How Bitcoin Mining Works: PoW, Hardware & Energy Use

Introduction

A miner can run one of the fastest Bitcoin machines ever built for 101 years and still expect to find no block on its own. That lottery arithmetic shapes everything about mining: why pools exist, why hardware turns obsolete within a few years and why electricity prices decide who survives. Network hashrate stood at about 1,020 exahashes per second (EH/s) on 4 November 2026, after falling 27% between October 2025 and February 2026. Fees supplied only 0.59% of block rewards, so the subsidy pays almost the whole bill, and on 4 November 2026 bitcoin sat 37.9% below its October 2025 high. Hashprice, the daily revenue of one petahash, recovered from a $27.7 monthly average in June to about $39 in October, then slipped to about $33 on 2 November. This guide follows a block from template to proof of work, then covers difficulty, ASICs, post-halving miner economics and the disputed energy numbers.

What Does a Bitcoin Miner Do, and Why Does Bitcoin Need Miners?

Bitcoin miners are the computers that agree on the order of transactions without a referee, and they earn new coins for doing it. Mining turns that agreement into a costly race: about 1,020 exahashes per second competed for each block on 4 November 2026 (mempool.space API, 2026-11-04).

Ordering transactions without a referee

Digital money has one hard problem: stopping the same coins from being spent twice. A bank solves it with a central ledger. Bitcoin lets any miner propose the next page of a public ledger, called a block, and makes each proposal expensive to produce. Nodes, the computers that enforce the rules, reject any block that breaks them, so miners decide only which valid transactions go first. Because a block costs real electricity, rewriting an old block means redoing that work. That cost is what proof of work (PoW) names. The pillar guide to Bitcoin ↗ covers the ledger itself; this guide covers the machines that extend it.

What miners are paid for

Each block pays its miner a block subsidy of newly issued bitcoin plus every fee attached to the transactions inside. The subsidy has been 3.125 BTC since block 840,000 on 20 April 2024 (Bitcoin protocol; mempool.space, 2026-10-20). Fees supplied 0.59% of rewards over the 144 blocks to 4 November 2026: 2.68 BTC of 452.68 BTC (mempool.space API, 2026-11-04). Miners do not process payments, verify balances or hold user funds. They sell one service, making the ledger expensive to alter, and the network pays for it in subsidy and fees.

Every hash attempt works on a block the miner has already assembled, and that assembly decides which transactions get paid.

How Does a Miner Assemble a Block Before Any Hashing Starts?

A block template is the miner's draft of the next block: a ranked selection of waiting transactions, one special reward transaction and an 80-byte header that commits to both. Building it takes milliseconds; the hashing that follows consumes the electricity. Consensus rules fix every part of the template except the transaction choice and one number.

The coinbase transaction and the fee queue

Unconfirmed transactions wait in each node's mempool. A miner ranks them by fee rate ↗, measured in satoshis per virtual byte (sat/vB), and fills the block up to the limit of 4 million weight units (Bitcoin protocol). Blocks averaged 1.54 MB on 14 October 2026 (Firsthand Bitcoin, 2026). The first transaction is the coinbase, which has no inputs and pays the miner the subsidy plus all fees. The block-building code ranks a low-fee parent together with the high-fee child that pays for it, so both enter the block when the pair clears the fee bar (Bitcoin protocol). With the fastest confirmation costing only 2 sat/vB on 4 November 2026, the queue barely rewards bidding (mempool.space API, 2026-11-04).

The Merkle root

A block holds thousands of transactions, yet the header must stay tiny. Miners solve this with a Merkle tree: each transaction is hashed, the hashes are paired and hashed again, and the process repeats until one 32-byte value remains, the Merkle root. Changing a single transaction changes the root, so the header commits to the whole list. The average block on 28 September 2026 held 4,728 transactions (Firsthand Bitcoin, 2026-09-29), all summarised in that one value. Nodes can later prove a transaction sits in a block with a short path of hashes instead of the full list. Each miner's coinbase pays a different address, so every miner's Merkle root differs and no two miners search the same set of headers.

The 80-byte header

Hashing never touches the full block, only an 80-byte header made of six fields. The version, previous block hash, Merkle root, timestamp and target bits come from the template, which leaves the nonce as the one field a miner changes freely on every attempt. The previous block hash inside the header chains each block to its parent, so altering an old block breaks every later header. The nonce field holds only four bytes, so machines also roll reserved bits in the version field, a technique known as version rolling, to widen the search without rebuilding the Merkle root (OCEAN DATUM Gateway documentation, 2025).

Data current as of November 2026.

Once the header is written, the search begins, and it has only one method.

How Does the Nonce Search Turn Electricity Into Proof of Work?

A header counts as valid proof of work when its double SHA-256 hash, read as a number, falls below the network target. No shortcut exists: a machine changes the nonce, hashes again and checks the result, billions of times per second. Difficulty stood at 148.36 trillion on 4 November 2026 (mempool.space API, 2026-11-04).

Hashing against a target

SHA-256 is a one-way function: the same input always returns the same 256-bit output, but nobody can predict the output without computing it. The target is a very large number the hash must beat, so valid hashes start with a long run of zeros. Difficulty measures how far the target has shrunk from its easiest setting. At 148.36 trillion, an average block takes about 6.4 × 10^23 hashes, which is difficulty multiplied by 2^32 (own calculation from mempool.space data, 2026-11-04). The 32-bit nonce offers only 4.3 billion values, so machines also vary an extra nonce inside the coinbase transaction, which changes the Merkle root and opens a fresh set of nonces.

Why a single machine almost never wins

Mining is a lottery in which every hash is a ticket and difficulty fixes the odds. A 200 TH/s machine computes 2 × 10^14 hashes per second, so it expects one solo block every 101 years (own calculation, same inputs). That is an average, not a schedule: the machine holds a chance near 0.003% of winning on any given day. Almost all hashrate therefore works through pools, which trade a rare large win for frequent small payouts. Luck also makes block arrival irregular, even though the average stays near 10 minutes.

Hash luck averages out only because the target moves with the network's total power.

How Does the Difficulty Adjustment Keep Blocks Near Ten Minutes?

Bitcoin aims for one block every 10 minutes, and every 2,016 blocks the protocol resets difficulty to match the time the previous 2,016 blocks took. Faster blocks raise difficulty; slower blocks lower it. This rule keeps block production steady while total hashrate swings by hundreds of exahashes.

The 2,016-block retarget

The rule is plain arithmetic. Nodes compare the actual time of the last 2,016 blocks with the expected 20,160 minutes, two weeks, and scale the target by that ratio, limited to a factor of four per adjustment (Bitcoin protocol). No committee, vote or calendar is involved: every node recomputes the same result from block timestamps. The formula counts 2,015 intervals between 2,016 blocks, an off-by-one inherited from the original code (Bitcoin Core consensus code). Between retargets, blocks run fast or slow with hashrate: the current period averaged 10.3 minutes per block on 4 November 2026 because hashrate was running below the level the new difficulty assumed (mempool.space API, 2026-11-04). The latest adjustment came at block 973,728 on 29 October 2026 and raised difficulty 5.54% to 148.36 trillion. The next falls at block 975,744, around 13 November, and mempool.space estimated a fall of about 2.6% on 4 November (mempool.space API, 2026-11-04).

A year of difficulty moves

Difficulty reached a record 155.97 trillion on 29 October 2025, then followed hashrate down. It fell 11.16% on 7 February 2026 and 10.09% on 14 June 2026, when it reached the year's low of 124.93 trillion (mempool.space adjustments series, 2026-10-20). The largest rise, 14.73% on 19 February, reversed the February cut within 12 days. Each cut handed the remaining miners a larger share of the same reward. Difficulty then recovered to 140.58 trillion by 16 October and 148.36 trillion on 29 October 2026, still about 5% below the 2025 record. The link between these moves and bitcoin's price has its own treatment in a companion explainer on why hashrate often moves before price ↗.

Falling difficulty helps only the miners whose machines still cover their power bill, and that depends on the hardware.

How Do ASICs and Mining Pools Decide Who Finds Blocks?

Bitcoin mining runs on application-specific integrated circuits (ASICs), chips that perform SHA-256 and nothing else, and on pools that merge thousands of machines into one entrant. Four pools found 71.3% of the 982 blocks mined in the seven days to 4 November 2026 (mempool.space pools API, 2026-11-04).

From CPUs to ASICs

Early miners used ordinary processors, then graphics cards. Dedicated ASICs replaced both because a chip built for one function wastes no silicon on anything else. Efficiency, measured in joules per terahash (J/TH), improved from 98 J/TH on the 2016 Antminer S9 to 9.4 J/TH on Bitdeer's SEALMINER A4 Ultra Hydro and 9.5 J/TH on Bitmain's Antminer S23 Hydro in 2026, roughly ten times less energy per hash (Sterling Charts, 2026; CoinLaw, 2026). Air-cooled machines such as the WhatsMiner M70S+ reach 12.5 J/TH. Older fleets stay on only while power is cheap. CoinShares expects sub-10 J/TH hardware from Bitmain and Bitdeer to scale through the second half of 2026 and rig prices to soften as a result (CoinShares, 2026).

What a pool does and how it pays

A pool coordinates machines and shares the winnings. It hands each miner work, counts shares, which are hashes that nearly meet the target and prove effort, and pays members in proportion to their shares. A 200 TH/s machine that expects one solo block in 96 years instead collects small, steady payouts. The pool charges a fee for this service; OCEAN lists 2% as standard and 1% for miners who build their own templates (Spark, 2026-10-07). Foundry USA and AntPool alone found 44.8% of the 982 blocks. Pools hold no extra power over consensus rules, since nodes still reject any invalid block they submit.

Who builds the template

Pool operators traditionally choose the transactions, so a few operators decide what the next block contains. The top six pools produced 82.1% of blocks in the week to 4 November 2026 (mempool.space pools API, 2026-11-04). Two designs return that choice to miners. Stratum V2 Job Declaration lets a miner build its own template while the pool only accounts for shares, and block 955,318, mined on 25 June 2026 through the DMND pool, was the first known block of that kind (Spark, 2026-10-07). OCEAN's DATUM gateway does the same through a miner's own node, and OCEAN found 49 of the 982 blocks, or 5.0%.

Data current as of November 2026.

Regardless of the template builder, every block pays the same reward, and its dollar value now decides which machines survive.

How Do Miners Make Money After the 2024 Halving?

Miners earn the block subsidy plus fees, paid in bitcoin, against costs paid in electricity and hardware. Revenue per unit of hashrate, called hashprice, stood near $33 per petahash per second per day on 2 November 2026 on Luxor's Hashrate Index, which replaces the Firsthand Bitcoin and Mining Disrupt readings used in earlier versions of this guide, with bitcoin near $78,300 on 4 November 2026, 37.9% below its October 2025 peak (Luxor Hashrate Index, 2026-11-02; CoinPaprika API, 2026-11-04).

Hashprice, the miner's revenue per unit of hashrate

Hashprice divides a day's block rewards by the hashrate competing for them; one petahash per second (PH/s) equals 1,000 terahashes per second. The monthly average fell to $27.7 in June 2026, an all-time low (CoinShares, 2026), then recovered to about $39 by October, against $47 a year earlier (Firsthand Bitcoin, 2026-10-20). It slipped to $32.86 on 2 November, down 8.7% in a week in which difficulty rose 5.5% and bitcoin fell 3.5% (Luxor Hashrate Index, 2026-11-02). Hashprice rises with bitcoin's price and falls as difficulty grows. Counted in bitcoin, it held up better than in dollars: 0.000476 BTC per PH/s a day on 10 October 2026 against 0.000419 at the October 2025 record, because difficulty fell 12% while bitcoin's price fell 34% (Firsthand Bitcoin, 2026-10-11). The April 2024 halving cut the subsidy from 6.25 to 3.125 BTC, and a separate explainer on the halving ↗ covers the schedule and its effect on supply.

Fees are under 1% of the pot

Fees made up 0.59% of block rewards over the 144 blocks to 4 November 2026 (mempool.space API) and averaged 0.60% across 2026, the lowest annual share since 2015 (Firsthand Bitcoin, 2026-09-29). On halving day, block 840,000 paid 37.63 BTC in fees on 40.75 BTC in total, about 92%, because the Runes token launch ↗ triggered a bidding war (OneMiners, 2026). That spike did not return, and Glassnode noted fees had stayed below 1% of miner revenue for almost a year (Cointelegraph, 2026-08-12). The average transaction paid 492 satoshis in 2026, about 41 cents, even as daily transactions hit a record 578,012 (Firsthand Bitcoin, 2026-09-29). The subsidy supplies about 99% of what every machine shares.

Break-even power price by machine

A machine earns hashprice times its hashrate and pays for the electricity it draws. Dividing hashprice by energy use gives the highest power price at which it breaks even. At $33 per PH/s per day, the 9.5 J/TH Antminer S23 Hydro breaks even at 14.5 cents per kilowatt-hour (kWh), the 12.5 J/TH WhatsMiner M70S+ at 11.0 cents and the 98 J/TH Antminer S9 at 1.4 cents (own calculation, before pool fees, cooling and hardware cost). Listed miners' weighted ex-tax cash cost reached about $75,500 per bitcoin in the second quarter of 2026, while bitcoin ended the quarter at $58,400 (CoinShares, 2026).

Data current as of November 2026.

Margins this thin push owners of cheap power to ask whether bitcoin is the best use of it, and in 2026 many answered no.

Why Did Bitcoin's Hashrate Fall in 2026, and Where Did Miners Go?

Network hashrate fell about 27% from roughly 1,160 EH/s in early October 2025 to roughly 850 EH/s by early February 2026, the first half-year decline since China's 2021 mining ban, and stood at about 1,020 EH/s on 4 November 2026 (CoinShares, 2026; mempool.space API, 2026-11-04).

The 2025-26 hashrate swing

Three forces pushed machines offline. Hashprice sank toward its $27.7 June average, leaving older fleets below break-even. An estimated 252 EH/s went offline in the first quarter alone as older machines became unprofitable. Texas power prices also rose at the start of the year (CoinShares, 2026). Hashrate remained about 12% under the October 2025 peak on 4 November 2026 (own calculation). Each difficulty cut then passed relief to the miners who stayed. When bitcoin recovered from $58,400 at the end of June, hashrate climbed back to 879 EH/s on 13 July (CoinShares, 2026) and about 952 EH/s by 28 September (MillionMiner, 2026-10-07).

AI contracts and the miners who stay

Power and grid connections became worth more than the machines using them. CoinShares estimates AI and high-performance computing (HPC) earns about $1.5 million per megawatt a year against $0.5 million for mining. Core Scientific paid $41.9 million to cancel 15 EH/s of next-generation hardware, Keel stopped mining on 29 June 2026, and at least 35 EH/s more is scheduled to leave the listed miners, including IREN by 31 December 2026 (CoinShares, 2026). CoinShares expects a price recovery to leave this transition intact, and names Riot, MARA, HIVE and Bitdeer as the listed miners with the most flexibility to keep mining. The machines that remain share the same reward among fewer competitors.

At any machine count, the electricity those machines draw is the most disputed number in mining.

How Much Energy Does Bitcoin Mining Use, and Where Does It Come From?

No meter records Bitcoin's electricity use, so every figure is a model or a survey, and published estimates run from 138 to about 203 terawatt-hours (TWh) a year. Citing one number without its method and date misleads in either direction, and the spread matters more than any single figure.

Why the estimates disagree

The Cambridge Centre for Alternative Finance (CCAF) surveyed 49 mining firms covering about 48% of hashrate and estimated 138 TWh a year and 39.8 million tonnes of CO2-equivalent (CCAF, Cambridge Digital Mining Industry Report, 2025). Preliminary follow-up work presented in July 2026 puts annualised use near 190 TWh in December 2025, with emissions near 48 million tonnes (TheEnergyMag, 2026-07-25). Digiconomist's economic model counted 152 TWh from January to September 2026, a pace near 203 TWh a year (Digiconomist, 2026). Surveys capture what miners disclose; economic models assume miners spend a fixed share of revenue on power. The Cambridge daily index, CBECI, models hashrate and hardware efficiency instead, and an academic study built on it put 2025 anywhere from 94 to 391 TWh (Papana and Katrakilidis, 2026).

The energy mix

The CCAF survey found sustainable sources supplied 52.4% of surveyed mining power: 42.6% renewables and 9.8% nuclear. Natural gas supplied 38.2% and coal 8.9%, down from 36.6% in 2022 (CCAF, 2025). The preliminary update raises the low-carbon share to 59.4%, with hydropower now the largest single source (TheEnergyMag, 2026-07-25). Both figures describe survey respondents, who represent roughly half the network and lean toward US firms, so the shares describe the sample rather than every machine. Method alone moves the answer: an IP-address-based model produced 69.6 million tonnes of emissions against the survey's 39.8 million, and counting miners that burn otherwise-flared gas would lower the survey figure to 32.9 to 37.6 million tonnes (CCAF, 2025, via GreenCalculus, 2026).

Flexible load and the critics' case

Defenders cite flexibility: surveyed miners curtailed 888 GWh of load in 2023, and 70.8% reported climate-mitigation measures, a self-reported figure (CCAF, 2025). Critics cite scale and direction. Cambridge's preliminary work shows consumption up about 38% in the 18 months to December 2025 and emissions up about 20% (TheEnergyMag, 2026-07-25), while Digiconomist's January to September 2026 total ran 14.4% above the same months of 2025. Digiconomist divides that energy by transactions to reach 1,084 kWh each, yet miners are paid to secure the ledger, not per payment, so the ratio falls whenever transaction counts rise.

Data current as of November 2026.

That spend is the price of a security guarantee, and the last question is what it buys.

Is Proof of Work Worth Its Cost, and What Changes at the 2028 Halving?

Proof of work converts electricity into a cost that an attacker must also pay, and that cost is the whole security model. Whether the price is worth paying is a judgement; the mechanics, revenue and energy figures above are the evidence for it.

What the security spend buys

Rewriting history means redoing the work of every block since the change, faster than the honest network adds new ones, so the cost scales with hashrate, about 1,020 EH/s on 4 November 2026. The security budget shrank with hashrate in 2025-26 and had recovered part of the loss by October. The Bitcoin Improvement Proposal 110 (BIP-110) episode tested this. Enforcing nodes split off at block 961,632 and produced about eight SHA-256 blocks before stalling, and Bitcoin Knots 29.4.1 later hard-forked that chain to a BLAKE2b proof of work (bitcoin/bips #2245; TXID, 2026-10-15). The main chain was unaffected. Pool concentration is the open question: four pools found 71.3% of blocks in the week to 4 November 2026.

The 2028 halving and what to read next

The next halving arrives at block 1,050,000, expected in spring 2028, and cuts the subsidy to 1.5625 BTC (Bitcoin protocol; mempool.space, 2026-10-20). If fees stayed at their current level, they would reach only about 1.3% of block rewards after that halving (OneMiners, 2026-10-19), so hashprice will keep following bitcoin's price and machine efficiency. Until then the subsidy pays about 450 BTC a day against fees near 2.6 BTC (Firsthand Bitcoin, 2026-10-15). Fee spikes such as the 2024 Runes launch show demand can surge, but the 2025-26 data show it has not held. Readers who want the practical side of owning hardware, from power contracts to payback, will find it in the beginner guide to mining ↗.

FAQ

?Can I still mine Bitcoin at home with a laptop, a GPU or a small ASIC?

A laptop or graphics card runs hundreds of thousands to millions of times slower than a modern ASIC, so it earns less than its electricity costs. A small ASIC can pay only where power is cheap: at $33 per petahash a day, the 9.5 J/TH Antminer S23 Hydro breaks even near 14.5 cents per kilowatt-hour, and older machines need far cheaper power.

?What happens when two miners find a valid block at almost the same moment?

The network briefly splits into two competing tips. Each node follows the chain with the most accumulated work, so the next block found decides the winner, usually within about ten minutes. The losing block becomes a stale block, its reward disappears and its transactions return to the mempool unless the winning chain already includes them.

?What is the difference between hashrate, difficulty and hashprice?

Hashrate measures the total speed of all mining machines, about 1,020 EH/s on 4 November 2026. Difficulty is the number that sets the target hashes must beat, 148.36 trillion after the 29 October retarget. Hashprice is revenue per petahash per second per day, about $33 on 2 November 2026. The first two describe the network; the third describes a miner's income.

?How is Bitcoin's proof of work different from proof of stake?

Proof of work secures blocks with electricity and hardware, which an attacker must also buy. Proof of stake secures them with locked coins that validators can lose if they cheat. Ethereum moved from proof of work to proof of stake ↗ in September 2022 (ethereum.org, 2022), while Bitcoin kept proof of work and its fixed 3.125 BTC subsidy schedule.

?What is a 51% attack, and could a mining pool pull one off?

A 51% attack means controlling a majority of hashrate to reorder recent transactions or reverse the attacker's own payments. It cannot steal coins without keys or change protocol rules. The largest pool, Foundry USA, found 25.3% of blocks in the week to 4 November 2026, and the top two together 44.8%. Pool members can switch pools in minutes, which limits long-term control.

?Does a mining pool take a cut, and what do FPPS and PPLNS payouts mean?

Pools charge a fee; OCEAN lists 2% as standard. With full pay per share (FPPS), the pool pays each miner the expected value of its shares, fees included, whether or not the pool finds a block. With pay per last N shares (PPLNS), payouts depend on blocks the pool actually finds, so income varies more but pool risk is lower.

?Why do some Bitcoin blocks arrive in under a minute and others take over an hour?

Finding a block is a random event with no memory, so intervals follow an exponential spread around the 10-minute average. About 10% of blocks arrive within a minute of the previous one, and roughly one in 400 takes longer than an hour (own calculation from a 10-minute mean). The difficulty adjustment only fixes the average, not individual gaps.

References / Sources