Only ~35 EH/s Is Actually Scheduled to Leave Bitcoin for AI: Listed Miners Bill ~550 MW of a 4 GW+ HPC Book, Realized Hashrate Fell 13.4% in Six Months, Yet Their Network Share Sits at a Record ~44%
The finding: the AI/HPC pivot has removed far less hashrate from the Bitcoin network than the contract headlines imply. CoinShares' Q2 2026 mining report (published September 15) counts at least 35 EH/s scheduled to exit from listed miners: IREN's 23.2 EH/s (roughly 380 MW) by December 31, 2026, Cipher Digital's 11.6 EH/s at Odessa likely by end-2027, TeraWulf winding down its remaining 145 MW, and Keel (ex-Bitfarms) which stopped mining on June 29. Most of that 35 EH/s is still hashing today; against a ~940 EH/s network the whole scheduled AI exit is under 4%.
The megawatt picture is the same story of backlog versus delivery. Listed miners have contracted more than 4 GW of IT capacity and disclosed over $100 billion of backlog, but only about 550 MW is actively billing, supporting roughly $1.1 billion of annualised AI/HPC revenue (CoinShares). Core Scientific is the bulk of it: 395 billable MW at quarter end, 437 MW by mid-July, $136.7 million of colocation revenue against $21.5 million of self-mining revenue in Q2, plus a new AMD framework for ~530 MW across five sites. TeraWulf booked $31.9 million of HPC lease revenue (71% of total); IREN's AI cloud revenue ($70.5 million) beat mining ($66.7 million) for the first time. Sector-wide, H1 2026 AI/HPC revenue was $341 million against $70 billion of announced contracts.
Where the hashrate did go is visible in the realized series, not the contract press releases. Miner Weekly's tally of public-miner realized hashrate fell from 368.3 EH/s in Q4 2025 to 344.4 EH/s in Q1 and 319.0 EH/s in Q2, a 13.4% six-month decline, versus a 10.6% network decline (1,071 to 957 EH/s). Strip out Bitdeer, which grew 44% to 63 EH/s on its own sub-10 J/TH hardware, and the rest of the cohort dropped 21.2% (324.6 to 255.9 EH/s). Hashrate Index's Q3 heatmap puts the US at 345 EH/s (36.7% of network), down 30 EH/s QoQ, with the network ~12% below its December 2025 peak.
What the headlines got wrong is attribution. The 100-plus EH/s that has left since the peak is overwhelmingly hashprice-driven curtailment of legacy machines at record-low economics (June hashprice $27.7/PH/s/day, an all-time low), not GPUs displacing ASICs. AI conversions are the slow, contractual tail of that same decision. A back-of-envelope check: at the ~16 J/TH implied by IREN's 380 MW running 23.2 EH/s, the entire 550 MW currently billing for AI would only ever have been ~34 EH/s, about 3.6% of the network, had it been mining instead. That is about one difficulty adjustment, and it is masked anyway by new deployments: the listed cohort's self-reported operating hashrate is 414 EH/s, ~44% of the network (ziven, September 22), a record share, while realized output runs nearly 100 EH/s lower. Listed miners are buying new-generation exahash and switching old exahash off at the same time, and the network series nets the two.
What to watch: IREN's December 31 mining shutdown is the first single removal large enough (23.2 EH/s, ~2.5% of network) to show up in a difficulty epoch on its own; Cipher's Odessa is the second, later. Core Scientific's billable MW moving from 437 toward its ~1.1 GW leased, and the sector's 550 MW closing on 4 GW, is the real conversion rate. And watch the spread between self-reported (414 EH/s) and realized (319 EH/s) hashrate: if it narrows while AI MW rises, the pivot is finally taking exahash off the network; if it widens, miners are still just curtailing and calling it strategy.
Sources (6)
AI Research
Key Takeaway
Listed miners have scheduled only ~35 EH/s (under 4% of the network) to leave Bitcoin for AI and are billing just ~550 MW of a 4 GW+ HPC book; the 13.4% six-month drop in their realized hashrate is mostly hashprice-driven curtailment, and new sub-10 J/TH deployments keep their self-reported share at a record ~44%.
The finding: the AI/HPC pivot has removed far less hashrate from the Bitcoin network than the contract headlines imply. CoinShares' Q2 2026 mining report (published September 15) counts at least 35 EH/s scheduled to exit from listed miners: IREN's 23.2 EH/s (roughly 380 MW) by December 31, 2026, Cipher Digital's 11.6 EH/s at Odessa likely by end-2027, TeraWulf winding down its remaining 145 MW, and Keel (ex-Bitfarms) which stopped mining on June 29. Most of that 35 EH/s is still hashing today; against a ~940 EH/s network the whole scheduled AI exit is under 4%.
The megawatt picture is the same story of backlog versus delivery. Listed miners have contracted more than 4 GW of IT capacity and disclosed over $100 billion of backlog, but only about 550 MW is actively billing, supporting roughly $1.1 billion of annualised AI/HPC revenue (CoinShares). Core Scientific is the bulk of it: 395 billable MW at quarter end, 437 MW by mid-July, $136.7 million of colocation revenue against $21.5 million of self-mining revenue in Q2, plus a new AMD framework for ~530 MW across five sites. TeraWulf booked $31.9 million of HPC lease revenue (71% of total); IREN's AI cloud revenue ($70.5 million) beat mining ($66.7 million) for the first time. Sector-wide, H1 2026 AI/HPC revenue was $341 million against $70 billion of announced contracts.
Where the hashrate did go is visible in the realized series, not the contract press releases. Miner Weekly's tally of public-miner realized hashrate fell from 368.3 EH/s in Q4 2025 to 344.4 EH/s in Q1 and 319.0 EH/s in Q2, a 13.4% six-month decline, versus a 10.6% network decline (1,071 to 957 EH/s). Strip out Bitdeer, which grew 44% to 63 EH/s on its own sub-10 J/TH hardware, and the rest of the cohort dropped 21.2% (324.6 to 255.9 EH/s). Hashrate Index's Q3 heatmap puts the US at 345 EH/s (36.7% of network), down 30 EH/s QoQ, with the network ~12% below its December 2025 peak.
What the headlines got wrong is attribution. The 100-plus EH/s that has left since the peak is overwhelmingly hashprice-driven curtailment of legacy machines at record-low economics (June hashprice $27.7/PH/s/day, an all-time low), not GPUs displacing ASICs. AI conversions are the slow, contractual tail of that same decision. A back-of-envelope check: at the ~16 J/TH implied by IREN's 380 MW running 23.2 EH/s, the entire 550 MW currently billing for AI would only ever have been ~34 EH/s, about 3.6% of the network, had it been mining instead. That is about one difficulty adjustment, and it is masked anyway by new deployments: the listed cohort's self-reported operating hashrate is 414 EH/s, ~44% of the network (ziven, September 22), a record share, while realized output runs nearly 100 EH/s lower. Listed miners are buying new-generation exahash and switching old exahash off at the same time, and the network series nets the two.
What to watch: IREN's December 31 mining shutdown is the first single removal large enough (23.2 EH/s, ~2.5% of network) to show up in a difficulty epoch on its own; Cipher's Odessa is the second, later. Core Scientific's billable MW moving from 437 toward its ~1.1 GW leased, and the sector's 550 MW closing on 4 GW, is the real conversion rate. And watch the spread between self-reported (414 EH/s) and realized (319 EH/s) hashrate: if it narrows while AI MW rises, the pivot is finally taking exahash off the network; if it widens, miners are still just curtailing and calling it strategy.