Bitcoin Hashrate Grew in 2026. The Power Bill Didn’t.

A year ago this week, Bitcoin set a price record near $126,000 and its mining network was working harder than it ever had, roughly 1.1 zettahashes per second. Twelve months later the price sits around $84,000, according to CoinGecko, and the hashrate has not returned to that peak. On a 30-day average the network runs near 940 exahashes per second, about 12 to 15 percent below the October 2025 high, and the headline reading has spent most of 2026 moving sideways or drifting lower.

Read quickly, that looks like a growth story that stalled. Read closely, it is a growth story that changed shape. The machines doing the work today are far more capable than the ones that set the 2025 record, and the network now performs almost as much computation while drawing roughly the same electricity it did a year ago. Hashrate growth did not stop in 2026. It moved off one axis and onto another.

For most of Bitcoin’s history, three numbers climbed together: hashrate (how much computing the network does), power (how much electricity that takes), and efficiency (how much work each unit of energy buys). When all three rise in lockstep, the phrase hashrate growth is a simple idea. In 2026 they came apart. Raw hashrate flattened, electricity use went roughly flat, and efficiency kept improving. The honest way to measure Bitcoin’s growth now is not exahashes, it is work per watt. For a broader tour of what the headline number does and does not signal, we covered that in a companion piece; here the focus is narrower and more physical: where the energy went, and why the chips, not the grid connection, are now the story.

What a hashrate number actually is

Hashrate is the rate at which miners run guesses through the SHA-256 hashing function, hunting for a block header below the network’s current target. Each attempt is one hash. The network’s hashrate is the sum of every machine’s attempts per second, and it is enormous, which is why it is quoted far up a metric ladder that runs from hashes to kilo, mega, giga, tera, peta, exa and, since 2025, zetta.

The number that matters here is a crucial catch: nobody measures hashrate directly. There is no meter on the network. Hashrate is inferred from two things the blockchain does record, the current difficulty and the observed time between blocks, using the relationship hashrate is approximately difficulty times 2 to the 32nd power divided by the target block time of 600 seconds. Because block discovery is random, short windows are noisy, so every published figure is an estimate with a margin of error. That is why trackers disagree on any given day: as this was written, CoinWarz showed a spot reading above 1.1 zettahashes, shattered.io put the network near 1,001 EH/s, and Hashrate Index listed roughly 978 EH/s, while the smoother 30-day average sat near 940. Pick a window and you pick a number. For a growth story, the trend and the averages matter more than any single print.

The three curves that used to move together

Think of Bitcoin’s mining buildout as three lines on a chart. The first is raw hashrate in EH/s. The second is total power draw, the gigawatts the fleet pulls from grids worldwide. The third is efficiency, measured in joules per terahash (J/TH): the energy each machine spends to perform a trillion hashes. Lower J/TH is better.

From the first ASICs in 2013 through the 2025 peak, those three lines tracked each other closely. Operators added hashrate mainly by adding machines, and more machines meant more megawatts. Efficiency improved steadily in the background, but demand for raw computation grew faster, so the power curve climbed alongside the hashrate curve. Growth was additive and physical: pour concrete, sign a power contract, rack more rigs.

That lockstep is why, for years, you could use any one of the three curves as a proxy for the others. A chart of network power told you roughly what hashrate was doing, and both told you the industry was expanding. The 2026 split breaks that shorthand. Anyone still reading the raw EH/s line as the whole growth story is now watching the one curve that stopped moving while missing the two that did not.

In 2026 the three curves split. Raw hashrate flattened near its record and even fell for two consecutive quarters. Total power draw went roughly flat. Efficiency, however, kept dropping fast as operators swapped older rigs for a new generation of sub-10 J/TH hardware. The result is a network that does close to record computation on close to flat electricity, because each watt now buys far more work than it did even two years ago. The rest of this article takes those curves one at a time, starting with the one everyone watches and ending with the one that actually grew.

Why the raw number stalled: the megawatts went to AI

The flat raw-hashrate line is not a sign of distress. It is a choice. The same buildings, substations and power contracts that would have hosted more mining rigs are increasingly hosting artificial-intelligence and high-performance-computing hardware instead, because a megawatt leased to an AI tenant can earn far more than a megawatt spent hashing. When a public miner signs a multiyear data-center lease, those megawatts leave the hashrate curve and never show up as new EH/s.

The scale of that diversion is large. James Butterfill, head of research at CoinShares, has described 2026 as one of the most challenging periods the sector has faced, with listed miners on track to draw as much as 70 percent of their revenue from AI by the end of the year, up from roughly 30 percent, against more than $70 billion in announced AI and HPC contracts across the industry. Those are gigawatts of planned capacity that, a cycle ago, would have been earmarked for SHA-256.

The individual deals make the diversion concrete. Riot Platforms signed a 20-year agreement to host AI developer Anthropic at its Rockdale, Texas campus, a lease covering 191 megawatts and about $9.1 billion of contracted revenue, rising to as much as $16.1 billion with extensions, with capacity phasing in between December 2027 and June 2028, according to CoinDesk. Core Scientific has contracted a multi-gigawatt AI buildout with AMD, and Hut 8 holds hundreds of megawatts under AI leases. Every one of those megawatts is capacity that formally leaves the hashrate curve, which is how the companies behind mining keep growing while the EH/s line stalls.

Price and the difficulty thermostat did the rest. Through the first half of 2026 Bitcoin spent months trading below many miners’ all-in production cost, so higher-cost machines powered down and the raw curve sagged. A Q3 rally back toward the mid-$80,000s pulled some idle rigs back online, but the point stands: the raw number now reads like a profit-and-loss statement, not a measure of how fast the technology is advancing. If you want to see the advance, you have to look at the second curve.

Efficiency is the axis where growth kept going

Mining hardware has become dramatically more efficient over a decade, and 2026 marked a milestone: the first mass-production machines to break the sub-10 J/TH barrier. Bitmain’s Antminer S23 Hydro runs at about 9.5 J/TH, and Bitdeer’s SealMiner A4 Ultra Hydro is quoted near 9.45 J/TH. Compare that to the 2016-era Antminer S9 at roughly 98 J/TH, and the fleet has improved about tenfold in ten years. Even the air-cooled crown, long held by the Antminer S21 XP at 13.5 J/TH, has moved on, with the air-cooled S23 near 11 J/TH.

Why does this matter for growth? Because efficiency is the lever that lets the network perform more work without adding power. Every time an operator unplugs a fleet of S19-class rigs at 21 to 34 J/TH and plugs in S23 Hydros at 9.5, the hashrate contributed per megawatt roughly doubles or triples. On a site with a fixed power contract, that is pure hashrate growth with no new electricity. Across the whole network in 2026, that swap-out is the dominant source of new capacity, which is exactly why the EH/s curve can hold near a record while the power curve stays flat.

There is a ceiling to this, of course. Chip efficiency gains are slowing as designs approach the physical limits of advanced silicon nodes, and the leading-edge fabrication capacity that mining chipmakers depend on is spoken for by phone and AI-accelerator customers who pay more. But those are constraints on the pace of future efficiency gains, not evidence that efficiency has stopped mattering. In 2026 it mattered more than ever, because it was the only growth engine still running at full throttle.

How the efficiency curve was built

The efficiency line did not appear overnight; it is the product of a fifteen-year hardware arms race. Bitcoin’s first blocks were mined on ordinary computer processors in 2009. Graphics cards took over around 2010, field-programmable chips briefly in 2011, and then the game changed in January 2013 when the first purpose-built ASIC, Canaan’s Avalon, shipped at roughly 66 gigahashes per second. Application-specific silicon does one thing, SHA-256, and does it at an efficiency general-purpose chips cannot approach.

From there the cadence was relentless. The 2016 Antminer S9 set the template near 98 J/TH. The 2020 S19 generation roughly tripled efficiency. The S19 XP and the first WhatsMiner M-series pushed into the low 20s and high teens. The S21 family crossed under 15 J/TH, and in 2026 hydro-cooled designs finally broke the single-digit barrier. Each generation made the previous one marginal, and that planned obsolescence is the engine: the network grows not by running more machines forever but by continuously retiring the least efficient ones. What looks like a flat hashrate chart is a fast-moving escalator of hardware underneath a nearly still surface.

The math: today’s hashrate on 2016 silicon

A single calculation makes the decoupling concrete. Power draw equals hashrate times efficiency. The network runs near 940 EH/s, which is 940 million terahashes per second. Run that on the 2016-era S9 at 98 J/TH, and the fleet would need about 92 gigawatts of continuous power. The network does not draw anything close to that. By Cambridge’s index it draws roughly 16 gigawatts, so running today’s hashrate on decade-old chips would cost nearly six times the electricity the network actually uses.

Turn that around and you get the fleet’s blended efficiency. About 16 gigawatts spread across 940 million terahashes per second works out to roughly 17 J/TH network-wide. That is the honest headline growth figure for 2026: not the exahash count, but a whole-network average that has fallen into the high teens, dragged down by waves of sub-15 J/TH machines even as a long tail of older gear keeps hashing wherever power is nearly free. These are first-principles figures, not official readings, but the arithmetic is simple and the direction is unambiguous. The network keeps getting more efficient faster than it gets bigger.

For scale, 16 gigawatts of continuous draw is roughly what a mid-sized industrial economy pulls, and it is a small fraction of the hundreds of gigawatts the world’s data centers already consume. The headline that Bitcoin uses as much electricity as some countries is true, but it misses the slope that actually matters: the network added almost no new power in 2026 while its computational output held near a record. That is the opposite of the runaway-growth story the comparison is usually meant to imply.

How much electricity the network really uses

Bitcoin’s energy footprint is one of the most argued-over numbers in the industry, and the honest answer is that it depends on the method. The Cambridge Bitcoin Electricity Consumption Index models the whole network from hashrate and an assumed hardware mix, and through mid-2026 it put annualized consumption near 140 terawatt-hours, with power demand around 16 gigawatts. Other trackers run higher: Digiconomist, which leans on miner-profitability assumptions, has pointed closer to 200 TWh. The gap is methodology, not a factual dispute, and any single figure should be read as a central estimate inside a wide band.

The most carefully built figure comes from the Cambridge Centre for Alternative Finance, whose Digital Mining Industry Report, published in April 2025, surveyed 49 mining firms covering roughly 48 percent of the network and estimated annual consumption at about 138 TWh, or around 0.54 percent of global electricity. That share is the key context the raw TWh number usually lacks. Bitcoin mining is a meaningful industrial load, comparable to a mid-sized country, but it is a rounding error against the world’s total, and in 2026 that share barely moved even as hashrate recovered, precisely because efficiency absorbed the growth.

The energy mix and the per-transaction myth

If efficiency changes how much energy the network uses, the energy mix changes what that energy is. The Cambridge report found that sustainable sources made up 52.4 percent of mining’s electricity, split between about 42.6 percent renewables and 9.8 percent nuclear, with natural gas the single largest individual source at 38.2 percent. That is a materially cleaner mix than the industry carried five years ago, and it reflects where new capacity is actually being built: hydro-rich Paraguay, flared-gas fields, curtailed wind and solar, and grids with surplus baseload.

The same Cambridge work put mining’s attributable greenhouse-gas emissions near 39.8 million tonnes of CO2-equivalent, a figure that moves with the fuel mix more than with the hashrate. As the sustainable share grows and gas displaces coal, emissions per unit of security computation keep falling, another curve bending in the same direction as efficiency.

It also undercuts the most common way the energy debate is framed. The familiar energy-per-transaction statistic, which divides the whole network’s power by the number of on-chain transactions, treats mining as if it existed to process payments. It does not. Miners are paid to secure the chain and win the block subsidy; the energy they spend is almost entirely independent of how many transactions sit in a block. ESG researcher Daniel Batten, who built a set of dynamic Bitcoin energy charts with analyst Willy Woo, has argued exactly this, noting that consumption is largely independent of transaction volume and that the sustainable share now sits above half. Batten has gone further, rebutting nine recurring criticisms of mining’s energy use with reference to peer-reviewed work. You do not have to accept every claim to take the structural point: a network whose energy is flat while its security computation holds near a record is getting cleaner and more efficient per unit of work at the same time.

Difficulty: the thermostat that reprices the growth

None of these curves move freely, because Bitcoin has a built-in governor. Every 2,016 blocks, roughly every two weeks, the protocol adjusts mining difficulty to steer the average block time back toward ten minutes. If the fleet got faster, difficulty rises; if hashrate left, it falls. Difficulty is the thermostat that keeps block production steady no matter how much computing power shows up.

The adjustment itself is mechanical. The protocol takes the time the last 2,016 blocks actually needed and compares it to the 20,160 minutes they should have taken at ten minutes each; if blocks came in 5 percent fast, difficulty rises about 5 percent, and vice versa, with a hard cap of a fourfold move in either direction that has never been hit. Hashrate is then backed out from difficulty, which is exactly why the number everyone quotes is a derived estimate rather than a direct reading.

That thermostat is why efficiency growth does not simply make mining easy money. When operators add efficient hashrate, difficulty climbs to match, and each machine’s share of the fixed 3.125 BTC-per-block reward shrinks. In 2026 the adjustments have been a tug-of-war: difficulty fell through the summer as rigs idled, then climbed back as the Q3 price rally pulled capacity online. As of early October it stood at about 132.72 trillion after a near-flat retarget on 3 October, with the next adjustment due around 17 October, per CoinWarz. It remains roughly 15 percent below the record near 156 trillion set in late 2025. The lesson for anyone reading a hashrate chart: the thermostat competes away the windfall from both price and efficiency, so the network tends back toward the marginal miner’s cost no matter how good the chips get. We walked through that feedback loop and the long-run security question in a dedicated look at difficulty and the 2028 halving.

What efficiency does, and does not do, for security

Here is where the decoupling gets misread most often. A common assumption is that a more efficient network is a more secure one. It is not, at least not directly. Bitcoin’s security budget is the total revenue miners earn, roughly hashprice times hashrate, or equivalently the block subsidy plus fees multiplied by the BTC price. At about 450 BTC paid out per day and a price near $84,000, that budget runs on the order of $14 billion a year. An attacker trying to rewrite history has to out-spend that floor.

Efficiency cuts both ways in that equation. Better chips lower the cost to defend the network, delivering more hashes per watt, but they also lower the cost to attack it, because an attacker buys the same efficient hardware. The dollar size of the security budget is set by price and fees, not by J/TH. So the right way to state 2026 is that the network is buying the same security for less energy, which is a sustainability win, not an automatic security upgrade. The security number still tracks price.

At current levels the floor is high. Duke finance professor Campbell Harvey, who models attack economics, has estimated that a majority-hashrate attack paired with a large short position in offshore derivatives could cost around $8 billion, or roughly 50 basis points of Bitcoin’s value. Whether that is comforting or alarming depends on your priors, but it is a direct function of the dollar budget, which is why the same question looks very different for proof-of-stake chains that price security in bonded capital and yield rather than energy, a contrast we worked through in our look at validator economics.

It is worth remembering what that budget has bought. Bitcoin’s main chain has never suffered a successful 51 percent reorganization in more than fifteen years, while smaller proof-of-work coins with thinner budgets, Bitcoin Gold and Ethereum Classic among them, have been reorganized more than once. The security spend is not theoretical insurance; it is the observable reason the largest chain has stayed immutable while cheaper imitations did not. Efficiency lets that protection be bought with less energy, but it is the dollar budget, underwritten by price, that does the protecting.

The 2028 halving makes the watt matter more

Everything about efficiency becomes sharper in the spring of 2028, when the block subsidy halves from 3.125 BTC to 1.5625 BTC around block 1,050,000. At a constant price, that cuts the subsidy portion of miner revenue in half overnight. The thermostat will adjust, less-efficient capacity will switch off, and difficulty will fall to rebalance, but the revenue earned per unit of work drops regardless. After 2028, the gap between a 9.5 J/TH hydro fleet on sub-five-cent power and a 25 J/TH fleet on grid power is not a margin difference, it is survival versus shutdown.

That is why efficiency growth is really survival growth. Fred Thiel, chief executive of MARA, one of the largest listed miners, has put the endgame bluntly, calling mining a zero-sum game where the floor is your energy cost and predicting that by 2028 a miner will either be a power generator, be owned by one, or be partnered with one. The transaction-fee market is supposed to make up the difference as the subsidy shrinks, but fees have run well under 1 percent of the block reward for much of 2026, so for now the burden falls on price and on the efficiency curve. The miners who make it to the next halving will be the ones who pushed their blended J/TH down hardest and locked in the cheapest power.

Where the efficient megawatts live

Efficiency is also quietly redrawing the map of where hashrate sits. Hashrate Index’s Q3 2026 global heatmap put the network near 940 EH/s, down about 6 percent on the quarter in its second straight quarterly decline, with the United States still dominant at roughly 36.7 percent, Russia near 17 percent and China around 12 percent, keeping the top three at about two-thirds of the total. The more telling detail is at the bottom of the table: Norway entered the top ten by simply holding flat, while Kazakhstan, once a magnet for cheap-power miners, dropped out.

That reshuffle tracks the efficiency story. When chips were power-hungry, the migration driver was the cheapest kilowatt-hour anywhere on earth, which sent fleets to frontier grids that later proved unstable. As machines get efficient and AI competes for the same sites and power, the driver shifts toward stable jurisdictions that can also host high-value compute and sell flexibility back to the grid. Miners have become a gigawatt-scale interruptible load that operators pay to power down, a role we explored in how the hashrate found a second job balancing the grid. Clean, stable, dispatchable power is winning the migration that cheap-but-fragile power used to.

Hashprice, margins and the marginal machine

The number that ties efficiency to profit is hashprice, the expected daily revenue per unit of hashrate. In early October it sat near $40 per petahash per day on Hashrate Index data, up roughly 45 percent from the sub-$28 low of late June, lifted by the price recovery rather than by anything the network did. Hashprice is the same for every miner on earth; what differs is the power bill, and the power bill is set by efficiency.

The arithmetic is unforgiving. Break-even power price is roughly hashprice divided by 24 times efficiency in J/TH. At $40 hashprice, an S23 Hydro at 9.5 J/TH breaks even near 17.5 cents per kilowatt-hour, an S21 XP at 13.5 clears at about 12 cents, an S19 XP at 21.5 needs power under about 8 cents, and a 29.5 J/TH relic from the last cycle needs roughly 5.6-cent power just to avoid losing money. The network’s economics are set at the margin, by the least efficient machine still worth running, and efficiency growth is the force that keeps pushing that marginal machine toward the exit. Read that way, hashrate growth and margin pressure are the same phenomenon seen from two sides.

A quick worked example shows why. One petahash per second of an S23 Hydro at 9.5 J/TH draws about 9.5 kilowatts, roughly 228 kilowatt-hours a day, and earns about $40 at current hashprice. At 5-cent power the daily electricity bill is near $11 and the machine clears almost $29; at 12-cent power it still nets around $13; at its 17.5-cent breakeven it earns nothing. Put the same petahash on a 29.5 J/TH veteran and it draws three times the power for the identical $40 of revenue, so it needs sub-6-cent electricity just to stay above water. The hashprice is the same in every case. Efficiency and the power contract are the whole difference between a profitable miner and a bankrupt one.

How to read hashrate growth from here

If the EH/s chart no longer captures the growth, what should you watch instead? Four things. First, blended network efficiency, the J/TH implied by dividing power draw by hashrate, which is the cleanest single measure of how fast the fleet is modernizing. Second, total energy use and its share of global electricity, because that is where the sustainability debate is actually settled. Third, AI and HPC lease announcements, which tell you how many future megawatts are leaving the hashrate curve. Fourth, the near-term levers on price and therefore on the security budget: the 17 October difficulty retarget and the Federal Reserve’s 27 to 28 October meeting both land inside the window this was written in.

Regulators, for their part, treat mining as the industrial activity it is. The Securities and Exchange Commission’s Division of Corporation Finance said in March 2025 that proof-of-work mining, whether solo or pooled, does not involve the offer or sale of securities, a stance that has held through the quieter enforcement environment we described in our review of SEC crypto enforcement in 2026. Mining sits with energy policy and chip supply chains, not securities law, which is the correct address for an industry whose growth now shows up in joules rather than headlines.

So the one-line answer to the question this article opened with: yes, Bitcoin’s hashrate grew in 2026, and no, the power bill did not follow. The growth left the megawatt axis and moved to the efficiency axis, where it is harder to see on a chart but easier to defend in an argument about energy. The network got faster without getting hungrier, and that, not the flat exahash line, is the real 2026 story.

Frequently Asked Questions

Why is Bitcoin’s hashrate not at an all-time high in 2026?

The 2026 plateau is voluntary rather than forced. Miners redirected megawatts that would have hosted new rigs toward AI and high-performance computing, which pay more per megawatt, and a first-half price slump idled higher-cost machines. The computing power did not vanish; much of the growth simply moved into more efficient hardware instead of new capacity.

How much electricity does Bitcoin mining use in 2026?

Cambridge’s index puts annualized consumption near 140 terawatt-hours, roughly 0.5 percent of global electricity, though other trackers estimate closer to 200 terawatt-hours using different assumptions. Notably, that figure stayed roughly flat through 2026 even as hashrate recovered, because efficiency gains absorbed the growth.

What is the most efficient Bitcoin miner in 2026?

The most efficient machines in 2026 are hydro-cooled units that broke the sub-10 joules-per-terahash barrier, led by the Antminer S23 Hydro at about 9.5 J/TH and the Bitdeer SealMiner A4 Ultra Hydro near 9.45 J/TH. For comparison, the 2016-era Antminer S9 used about 98 J/TH, roughly ten times more energy for the same work.

Does a higher hashrate make Bitcoin more secure?

Higher hashrate raises the raw computing power an attacker must match, but Bitcoin’s dollar security budget is set by the block reward and the BTC price, not by chip efficiency. At current levels miners earn on the order of $14 billion a year, and more efficient chips lower the cost to attack the network just as much as the cost to defend it.

What does joules per terahash (J/TH) mean?

Joules per terahash measures how much energy a miner spends to perform one trillion hashes, so a lower number means a more efficient machine. Dividing the network’s power draw by its hashrate implies a blended average near 17 J/TH in 2026, down from around 98 for a flagship machine a decade ago. It is the cleanest single gauge of how fast the fleet is modernizing.

By Marcus Okafor, mining and energy desk, HOGE Wire.