The Miners Who Missed the Rally: Hashprice, Heat, and the Quiet Pivot to AI

On a Tuesday in late spring, a listed Bitcoin miner in West Texas filed a routine disclosure. It did not mention Bitcoin once. It mentioned a twelve-year power purchase agreement, a substation upgrade, and the phrase that has become the industry's quiet confession — high-performance computing. The same week, spot Bitcoin printed a local high, exchange tokens caught a bid, and the aggregate float of dollar-pegged stablecoins crossed a threshold that would have sounded like fiction in the winter of 2022. The miners — the ones who underwrite the entire security model of the network, the ones whose electricity bills are, in the most literal sense, the cost of trust — watched from the other side of the glass. This is not a story about bad timing. It is a story about what happens when an industry discovers that the most valuable thing it owns is not its hardware, but its place in the queue. To understand why, you have to hold two facts in your head at once. The first is mechanical. Bitcoin mining is not a technology business in the way most people imagine. It is a power arbitrage business wearing a technology costume. Each ASIC — an application-specific integrated circuit — performs exactly one operation, the double SHA-256 hash, and nothing else. It cannot be reprogrammed. It cannot be repurposed. It is silicon shaped into a single verb. The entire art of mining is acquiring electricity at a price below what that silicon can convert it into. The second fact is arithmetic, and it is brutal. In April 2024, the network's block subsidy halved to 3.125 BTC. An industry that had built its cost models around a certain subsidy woke up to half of it, with the same machines, the same debt, and the same megawatts. Hashprice — the revenue a unit of hashing power earns per day — compressed. Everyone in the business knew what followed: the least efficient operators curtail, then sell, then die. I have spent four winters reading miner filings the way other people read poetry. The 2022 credit crisis taught me the cadence of it. Core Scientific, once the largest public miner by hashrate, filed for Chapter 11 in December 2022 — not because the machine was broken, but because the capital structure was. Compute was never the constraint. Weather was never the constraint. The constraint was always the price of power against the price of the reward, and the ability to service debt in the gap between them. Then in January 2024, the spot Bitcoin ETFs began trading, and a new kind of buyer entered the market — one that wanted exposure without a single transformer, without a single curtailment contract, without a single megawatt-hour of operational risk. The miners did not get that bid. They got the consequences of it. And so, quietly, beginning in 2023 and accelerating through 2025, they began signing a different kind of contract. Here is the part that gets reported badly, and it matters more than anything else in this piece. You cannot turn a Bitcoin miner into an AI datacenter, because you cannot turn an ASIC into a GPU. This is not a matter of software. A SHA-256 ASIC is a fixed-function pipeline: transistors wired to do one thing, thousands of times a second, at a thermal and power envelope that assumes it will do nothing else, forever. General-purpose matrix multiplication — the thing that large language models actually consume — requires programmable units, high-bandwidth memory, and the freedom to be reconfigured every few months as model architecture drifts. A mining ASIC has the approximate flexibility of a stapler. A hundred thousand staplers do not become a printing press. So when a miner announces a pivot to "AI and compute," it is not pivoting its silicon. It is pivoting everything around the silicon. That list is longer and more valuable than people assume. It includes land already zoned for industrial load. Interconnection rights already granted by an increasingly reluctant grid. Substations and transformers — a component whose lead time stretched past two years during the boom, making an installed transformer a genuinely scarce asset. Long-dated power purchase agreements, often signed at prices no new buyer could obtain today. Immersion and air-cooled shell space. And finally, the only genuinely transferable piece: the balance sheet and the engineering culture that knows how to keep a thousand machines alive through a Texas summer. When I audited a small hosting operator's cost model back in 2021, the single line that mattered most was not hashrate or uptime. It was the interruptibility clause. Everything else was downstream of that. Now the numbers, because the numbers are where the poetry dies. Take a plausible modern fleet: machines running around 20 joules per terahash, so a megawatt of load produces roughly 50 petahash per second. At a hashprice near $45 per petahash per day — a level the industry has hovered around, sometimes painfully, through 2025 — that megawatt grosses about $2,250 a day. Power at five cents per kilowatt-hour costs roughly $1,200 a day. Before labor, before debt service, before the curtailment credits many operators quietly survive on, that is a gross margin of about a thousand dollars a day per megawatt — on a facility that cost several million dollars to build and needs refurbishing every three to four years as machines age out. Now put the same megawatt in front of a hyperscaler. Reported contracts for AI and HPC colocation in 2024 and 2025 cleared in a range that, annualized, is several times that gross margin per megawatt — with one crucial difference. The revenue arrives under a ten- to fifteen-year take-or-pay agreement, indexed, and largely indifferent to the spot price of anything. This is the trade. Not silicon, not sophistication — certainty for optionality. It explains something the market read as failure. When a miner sells Bitcoin to fund the build-out of an HPC shell, or issues convertible notes against a compute contract, or — most painfully — dilutes equity at a depressed price to make payroll through a halving, it is not missing the rally. It is choosing a different yield curve. The rally pays you if you can hold. The compute contract pays you if you can build. The two require opposite postures. That is the whole conflict, in one sentence. There is a second-order effect almost nobody prices correctly. The scarce input in AI infrastructure is not chips. Chips are hard, but they come from a supply chain that responds to money. The genuinely scarce input is a grid connection. In the major data center corridors — Northern Virginia, Texas, parts of the Nordics and Ireland — interconnection queues run years long, and utilities have begun telling new entrants, politely, that there is no power for them until the 2030s. A Bitcoin miner sitting on a 300-megawatt interconnect approved in 2019 is, in the most mundane possible sense, holding a queue position that money cannot buy. That is the asset. Not the machines. The queue. Which reframes the entire opportunity-cost narrative. The public conversation says miners missed the rally. The internal conversation — the one I hear when I talk to operators — says something closer to: we were always in the power business, and the power business just repriced. It also explains the politics. A large flexible load that can drop 200 megawatts in ninety seconds when the grid tightens is not an adversary of the grid; it is the grid's shock absorber, and increasingly it is paid to be one. Demand response revenue, ancillary service payments, curtailment credits — these lines now appear in miners' revenue breakdowns, and in some quarters they matter more than the block reward. In that framing, mining does not compete with the grid. It is grid infrastructure with a hash function bolted on. But — and this is the ethical debt I keep returning to — if the flexible load earns more selling capacity to a language model than securing a monetary network, the flexible load will do the more profitable thing. Nobody in this industry is sentimental. The block reward does not pay a loyalty premium. This is the part that should sit uncomfortably with anyone who believes Bitcoin's security is a solved problem. Bitcoin's difficulty adjustment is a remarkably elegant immune system. When hashrate leaves, difficulty falls, and the remaining machines earn more per unit of work. When hashrate arrives, difficulty rises, and the marginal operator is squeezed out. It is a self-correcting market for security, and it has worked through every crisis since 2010. What it does not do is guarantee amount. Difficulty adjusts the split of the pie. It does not adjust the size of the pie. The size of the pie is the block subsidy plus fees, divided by the price of electricity. If a structural share of the world's cheapest, most grid-advantaged power migrates permanently toward AI workloads, the marginal cost of producing a unit of security rises — not because Bitcoin failed, but because Bitcoin now competes for its most critical input against the most capitalised industry on earth. Fees are the theoretical answer. Fees have not yet been the practical answer, and I say that having watched every fee-spike narrative since 2017 evaporate once the mempool cleared. A network whose security is funded by intermittent cultural enthusiasm is a network with a seasonal budget. Meanwhile, the assets that did catch the bid tell you where the capital went. Exchanges rallied because a rally is the only product they sell, and because their revenue is a function of volatility, not direction. That is an old story. The stablecoin story is the interesting one. The float of dollar-pegged tokens expanded through a period when the underlying reserves — short-dated Treasuries — were still yielding meaningfully. In that configuration, an issuer earns the difference between what the reserves pay and what the holder receives, which for most holders is zero. It is one of the cleanest carry trades ever assembled, and it requires no lending, no leverage, no liquidity mining. It requires only that people keep dollars on-chain and stay still. Compare that to what passes for yield in DeFi lending. The rates on the largest money markets are not discovered; they are administered. They are governance parameters — a slope, a kink, a target utilisation — set by a vote and adjusted when the vote decides the number was wrong. I have never understood why we describe this as a market rate. A price that a committee can move with a transaction is not a price. It is a policy. Which brings me to the fork I think about most in this sector. The stablecoin that wins the next cycle may not be the one with the best collateral or the cleanest attestations. It may be the one the state can freeze on request. Convenience and control are the same feature, seen from two ends. Everything else — the peg, the chain, the audit — is downstream of that single design choice. There is a detail in the flow data that gets flattened into a headline. Through 2024 and 2025, aggregate miner reserves trended down. Some of that was operational — Bitcoin sold to pay for power, which is what mining has always done. Some of it was strategic — treasuries liquidated to fund HPC capex, which is what mining is newly doing. And some of it was forced — the same refinancing wall that broke operators in 2022, arriving on schedule for the class of 2021 debt. The ETF absorbed that supply and more. Which means the rally miners missed was partly made by their own distribution. The coins they sold into strength became the inventory that institutions now hold in custodial form, inside products with a fee ratio and a closing bell. I want to be precise here, because it is easy to be dramatic. Miners did not orchestrate their own exclusion. But the structure is hard to unsee: an industry sells its accumulated asset to survive a subsidy cut, and the buyers of that asset experience the appreciation the sellers needed. The seller gets runway. The buyer gets upside. Both are rational. Only one of them is described in the press as having missed out. And now the counter-argument, because I do not fully believe the pessimistic reading either. The consensus framing — miners lost, exchanges won, stablecoins won more — assumes the objective was to capture the rally. But an operator who survived the 2022 credit crisis and the 2024 halving and is still solvent in 2026 did not have capture the rally on the risk committee's agenda. They had do not be the next Chapter 11. Judged against that objective, signing a fifteen-year take-or-pay contract with an investment-grade counterparty is not a failure. It is the most adult decision this industry has made in a decade. There is also a version of this story where the miners who did not pivot win. If AI capex is a genuine multi-decade build-out, the pivot works. If it is a capital cycle — and every capital cycle in my lifetime has been a capital cycle — then operators who signed fifteen-year contracts at the top of a build-out will be renegotiating them in a decade, while the ones who kept their machines pointed at SHA-256 will be buying hashrate from liquidators at ten cents on the dollar. Hashrate is not a depreciating asset. It is a transferable one, and transfers happen at the bottom. From the ashes of 2022, we planted seeds for 2030. Some of those seeds were machines. Some of them were substations. I am no longer certain which harvest we should be praying for. The question I keep returning to is not whether miners will make money. It is who becomes the landlord of the world's most important ledger — and what they charge for the privilege of keeping it honest. A network whose security depends on the continued profitability of the least glamorous job in technology has always been, quietly, a bet on energy economics. That bet is now competing, megawatt for megawatt, with the most fashionable industry of the decade. Bitcoin does not need the miners to believe in it. It needs them to be paid. Watch the interconnect queues. They will tell you what happens next.