Headlines say bitcoin miners are switching to artificial intelligence. The line is convenient. It mixes three distinct things: the earnings of a few listed companies, the computing power of the Bitcoin network, and the fate of the machines. The three do not move at the same pace.

The starting point is prosaic. After the 2024 halving and the drop in bitcoin’s price from the October 2025 highs, revenue per unit of compute — hashprice — has settled around $29 to $32 per petahash per day, according to CoinShares and Luxor.

At the same time, AI operators pay more for a megawatt that is already connected. U.S. interconnection queues run to thousands of gigawatts. A new site often takes years to be energised. Miners already have the substation.

The contracts speak. The halls, not yet

CoinShares counts more than $70 billion in AI and high-performance computing contracts announced by public miners. The firm says those groups could draw up to 70% of revenue from AI by the end of 2026, against about 30% at the start of the year.

At Core Scientific and TeraWulf, hosting has already overtaken mining in some quarters. IREN published a quarter where AI cloud services beat mining. Those figures describe a mix of revenue. Not the share of the Bitcoin network that would have changed use.

VanEck recalls the execution gap: only about a quarter of leased capacity would be delivered. The cost is not the same. A mining site often runs $700,000 to $1 million per megawatt. A liquid-cooled AI hall, $8 to $15 million, according to the same house and CoinShares. Leases sometimes run 12 to 20 years. Once signed, going back to mining on those square metres is no longer a quarterly decision.

The network has receded. For several reasons

Global hashrate peaked around 1.05 to 1.16 zettahash per second in late 2025. In early September 2026, current estimates sit around 900 to 923 exahash. Luxor puts difficulty about 19% below its November 2025 peak, one of the steepest drops of the specialised-machine era, and the second negative annual reading after China’s 2021 ban.

BlocksBridge measures, for listed miners, a drop of about 13 to 15% in realised hashrate in the first half of 2026 — closer to 21% excluding Bitdeer, which kept expanding mining. The global network receded less than that cohort. In other words: the companies visible on the stock market reallocate faster than miners as a whole.

Blaming the whole drop on AI would be excessive. Luxor also cites bitcoin’s price, ageing hardware, Texas curtailments, local accidents. In August the firm estimated about 1,150 EH of installed capacity against nearly 915 EH active. A non-negligible share is idle. Not necessarily converted.

The machines do not change jobs

A bitcoin mining machine does not train a model. D-Central and other operators say it plainly: what is reused is the interconnection, the transformer, the land. The boards come out.

Ethan Vera, of Luxor, expected liquidations of tens of thousands to more than 100,000 units per company over two years. The second-hand market has sunk: S19s sometimes trade under $2 per terahash. That can feed smaller sites, if the power is cheap enough. Otherwise the silicon goes dark.

Neither a sudden centralisation, nor an inverse miracle

The map moves without really diluting. In the third quarter of 2026, Hashrate Index put the United States at 36.7% of hashrate, about 345 EH, ahead of Russia and China. The top three countries still hold nearly two thirds. The United States lost the most volume on the quarter (−30 EH), a mix of margins and conversions. The U.S. share has slipped a little since the start of the year. It remains first.

Pools, another layer, barely move. Foundry USA still shows about a quarter of blocks. Unplugging machines in Texas does not, on its own, redistribute block production. ViaBTC’s CEO, Haipo Yang, describes a sort: the sites AI values most go to it; mining looks for the power that is left. Difficulty adjusts. The cost of attack falls with total hashrate, without that amounting, at this stage, to a security break.

The question the share prices do not price in

Mining can stop in seconds. A former ERCOT monitor, Beth Garza, saw “the best customers” of the grid: they take the surplus, they leave when the price explodes. In Texas a large share of emergency-reserve programmes long rested on those loads. CryptoSlate notes that conversion to high-availability leases removes precisely that option.

A Duke University study (2025) put the potential at 76 GW of new load integrable in the United States with 0.25% annual curtailment, 98 GW at 0.5%. AI, as it is contracted today, aims at the opposite: staying on. Trials exist to make it a little flexible. That is not yet the standard clause of a hyperscaler lease.

A table nobody publishes clearly remains: megawatts still mining, megawatts converted and lit, machines resold against machines switched off, flexibility lost by zone. Until those four columns exist, the “pivot to AI” will remain a valuation story more than a census.

Sources: CoinShares, Luxor, VanEck, BlocksBridge, Hashrate Index, D-Central, CryptoSlate, Duke University (2025). Nothing here is investment advice.