earns a single satoshi.

How much of that embodied energy should be charged to Bitcoin? Honesty requires a caveat here that the loudest critics skip. For a device that runs flat out, 24 hours a day, for its whole life, the energy spent making it is small next to the energy it consumes operating. A mining ASIC is the opposite of a smartphone, where manufacturing dominates lifetime emissions precisely because the phone sips power. So embodied energy is a real line item, not a dominant one. But it is not zero, and crucially, it is re-incurred every time the fleet turns over. A network that replaces its machines every eighteen months pays that manufacturing cost far more often than one that runs hardware for five years, which drags the embodied-energy question back toward the churn problem at the center of this piece.

Mining is not alone in this. The same specialized-silicon dynamic drives the AI build-out and the cryptographic hardware behind zero-knowledge proofs, where purpose-built provers are becoming their own small arms race, as we covered in our look at zk-ML proving markets. Wherever custom chips chase a fast-moving performance frontier, embodied energy and premature obsolescence travel together.

The counter-case: machines don’t die, they migrate

Here is where the e-waste narrative needs a hard cross-examination, because the 1.3-year figure describes a decision to replace, not a machine hitting a landfill. The two are very different. An ASIC is physically good for five to seven years, and often ten. What de Vries and Stoll modelled was how fast operators at the technological frontier swap gear, then assumed the displaced machines are scrapped. In practice, a large share of them are not scrapped. They move.

A rig that is uneconomic at 12-cent power in Germany can be very economic at 3-cent hydro power in Paraguay, at flared-gas sites in Texas, or behind curtailed wind in the Nordics. The second-hand market that moves them is real and liquid: resellers and refurbishers advertise thousands of used units, a used S19 XP changes hands for roughly $800 to $1,500 against $5,000 to $7,000 for a new S21 Pro, and guides to the second-hand ASIC market describe payback periods of a few months for buyers with cheap electricity. Jaran Mellerud, co-founder of the mining-services firm Hashlabs, which sources hardware and coordinates repairs and hosting, is among the analysts who argue that access to cheap or stranded power, not chip age, is what decides whether an older machine keeps earning. Even the venerable S9 still runs in 2026 where power is nearly free or where its heat is the actual product.

This migration does not make the e-waste problem disappear; it defers it and relocates it, often to jurisdictions with weaker recycling infrastructure, which is its own concern. But it does mean the top-down estimates almost certainly overstate how much hardware is actually discarded each year. The true figure sits somewhere between the frontier-replacement model and the five-to-seven-year physical reality, and no global registry of retired ASICs exists to settle it.

The second life runs on the world’s cheapest power

The migration is not random. It follows the global power-cost map almost exactly. A machine that is dead weight on a US or German grid becomes a going concern wherever electricity is nearly free: sub-4-cent hydro in Paraguay, state-backed mining deals in Ethiopia, Himalayan runoff in Bhutan, flared-gas pads in the Permian Basin, or capacity behind curtailed wind in the Nordics. There is an irony worth sitting with. The oldest, least efficient boxes in the global fleet, the very ones a worst-case e-waste model has already written off, tend to end up running on the cheapest and often the cleanest marginal power on earth, because only that kind of power can make a 21 or 98 J/TH machine pencil out at all.

This is where the two ledgers reconnect. Keeping a paid-off machine alive on stranded energy avoids both the waste of scrapping it and the embodied energy of building its replacement, while putting power to work that had no better buyer. It is not a clean win. The same flexibility can prop up fossil baseload that would otherwise shut down, and it quietly exports the eventual disposal problem to the places least equipped to handle it. Kazakhstan’s 2021 boom, when cheap coal power drew a flood of machines only for a grid crunch and a government crackdown to strand them within a year, is the cautionary version of the same story. A second-life haven is only as durable as its power and its politics.

Immersion, firmware, and the repair economy

Operators have practical reasons, all of them financial, to make machines last. Cooling is the biggest lever. Heat is what kills hashboards, and the physics are unforgiving: by the Arrhenius rule of thumb, every 10 degrees Celsius cooler roughly doubles component life. Immersion cooling, which submerges rigs in dielectric fluid, can run chips 25 to 30 degrees cooler than air and, by industry accounts such as Braiins’ analysis, stretch useful life from the two-to-three years typical of hard-run air cooling to five, seven, or more. A machine that lasts twice as long is, by definition, half the e-waste.

Then there is the repair and refurbishment trade. Firms like D-Central rebuild dead hashboards, swap failed chips, reball controllers, and harvest parts from write-offs, giving a refurbished S19-class unit roughly a third of the cost-per-terahash of a brand-new S21. Custom firmware adds another dimension: undervolting and tuning can wring more efficiency out of aging silicon, nudging a marginal machine back over the profitability line without replacing it. None of this is charity. It is simply that in a business where the floor is your energy cost, a cheap machine that is already paid off and sips a little less power is a genuine asset, and the market has built an entire layer to keep those machines alive. Resellers move large volumes of used ASICs, and dedicated repair shops now stock replacement chips, fans, and hashboards the way a garage stocks parts, turning what critics picture as a disposable gadget into something closer to a maintainable industrial asset.

None of this fully solves the problem, and it would be naive to pretend otherwise. Repairs hit diminishing returns once a generation is several steps off the efficiency frontier, firmware cannot rewrite physics, and a machine kept alive for a decade still ends as waste eventually. The point is narrower but it matters: the interval between manufacture and landfill is a choice, not a constant, and the industry’s own economics now push that interval longer rather than shorter wherever the power is cheap enough to reward patience. That is the opposite of how the 1.3-year headline makes it sound.

Why Proof of Stake barely has this problem

The sharpest way to see that the hardware ledger is distinct from the energy ledger is to look at the chain that erased both at once. When Ethereum switched from Proof of Work to Proof of Stake in the Merge of September 2022, its electricity use fell by more than 99.98%, from roughly 21 TWh a year to about 0.0026 TWh, according to ethereum.org. That number is famous. Less discussed is what happened to the hardware.

Proof of Stake secures the network with capital, not computation. A validator is a general-purpose computer, often a small server or even a mini-PC, running ordinary software. It is not racing anyone on joules per terahash, so there is no arms race forcing its replacement every eighteen months. It lasts as long as a normal server, five to seven years, and when it retires it can be sold, repurposed, or recycled through the same channels as any other PC because the silicon inside it is not single-function. The economic security comes from staked ether, not from specialized chips, which is also why the risks in that world are financial and governance-shaped rather than material, a theme we explored in our coverage of restaking in 2026. The upshot is that Proof of Stake has essentially no mining-specific e-waste stream. Whatever you think of the security trade-offs, on the hardware ledger it is not close.

Does the e-waste number even hold up?

If the energy-mix figures are contested, the hardware figures are barely measured at all. The operating ledger at least has rival methodologies arguing over a known quantity of electricity. The hardware ledger has a single foundational academic paper, a handful of models built on top of it, and no primary data on how many machines are actually scrapped versus resold, refurbished, or shipped abroad. Every headline e-waste tonnage is an inference from assumed replacement rates, not a count.

That is worth keeping in mind precisely because the same caution applies to the energy side. Alexander Neumueller, who leads the digital-assets research at CCAF, framed the 2025 Cambridge study as an effort to close exactly this kind of hole, saying it “directly addresses a persistent data gap by relying on direct practitioner insights rather than abstractions.” Cambridge surveyed dozens of firms to get at the energy mix. No comparable survey exists for hardware end-of-life. So the intellectually honest position is a range: the e-waste problem is real, it is unusual because the chips cannot be repurposed, and it is almost certainly smaller than the frontier-replacement models imply because so much gear migrates instead of dying. For scale, even the highest mining estimates are a sliver of the tens of millions of tonnes of electronic waste the world throws out every year; the trouble with mining hardware is less its gross weight than that almost none of it can be reused, which turns a small pile into a stubborn one. Anyone quoting a single confident tonnage, in either direction, is selling something.

What would it take to do better? A real measurement would track machines, not models: import and customs records for new rigs, resale and export flows for the old ones, and end-of-life reporting from the large public miners who already file detailed accounts with their regulators. None of that is technically hard, because the data exists, inside companies and at borders. What is missing is any obligation to publish it, which is exactly the gap that defines the next part of the story.

The disclosure gap: Europe made e-waste optional, the US erased the question

If the hardware ledger is poorly measured, it is also poorly mandated. Europe came closest to forcing the issue. Under the Markets in Crypto-Assets regulation, the Commission’s sustainability-disclosure technical standard, Delegated Regulation (EU) 2025/422, requires crypto-asset white papers to publish the consensus mechanism’s total annual electricity use as a mandatory field, with renewable share, energy intensity, and greenhouse-gas emissions becoming mandatory above a usage threshold. Any consensus mechanism drawing more than 500,000 kilowatt-hours a year, which Bitcoin exceeds many thousands of times over, trips those supplementary requirements, among them an energy-intensity figure reported per transaction, the very metric that makes the least physical sense for a network whose power draw follows hashrate rather than payment volume. Electronic waste, however, sits in the optional category. A firm may disclose it; nothing compels it to. The one hard number regulators demand is a watts-in number. The hardware ledger remains, by design, a nice-to-have.

The United States is moving in the opposite direction, which for a largely US-hosted industry matters. The Securities and Exchange Commission’s climate-disclosure rule, adopted in 2024 and then stayed, is now the subject of a 2026 proposal to rescind it outright. Whatever one thinks of that rule’s scope, its retreat means the one plausible federal lever that might have asked large miners to quantify any part of their physical footprint, energy or hardware, is being switched off. Between an EU regime that makes e-waste optional and a US regime erasing the broader question, the machines-out side of the ledger is effectively invisible in official filings. Investors and the public are left with academics and activists on one side and industry blog posts on the other.

There is a defensible argument that hardware does not belong in a securities disclosure at all, that it is an environmental question for a different agency. Fair enough. But the practical result is a vacuum. No US agency currently asks, Europe treats the answer as voluntary, and the manufacturers volunteer nothing. For an asset class that spent 2026 courting institutional money and exchange-traded funds, the silence on half of its physical footprint is a strange look, and one a future rulebook in Brussels or Washington could close with a single line item.

What the treadmill does to a balance sheet

The hardware ledger is not only an environmental story. It is the central line on a miner’s income statement, which is why the industry behaves the way it does. Electricity is the cash cost everyone talks about, but the gap between a cash-profitable quarter and a reported loss is mostly depreciation: the steady write-down of machines that lose value on a schedule set by the efficiency race. Rigs are capital that decays whether or not they run, and the sub-9 J/TH frontier forces continual reinvestment just to hold position. That is the financial face of the same churn that produces the waste.

It also explains why the environmental fix and the economic fix point the same way. Running a machine longer, cooling it better, buying it used, and redeploying it to cheap power all improve both the hardware ledger and the margin. The pressure cutting against them is relentless: Bitcoin’s block subsidy halves roughly every four years, so miners must either ride a rising price or keep squeezing more hashes per watt out of ever-newer silicon, a dynamic we unpacked in our halving-cycle math. With Bitcoin stuck around $81,000 and the Federal Reserve still the dominant variable for the price, as our year-end outlook laid out, 2026 is a year when margins are tight enough to make every one of those levers worth pulling. The same scarcity logic that supports the asset’s long-run case, explored in our piece on stock-to-flow and liquidity, is what keeps the hardware treadmill spinning underneath it.

This is also why the AI pivot sweeping the mining sector cuts both ways for the hardware ledger. Converting a site to host graphics processors for artificial-intelligence tenants retires Bitcoin ASICs faster, which adds to the near-term waste, but it also extends the life of the machines that remain by thinning the field competing for the next block. The hardware does not disappear; it is redistributed, retired, or redeployed according to whichever kind of computation pays best that quarter. The treadmill does not stop. It just changes what it is running toward.

The bottom line: count the machines

Bitcoin’s hardware footprint is real, genuinely unusual, and routinely both exaggerated and ignored. It is exaggerated when a 1.3-year replacement decision is treated as a 1.3-year trip to the landfill, when per-transaction grams imply a cost that scales with usage, and when embodied energy is talked up for a device whose operating energy dwarfs it. It is ignored when energy-mix reports stop at the meter, when regulators make e-waste an optional checkbox, and when an industry that imports thousands of tonnes of new chips a year stays out of every mainstream recycling program.

The honest version sits in between. Tens of kilotonnes of specialized, hard-to-recycle silicon are retired each year; a large and uncounted share of it migrates to cheaper power rather than dying; immersion cooling, refurbishment, and firmware are quietly extending lives because doing so also protects margins; and Proof of Stake sidesteps the whole stream by securing itself with capital instead of chips. The levers that shrink the footprint are the same ones that defend a miner’s balance sheet, which is the most durable reason to expect them to be pulled. What to watch from here is not another consumption estimate but the plumbing: how deep the second-hand market runs, how fast immersion spreads, whether any recycler ever earns an ASIC manufacturer’s business, and whether a single disclosure regime ever asks miners to count their machines the way they are now forced to count their watts. Until one does, the energy mix will keep describing half of mining’s physical cost and calling it the whole.

Frequently Asked Questions

How long does a Bitcoin mining ASIC actually last?

It depends on whether you mean economics or physics. The most-cited research puts the average economic life at about 1.29 years, the point at which a machine stops being worth its electricity, while the hardware itself is physically sound for five to seven years and sometimes ten. Obsolete and broken are not the same thing in mining.

How much e-waste does Bitcoin mining produce?

Estimates range from roughly 23 kilotonnes a year on Digiconomist’s current monitor to about 30.7 kilotonnes in the foundational de Vries and Stoll study, comparable to the small-IT-equipment waste of a country like the Netherlands. All of these are modelled inferences from assumed replacement rates, not direct counts, so treat them as a range rather than a precise figure.

Can old Bitcoin miners be recycled or reused?

A working machine is usually resold and redeployed to a site with cheaper or stranded power rather than scrapped, which is the industry’s main answer to the waste problem. The catch is that a SHA-256 ASIC cannot be repurposed for any other computing task, so at true end-of-life it is hard to recycle, and only about 17% of global e-waste is formally processed.

Does Proof of Stake have an e-waste problem?

Essentially no. Ethereum’s validators run on ordinary general-purpose computers with normal five-to-seven-year lives and no efficiency arms race forcing replacement, so there is no mining-specific hardware waste stream, and the network uses about 0.0026 TWh a year after the 2022 Merge.

Why do Bitcoin miners replace hardware so quickly?

Two forces. New chips keep getting more efficient (from about 98 joules per terahash in 2016 to under 9 in 2026), and rising network difficulty keeps shrinking each machine’s share of rewards, so an older rig becomes unprofitable long before it breaks. Thin margins then push operators toward the newest silicon.

Written by the HOGE Wire mining and energy desk.