Five Miners Cut Bitcoin Output 47%. The Network Added 105 EH/s Anyway.

On August 6, the seven-day hash rate print on the Bitcoin network dropped below 900 EH/s. By the third week of October it was back near 1,005 EH/s. That is an 11.7% recovery, and roughly half of it landed inside a three-week window in early October.

Now put a second number next to it. The five listed miners most publicly committed to converting their power contracts into AI and HPC capacity produced 13,263 BTC across the trailing year. In the year just closed, the same cohort produced 7,075 BTC. A 46.6% collapse in output.

Read those two paragraphs together and you get the headline every outlet ran: miners are fleeing to AI, and yet the hash rate is surging. The framing implies a paradox. It implies that something unexplained is happening, that the network is somehow defying the departure of its most visible participants.

There is no paradox. There is a measurement problem, a fungibility property, and a capital allocation decision that any rational operator would have made eighteen months earlier. The only genuinely surprising thing in this story is how badly the arithmetic has been reported.

I have been through three of these cycles with my own capital on the line. In 2022 I watched Terra's incentive structure fail from the inside — not from the price chart, from the mechanism — and shorted LUNA into the cascade because the failure was legible in the code weeks before it was legible in the market. What I learned that month is the thing I keep relearning: when a headline number and a mechanical number disagree, the mechanical number is right and the headline is being computed wrong.

So before I analyze anything, I audit the data chain. If the inputs are inconsistent, every conclusion downstream is garbage. That is not caution. That is the job.

The arithmetic nobody ran

The source material for this piece contains a clean number set. I recomputed all of it. Here is what survives.

| Check | Calculation | Result | Consistent | |---|---|---|---| | Implied hash ATH | 900 ÷ (1 − 0.18) | ≈ 1,098 EH/s | — | | Post-bounce hash rate | 900 × 1.117 | ≈ 1,005 EH/s | Yes | | Distance from ATH | 1,005 ÷ 1,098 = 91.5% | 8.5% gap | Yes, self-consistent | | Net new hash | 1,005 − 900 | +105 EH/s | Yes | | Five-miner production decline | (13,263 − 7,075) ÷ 13,263 | −46.6% | Yes, rounds to "47%" | | Five-miner share of blocks, this year | 7,075 ÷ (3.125 × 144 × 365) | ≈ 4.3% | — | | Five-miner share of blocks, last year | 13,263 ÷ 164,250 | ≈ 8.1% | — | | Exited hash as share of network | 35 ÷ 1,005 | 3.5% | — | | Announced AI contracts vs annual mining revenue | $50,000M ÷ (7,075 × $80k ≈ $566M) | ≈ 88× | — |

The chain holds. The internal consistency is real. But two claims in the underlying narrative do not survive scrutiny, and both of them are load-bearing for the popular interpretation.

Flag one: the "deepest drawdown since the 2021 China ban" claim is almost certainly an artifact of smoothing. The 2021 mining ban knocked more than 50% off measured hash rate inside a few weeks. The post-halving period in 2024 produced multiple double-digit retracements. An 18% drawdown being described as the deepest since 2021 only works if you are looking at a 30-day moving average or a monthly mean, not at the raw estimate. The superlative is doing rhetorical work that the data does not support. I do not accept it.

Flag two: +105 EH/s in roughly ten weeks is physically implausible as new manufacturing. Run the units. Gross additions were about 140 EH/s (140 net of the 35 EH/s that exited). At Bitmain's S21 XP spec of 270 TH/s, 140 EH/s requires roughly 518,000 machines. At a more typical 200 TH/s fleet average, it requires over 700,000. Global ASIC annual shipments do not support that volume landing in a single quarter, let alone ten weeks, and the machines would have to be delivered, racked, powered, and hashed simultaneously.

There are three explanations that actually fit, and none of them is "new machines." First, idle fleets coming back online — rigs that were curtailment-stranded or shut down post-halving when the marginal joule stopped paying. Second, overclocking of existing fleets, which raises reported hash per machine without adding a single unit. Third, and most important, the hash rate figure itself is a smoothed estimate rather than a measurement.

That third point is the one that reorganizes the entire story. Hash rate is not a physical quantity you can weigh. It is an inference derived from block arrival times against a difficulty target that only updates every 2,016 blocks. During a migration window, that inference goes stale in a specific and predictable direction. I will come back to it.

Methodological premise for everything that follows: hash rate ≠ energy consumption ≠ new capital expenditure. Collapse those three into one variable and you will overstate the significance of the miner migration by an order of magnitude.

Context: what is actually being repriced

To understand why five miners walked away from 47% of their production and the network did not blink, you have to be precise about what a mining company owns.

A Bitcoin miner owns four things, in descending order of scarcity: an executed interconnection agreement and power purchase agreement, a physical shell with cooling and switchgear, a fleet of application-specific integrated circuits, and a treasury of BTC. The ASICs are the least scarce of the four. They are manufactured at scale, they depreciate brutally, and their resale market is thin. The interconnect is the scarcest. In most North American and European grids, the queue to connect a hundred megawatts runs three to seven years, and the queue is not first-come-first-served — it is prioritized by creditworthiness, by location, and increasingly by whether your load can be curtailed on demand.

That asymmetry is the whole story. A mining company that converts a 300 MW interconnect from proof-of-work hashing to HPC colocation is not abandoning its business. It is taking the single asset it owns that has a moat and pointing it at a customer with a twelve-year contract and a credit rating.

The demand side is not speculative anymore. Announced AI and HPC contracts across the sector exceed $50 billion. IREN's agreement with Microsoft alone is reported at $9.7 billion — a figure that, I want to note, has been misreported in at least one version of this story as $970 million, a decimal error that inverts the entire scale of the comparison. $9.7 billion is larger than the total market capitalization of most mid-cap Bitcoin miners. That is not a contract. That is a re-rating event.

Against that, look at what mining actually pays. A miner's revenue is block subsidy plus fees. The subsidy is 3.125 BTC per block post-halving. Fees, in normal conditions, run between 1% and 5% of total revenue. Which means 95% or more of a miner's revenue is a function of a price it does not control, produced by an asset that is perfectly substitutable, sold into a market where its marginal cost is the only differentiator.

That is not a business. That is a leveraged long position on BTC with a physical plant attached and a fixed electricity bill.

I have run this math on my own book. In 2020 I staked $15,000 into Synthetix, calculated the collateralization ratio by hand on a local Ethereum node rather than trusting a dashboard, and watched the DeFi Summer liquidity fragmentation create a window where the same asset traded at meaningfully different prices on Uniswap and Sushiswap. I captured 42% in three weeks. The lesson was not that arbitrage works. The lesson was that the edge exists precisely where two markets price the same underlying with different mechanisms. Mining and HPC colocation are the same underlying — a grid connection — priced by two different mechanisms. The spread is enormous. The migration is the arbitrage.

Core I: fungibility is a property, not a coincidence

The most common misreading of this story is that miners leaving represents a threat to Bitcoin's security budget. Let me kill that with arithmetic before it wastes anyone's time.

The exiting hash is 35 EH/s. The network is running at roughly 1,005 EH/s. That is 3.5% of total. The five AI-transition miners combined represent about 4.3% of blocks mined this year, down from 8.1% last year. For their departure to constitute a security concern, you would need to argue that 3.5% of hash rate migrating from one operator to another represents a structural change in the cost of a 51% attack. It does not. It represents a change in who is paying the electricity bill.

Here is the property that makes this true, and it is not an accident of market conditions. Proof-of-work networks do not identify who is hashing. They identify whether valid work was performed. The protocol has no field for operator identity, no permissioning layer, no allowlist. When 35 EH/s disconnects, the remaining network produces blocks slightly slower than the difficulty target implies, and 2,016 blocks later the difficulty adjusts downward to restore the ten-minute cadence. At that point, any operator with a marginal cost below the new revenue-per-hash can profitably plug in. The system is self-healing by construction.

This is why the "miners are abandoning Bitcoin" narrative fails at the level of mechanism. There is no abandonment available. There is only a price signal that gets transmitted through difficulty and picked up by whoever has the cheapest electrons.

And the pick-up happened fast. The fill came from three sources: private miners deploying new-generation rigs, sovereign-adjacent entities with access to subsidized power, and non-US operators outside the reporting perimeter. That composition matters more than the volume, and I will get to why.

What I want to establish here is the strategic consequence for the miners themselves. Because hash rate is fungible, a mining company has zero pricing power over its own product. It cannot differentiate. It cannot build a moat. It cannot raise price. It can only lower cost, and cost reduction in a commodity with a fixed reward schedule is a race to the bottom that terminates, inevitably, in utility economics — thin margins, regulated returns, consolidation.

The terminal state of industrial mining is a regulated utility. The miners who understand this are leaving before the terminal state arrives. The ones who do not are the ones whose production is up 16% year over year while their unit economics compress. I do not say that as criticism. I say it as a description of a trap that looks like growth from the inside.

Core II: the measurement is the story

Now the part that almost nobody has written, and the part I care most about.

When a report tells you hash rate fell 18% and then recovered 11.7%, the first question a competent analyst asks is: measured how? Instantaneous? Seven-day moving average? Thirty-day? Difficulty-implied? These are not equivalent, and during a miner migration window they diverge violently.

The reason is structural. Difficulty only adjusts every 2,016 blocks — roughly two weeks at target cadence, longer when blocks are slow. Hash rate estimates are derived by taking observed block times and inverting them against the current difficulty. If a large cohort of hash disconnects, blocks slow down. But the difficulty target has not yet adjusted. So the estimate computed from those slow blocks understates the true hash rate, because the estimator is dividing an unchanged difficulty by an inflated inter-block time.

Now reverse it. Hash comes back online. Blocks speed up. The difficulty target is still the old, low one — the downward adjustment has not yet landed or has just landed. The estimator divides the old difficulty by a compressed inter-block time and reports a hash rate spike.

During a migration window, hash rate readings can rise for accounting reasons before any new capacity exists. That is not a conspiracy. It is what happens when you infer a physical quantity from a lagged control variable and then report the inference as if it were a measurement.

I ran a Freqtrade bot through the first quarter of 2025 — 1,200 trades, 28% net after fees, and three manual overrides where the sentiment layer hallucinated a signal that the order book flatly contradicted. The single most useful thing that project taught me had nothing to do with returns. It taught me that any metric computed from smoothed inputs will confirm whatever hypothesis you brought to it. I had to rebuild my backtest three times before the results stopped flattering me. Hash rate is the same problem at network scale.

So when a headline says the network recovered 11.7% in ten weeks, half of it in a three-week window, the honest reading is: some of that is real re-energization, some is fleet overclocking, some is a stale difficulty target being divided into a fresh block cadence, and some is genuine new deployment. The published number does not decompose those. It cannot.

This has an immediate, practical implication for anyone trading the sector. The "hash rate surging while miners flee" headline is being read as evidence of network strength. It is more accurately evidence that the network's most-cited health metric is noisy exactly when it matters most. The chart is a map, not the territory. And this particular map has a two-week lag baked into its grid lines.

There is a second-order effect worth flagging. Fleet refresh — replacing 200 TH/s machines with 270 TH/s machines at identical power draw — raises hash rate substantially with zero increase in energy consumption. If a meaningful share of the gross 140 EH/s additions came from fleet refresh rather than new deployment, then the implied new electrical load collapses. At 15 J/TH, 140 EH/s implies about 2.1 GW of gross new capacity. That number is not credible. It is roughly the scale of a mid-sized national grid addition, commissioned in a quarter, in a market where transformer lead times run 100 weeks.

So the physical reality is almost certainly: modest new capacity, substantial re-energization of curtailed fleets, aggressive overclocking, and a measurement artifact. The narrative reality is "hash rate surging." Those are not the same claim.

Core III: the tokenomics of an orderly exit

Here is the conclusion I think is genuinely underrated, and I will state it plainly because the mechanism is clean.

Five miners moved from 8.1% of network block production to 4.3%. That is a halving of their share in a single year. And critically, this cohort's BTC was the most price-insensitive, most structurally forced selling in the entire market. A listed miner with a fixed power bill and a quarterly reporting obligation does not sell BTC because it wants to. It sells because payroll clears on the 15th and the utility invoices on the 30th. That is supply that does not respond to price. It responds to the calendar.

Removing 6,188 BTC of annual production from that category is a structural reduction in forced supply. That is real. It is the cleanest bullish mechanism in this entire story, and the source material gives it one sentence.

But I am not going to let the bull case run unchecked, because the magnitude does not support the enthusiasm. 7,075 BTC per year is roughly 19.4 BTC per day. Global daily BTC spot volume across venues is measured in the tens of thousands of coins. Nineteen coins a day is not a liquidity event. It is a sentiment event. The impact is on marginal pricing and narrative, not on depth. Anyone modeling this as a supply shock is modeling the wrong variable.

Now the countervailing force that nobody has priced, and I will be direct about it because it cuts against the clean bull story.

A miner that has signed a $9.7 billion HPC contract has to fund the buildout. Interconnect upgrades, transformers, liquid cooling retrofits, GPUs if they are going that far — none of that is free, and none of it is paid for by the customer up front. The capital structure of these transitions runs through equity issuance, convertible debt, and — this is the part that matters — liquidation of the BTC treasury.

A miner that no longer needs BTC as working capital has converted its treasury from an operating asset into a liquidatable asset. That is a category change, not a gradual drift. And the same five miners who reduced production by 47% are sitting on accumulated coins that now have a better use: funding a 15–30x EBITDA business instead of a 5–12x one.

So the net supply effect is not obviously negative. It may be a wash: less forced selling from operations, more discretionary selling from treasury to fund capex. The direction of the net is unknowable from public data, which is exactly why the confident bullish readings should be treated with suspicion. Yield is just risk wearing a smiley face. The reduced selling pressure is real. The treasury liquidation risk is also real. Anyone quoting only the first half is selling you something.

Core IV: eighty-eight times

Let me put the capital allocation decision in one comparison, because it is the single most clarifying number in the dataset.

Announced AI and HPC contracts across the sector: over $50 billion.

Annual mining revenue for the five transitioning miners, at $80,000 per BTC: approximately $566 million.

That is a ratio of roughly 88 to 1.

At 88 to 1, the word "transition" is wrong. This is not a company diversifying. This is a company exiting a business with a structurally capped return profile to enter one with contracted, escalator-bearing, investment-grade-counterparty cash flows. If you ran this comparison past any capital allocation committee on earth, the decision would take eleven minutes.

The only reason it reads as surprising is that the market spent a decade pricing these companies as BTC beta vehicles, and the beta framing became identity. When identity and arithmetic conflict, identity usually loses — but it loses slowly, and the lag is where retail money gets destroyed.

I want to be precise about what this ratio does and does not say. It does not say AI contracts are guaranteed. It does not say the counterparties are infallible. It says that relative to the alternative available to the same management team with the same asset base, the expected value calculation is not close.

And here is the part that makes the migration self-reinforcing rather than a one-off. When the market awards 15–30x EV/EBITDA to the converted and 5–12x EV/hash to the unconverted, the converted get cheaper capital. Cheaper capital lets them sign more contracts. More contracts widen the valuation gap. The unconverted watch their cost of capital rise in relative terms every quarter they delay. This does not stop at five companies. The peer pressure is mechanical, not psychological.

Which brings me to the mismatch that has gone almost entirely unremarked.

Cipher's production declined 18%. Bitfarms' declined 49%. These are the two extremes of the transition cohort, and the market is about to punish the wrong one. If the sector re-rates on "AI transition premium," then the company that has transitioned least gets hit twice: it does not earn the premium, and it still eats the post-halving margin compression. Being slow is not neutral in a bifurcating market. Being slow is the worst available position.

I have seen this exact pattern before. In 2024 I pulled 40% of my spot BTC exposure after watching IBIT custodian flows on-chain rather than reading the fund's press releases, verified the withdrawal proofs on Etherscan, and moved the remainder to a Ledger Nano X. The insight was not that institutions were selling. The insight was that the reporting layer and the settlement layer had diverged, and only one of them was verifiable. The same divergence is happening here: the reporting layer says "miners fleeing, hash surging." The settlement layer says difficulty is lagged, fleets are being refreshed, and half the new hash has no identifiable operator.

Core V: the hash nobody can audit

The fill for the 35 EH/s that left came from private miners, sovereign-adjacent entities, and non-US operators. Every one of those categories is outside the disclosure perimeter.

A listed miner files quarterly production reports. You can verify its hash rate against its financials. You can count its machines in an SEC filing. When MARA or CleanSpark or Riot adds capacity, it appears in a document with a signature attached.

A private miner in a jurisdiction with subsidized power files nothing. A sovereign-adjacent entity with a policy-driven electricity price files nothing. Their marginal cost is not a market price. It may be a political one.

This has two consequences, and the second one is the one that should worry people who care about network integrity.

First, the difficulty adjustment mechanism is being partially decoupled from market electricity prices. If a growing share of network hash has a near-zero marginal cost because its power is subsidized, then difficulty will rise past the point where market-priced miners can survive. That is not a bear market. That is a structural transfer of the security budget from commercial operators to political ones. It happens slowly, and it does not show up in any price chart.

Second — and this is the transparency problem that the source material does not mention at all — the auditability of Bitcoin's hash rate is declining even as the hash rate rises. Network security has always been defended on the grounds that it is observable. You can watch it. You can verify it. But observation requires knowing who is producing, and the composition shift is moving production from auditable entities to un-auditable ones.

I spent 2017 reading Status Network's token sale contracts in the final hour before mainnet, found an integer overflow in the minting function, and reported it privately. The bounty was modest. The lesson was not. The lesson was that unverified claims are not claims, they are advertisements. A hash rate figure with no operator disclosure is an advertisement. It is directionally useful and it is not auditable, and anyone building a security thesis on top of it is building on sand.

None of this means the network is unsafe. At 3.5% exited and 4.3% share, the security budget is not remotely threatened by this migration. What it means is that the measurement of the security budget is degrading, and degradation is invisible until the moment it is not.

The contrarian read: everyone is worried about the wrong tail risk

The consensus interpretation of this story has two halves, and both are wrong in opposite directions.

The bear half says: miners leaving weakens Bitcoin's security. This is falsified by arithmetic. 35 EH/s out of 1,005 EH/s is 3.5%. Five companies at 4.3% of blocks cannot threaten a network that re-prices hash automatically every 2,016 blocks. The security argument is not weak. It is off by more than an order of magnitude. Anyone making it has not done the division.

The bull half says: reduced miner selling is structurally bullish for BTC. This is directionally true and quantitatively trivial. Nineteen coins a day against tens of thousands of daily spot volume is a rounding error in the order book. The mechanism is real, the magnitude is not. And it is offset by a treasury-liquidation force that the bulls conveniently omit.

Here is what I think the actual tail risk is, and it is not on either side of that debate.

The real risk is on the AI side of the balance sheet, not the BTC side. A miner that signs a $9.7 billion contract has to build $3–5 billion of infrastructure before the first dollar of revenue arrives. That build is funded by equity and convertible debt against a counterparty that has not yet taken delivery. This is a capital structure with a multi-year duration gap between cash out and cash in, in a sector where GPU depreciation curves are brutal and hyperscaler procurement strategies can shift in a single earnings call.

If AI capex decelerates — and it will decelerate at some point, because every capex cycle does — those contracts get renegotiated, delayed, or absorbed at a loss. And the miner that has converted 80% of its interconnect to HPC colocation has no mining business left to fall back on. Its ASICs are sold. Its power contract is committed. Its revenue is a single counterparty.

That is a concentration risk that is structurally worse than the one it replaced. Mining had a hundred thousand counterparties — the entire global market — and a bad month meant lower revenue. HPC has one counterparty and a bad quarter means a covenant breach.

I am not predicting that outcome. I am noting that the sector is being re-rated on the assumption that the new business is strictly better, when the honest comparison is that it is differently risky — lower volatility, higher concentration, longer duration. The multiple expansion from 8x to 22x is not a reward for safety. It is a reward for a different risk profile that the market currently finds more legible.

And there is one more blind spot worth naming. Emotion is the only variable I cannot hedge. The miners are not making an emotional decision — they are making an obviously correct financial one. The emotional decision is being made by the retail holder who bought a "Bitcoin miner" at 8x, watched it re-rate to 22x on AI news, and now believes they own an AI infrastructure company. They own a company in the middle of a construction project with a counterparty concentration problem. The ticker did not change. The business did.

What I am actually watching

I do not trade narratives. I trade levels and I trade mechanisms, and this story has three mechanisms that will resolve on a schedule.

The difficulty epochs. Two adjustments is roughly four weeks. If the +105 EH/s was real new capacity, difficulty ratchets up twice and unit revenue per hash compresses. If it was largely smoothing artifact and fleet overclocking, the first adjustment will come in softer than the hash rate print implies. That single data point decomposes the entire story. Watch the adjustment percentage, not the hash rate chart.

The valuation spread. 15–30x EV/EBITDA for converted miners against 5–12x EV/hash for pure plays. If that spread holds or widens for two more quarters, the migration is not five companies, it is fifteen, and the supply-side story changes materially. If it compresses, the AI premium was a trade, not a re-rating, and the converted miners are now leveraged infrastructure plays with concentration risk and no mining fallback.

The miner treasury flows. This is the one that decides the net supply question, and it is verifiable on-chain. Watch the wallet clusters of the five transitioning miners. If BTC moves from cold storage to exchange deposit addresses while HPC capex ramps, the "reduced selling pressure" thesis is fully offset. If those balances hold flat or grow, the bull case survives. That is not a forecast. That is a balance to check.

What I am not watching is the hash rate headline. It is a lagged estimate of a fungible quantity, computed against a difficulty target that has not yet caught up, published without a stated methodology, in a market where a growing share of production has no reporting obligation at all.

Liquidity doesn't announce itself until it's gone, and neither does hash rate. The number that matters is not how much hash the network has. It is how much of that hash anyone can actually verify.

Right now, nobody can verify about half of it. And the price of everything downstream — the difficulty adjustment, the security budget, the miner equity multiple — is being set by a number that no longer has an owner's signature on it. That is the thing to sit with. Not the 47%. Not the 105 EH/s.

The 47% is arithmetic. The 105 EH/s is an estimate. The chart is a map, not the territory — and the people drawing this particular map have stopped telling us which projection they used.