Have you ever watched a number that is supposed to look unstoppable suddenly lose its nerve? That is the feeling this week around Bitcoin’s computing power. The seven-day average hashrate slipped to about 915.8 EH/s, the softest print in three weeks, while wallets tied to miners shed another 1,530 BTC. On paper that sounds like a simple retreat. In practice it is a tangle of power bills, treasury choices, weather at distant dams, and a growing temptation to rent the same electricity to artificial intelligence instead of chasing block rewards.
Why This Hashrate Dip Matters More Than The Headline
I keep coming back to a basic point. Hashrate is not a vibe. It is the raw computing muscle that tries to find the next block, lock in transactions, and make rewriting history expensive. When that muscle flexes, the network looks crowded and hard to bully. When it eases, people start asking whether miners are tired, broke, or simply shopping for a better customer than Bitcoin.
The latest seven-day moving average landed near 915,844,520 TH/s on September 26. That is roughly 34.86 million TH/s lighter than a week earlier and the lowest reading since around September 3. Daily snapshots jumped around as they always do. One dashboard put the network near 954 EH/s on September 25 after something closer to 984 EH/s the day before. Earlier in the month, a few sessions even poked above one zettahash. Then late September cooled off again and several windows stayed under that round number.
Does a three-week low equal a crisis? No. Bitcoin still produced blocks on schedule. Pools kept swapping share. The difficulty retarget still cares about a full 2,016-block window, not one tidy weekly average. Still, the direction is hard to shrug off if you have been watching the longer slide from the late-2025 peak. Some operators now talk about a hashrate bear market, with the drop from that peak estimated around 22% to 24%. I find that phrase a little dramatic, but the arithmetic is not imaginary.
What Hashrate Actually Measures When The Chart Wiggles
Hashrate is the combined guesswork miners throw at the proof-of-work puzzle. More machines, more joules, more attempts per second. A higher figure usually means more hardware is online and competing. A lower figure can mean machines went dark, power got rationed, pools reported late, or luck in block timing made the estimate look thinner than the physical fleet.
That last part matters. Short-term hashrate is inferred from how fast blocks arrive and from pool reports. If blocks bunch up, the estimate can look heroic for a day. If they stretch out, the same fleet can look weak. The seven-day average is meant to sand off that noise. Even then, it is a smoothed guess, not a census of every rack in Texas, Kazakhstan, or the Ethiopian highlands.
- Daily readings can swing tens of EH/s without a matching change in installed machines.
- Power curtailment can pull whole sites offline for hours and then bring them back.
- Difficulty only reacts after a full adjustment period, so price and hash can diverge for weeks.
- Pool mix shifts can change reported share even when total work is stable.
In my experience, the useful question is not “did the line go down.” It is “did the line go down for a reason that persists.” Weather, policy, and a pivot toward high-performance computing can persist. A quiet Sunday with slow blocks usually does not.
Miner Reserves Fell By 1,530 BTC In Seven Days
While hashpower cooled, coins left miner-linked wallets. Tracked reserves sat near 1,192,766 BTC on September 26, down 1,530 BTC from seven days earlier. That is not a collapse of the entire miner stash. It is a noticeable weekly leak from a pile that still sits around 1.1928 million BTC.
A falling reserve means coins moved out of addresses that on-chain analysts associate with miners and pools. It does not automatically mean every coin hit a spot exchange and smashed the bid. Transfers can go to custodians, lenders, collateral wallets, over-the-counter desks, or simply a new internal treasury address that the model has not labeled yet. I wish the market would treat that distinction as gospel. It rarely does.
A thinner miner wallet is a clue, not a confession. The destination of those coins decides whether the move is selling pressure or just housekeeping.
Earlier in September the flow even flipped. Holdings were reported near 1.1919 million BTC on September 5, up 261 BTC from the prior week. So this is not a one-way conveyor belt. Miners accumulate when cash flow is comfortable and distribute when bills, expansions, or buybacks of their own equity demand fiat.
Public Miners Are Not Running The Same Playbook
Listed operators make the split visible because they file numbers. One large public miner produced 593 BTC in August and still sold 821 BTC in the same month, finishing with 13,703 BTC on the balance sheet and an average operating hashrate around 38.3 EH/s. That is a firm that mines and also funds the business by letting inventory go.
Another operator shut Bitcoin mining at a Michigan site on September 1 to ready the building for an AI computing contract. Holdings there dropped roughly 79% from 1,006 BTC in July to around 215 BTC. That is a different story altogether. The machines are not merely paused. The building is being recast as a different kind of factory.
I have found that readers often mash these cases together. They are not the same. Selling newly mined coins to cover power and payroll is ordinary. Liquidating most of a treasury while converting a hall to inference clusters is a strategy change. Both can lower reported miner reserves. Only one tells you the company is leaving the hash wars.
| Signal | What it can mean | What it does not prove |
| Seven-day hashrate down | Less work arriving on chain this week | Every farm is permanently offline |
| Miner reserves down | Coins left labeled miner wallets | Every coin was dumped on an exchange |
| Puell Multiple up | Issuance value beat its one-year average | Every miner is suddenly profitable |
| Site converted to AI | Power and capex found a new buyer | The whole network is abandoning Bitcoin |
The Puell Multiple Rose Even As Hash And Reserves Softened
Here is the wrinkle that keeps the week from being a simple risk-off story. The Puell Multiple climbed by 0.24 over seven days to 1.13 on September 26 in one widely cited set of figures. Another public dashboard put it closer to 1.01 the same day and called the zone normal, with daily issuance near 450 BTC and daily miner revenue around $37.8 million.
Those two readings do not have to fight. Timing, price feeds, and smoothing can split a metric that looks simple. The idea behind Puell is straightforward. Compare the dollar value of newly issued bitcoin today with the 365-day average of that same daily issuance value. Above one, today’s coinbase is richer than the typical day over the past year. Below one, miners are issuing into a thinner paycheck than they are used to.
A print above one does not mean every site is printing cash. Electricity, debt service, hosting fees, and the age of the ASIC fleet never show up in Puell. A hydro plant with cheap night power and a leveraged public miner running older boxes can live in opposite worlds on the same reading. Perhaps the most interesting aspect is how often the market treats one oscillator as a verdict on an entire industry.
Puell at a glance: Above 1 = issuance revenue richer than the one-year norm Near 1 = roughly average conditions by that yardstick Below 1 = thinner issuance dollars than miners are used to Missing = site-level power cost, debt, and uptime
Block Subsidy, Halving Math, And Why Revenue Still Feels Tight
The current block subsidy is still 3.125 BTC after the April 2024 cut. Fees sit on top of that, but they are moody. The next programmed cut is expected around 2028, when the subsidy falls to 1.5625 BTC. That calendar is not a secret and every serious operator has it in a spreadsheet.
What changes the mood is the mix of price, difficulty, and alternative uses for megawatts. If bitcoin’s dollar price holds up, a Puell reading can drift higher even while some machines unplug. If price sags and difficulty stays elevated from earlier hash growth, weaker sites bleed first. The late-2025 peak in network work set a high bar. Coming off that bar by roughly a fifth looks, to some executives, like the first real hash winter since the last expansion cycle.
I’ve found that people outside mining underestimate how local this business is. A national grid scare, a reservoir that misses its rainy season, or a town that renegotiates an industrial tariff can do more in a month than a global narrative does in a quarter.
Power Politics: When A Dam Decides Who Gets To Mine
September made that local reality blunt in Ethiopia. The utility cut electricity delivered to Bitcoin miners to 23% of contracted volumes after reservoir inflows dropped about 20%. Mining had been using nearly a third of the country’s electricity and generating about 35% of the utility’s revenue in the prior fiscal year. That is an awkward pair of facts. The sector is both a customer the grid likes on paper and a load the grid cannot always feed when water runs short.
Officials plan to reassess in October. Further cuts will depend on reservoirs and domestic demand. From a network-wide view, one country’s rationing will not freeze Bitcoin. From a miner’s view, it can strand containers, wreck utilization, and force sales of coins that were meant to sit as treasury. Hashrate is global. Cash flow is painfully local.
- Watch contracted power versus delivered power, not just nameplate megawatts.
- Track rainy-season and reservoir headlines in hydro-heavy mining regions.
- Separate temporary curtailment from a site that is being rewired for other compute.
- Expect difficulty to lag the physical change by one or two adjustment windows.
The AI Detour: Same Warehouses, Different Workload
A growing share of the conversation is not about bitcoin at all. It is about whether the same substations, land, and cooling loops should serve model training and inference. Some mining chiefs have said openly that capital and power are being steered toward AI. That helps explain why a drop from the late-2025 peak can look structural rather than seasonal.
The economics are crude but sticky. If a cluster of GPUs or specialized accelerators can pay more per megawatt-hour than an aging ASIC row, the ASIC row loses the argument. Bitcoin still has a built-in buyer of last resort: the protocol pays whoever finds the block. AI has corporate buyers who sign multi-year offtake deals. Those contracts can look safer to a board than a coin that prints a new high or a new scare every other month.
That does not make Bitcoin mining obsolete. It makes the marginal watt more contested. Efficient new machines in cheap-power pockets will keep running. Older fleets in tight grids will face a fork in the road. Keep hashing, host someone else’s compute, or sell the site. The Michigan shutdown is one data point in that fork, not a eulogy for proof of work.
When two industries bid for the same electron, the protocol with the weaker bid does not die. It just gets pickier about where it lives.
Network Health Versus Miner Mood
It is easy to confuse the two. Network health is about whether blocks keep arriving, whether fees clear, and whether an attacker would need an absurd amount of hardware and energy to rewrite recent history. Miner mood is about margins, equity prices, and whether the CFO wants more coins or more cash.
This week the network still looked functional. Mempool views on September 26 showed blocks landing and pool shares rotating through the day even while short-term hash estimates bobbed. That is the system working as designed. Difficulty will eventually shrug if the weaker weekly average persists across a full epoch. If hash recovers, the next retarget may barely blink.
Miner mood is more mixed. Some firms still treat bitcoin as a core treasury asset and only sell what operations require. Others have become habitual sellers. A few are leaving the category. If you only watch the reserve number, you will flatten those differences into a single scare headline. I would rather keep them separate.
How To Read The Next Few Weeks Without Overfitting
A three-week low is a data point. It becomes a trend if October’s seven-day average fails to reclaim the early-September range and if miner wallets keep leaking at a similar pace. It fades if hash snaps back after maintenance, weather, or a difficulty dip that brings weaker machines online again.
I would watch four threads at once rather than worship one chart.
- Seven-day and thirty-day hash, not a single heroic or ugly daily print.
- Miner reserve changes together with known public-company sales, so unlabeled outflows do not get over-interpreted.
- Puell and realized issuance dollars against spot price, because a higher multiple can arrive from price alone.
- Physical headlines on curtailment, AI conversions, and new efficient deployments.
If those four disagree, believe the physical headlines first. Charts follow warehouses more often than warehouses follow charts.
A Practical Frame For Investors Who Are Not Miners
You do not need a container of machines to care about this. Hashrate is one of the few industrial stats attached to a monetary asset. When it rises for months, the security budget is thick and the industry is expanding. When it stalls, either the asset is less profitable to mint or the industry found another tenant for its power.
A modest dip with rising Puell can mean price is doing more work than hardware. A dip with falling reserves and visible AI conversions can mean supply from miners will stay noisy even if the protocol itself is fine. Neither setup is a trading signal by itself. Both are context. Context is what keeps people from turning every weekly print into a prophecy.
In my view, the healthy stance is slightly boring. Assume Bitcoin will keep producing blocks. Assume some miners will sell because they run factories, not museums. Assume a slice of power will keep leaking toward other compute until the bitcoin bid wins again in specific regions. Then update those assumptions when the physical world actually changes.
The Quiet Work The Difficulty Algorithm Still Does
People love a crisis narrative. The protocol prefers arithmetic. About every 2,016 blocks, difficulty moves so that block times drift back toward the ten-minute target. If late September’s softer work persists through a full window, the next retarget eases the puzzle. Easier puzzles can invite sidelined machines back, especially if power prices cooperate. If hash was only noisy, the retarget barely moves and everyone forgets the weekly average.
That feedback loop is why a three-week low is interesting and still not fatal. The network is allowed to breathe. Miners are allowed to leave. The remaining set inherits a slightly easier contest if enough of them do. It is an unsentimental design, which is part of why it has lasted.
Putting The Week In One Place Without The Panic
So where does that leave the story? Bitcoin’s seven-day average hashrate eased to roughly 915.8 EH/s, the weakest mark in about three weeks. Miner-linked wallets were 1,530 BTC lighter. Daily estimates spent several late-September sessions under one zettahash after flirting with that threshold earlier in the month. The Puell Multiple, depending on the feed, sat a little above one. Public miners kept splitting between those who sell inventory to fund operations and those who are walking away from mining halls altogether. One major hydro region rationed power hard enough to make utilization a political number, not just an engineering one.
None of that requires you to declare a new era every Sunday. It does require you to stop treating hashrate as a straight-line badge of health and miner reserves as a pure sell ticker. The machines are competing with other machines now, and not all of those other machines speak Bitcoin. That contest will shape the next year of hash more than any single weekly print.
If the October review in constrained grids restores power, some of this dip simply vanishes. If more sites sign compute contracts that pay better than block rewards, the late-2025 peak may stand as a local high for a while. I would rather watch the substations than the slogans. The substations, inconveniently, are where the truth still lives.
And if you only remember one thing from this stretch of numbers, make it this. A network can look quieter on a seven-day chart and still be doing its job, while the people who plug in the racks quietly renegotiate what those racks are for. That renegotiation is the real plot. The 915.8 EH/s print is just the latest line in the script.
