Listed Bitcoin miners have signed over $70 billion in cumulative AI and high-performance computing (HPC) contracts since mid-2025, converting grid-connected power assets originally built for proof-of-work hashing into infrastructure for large language model training and inference. CoinShares proj...
Listed Bitcoin miners have signed over $70 billion in cumulative AI and high-performance computing (HPC) contracts since mid-2025, converting grid-connected power assets originally built for proof-of-work hashing into infrastructure for large language model training and inference. CoinShares projects AI-derived revenue among public miners will reach 70% of total revenue by year-end 2026, up from approximately 30% in Q1 2026. At IREN, AI cloud revenue overtook Bitcoin mining revenue for the first time in Q2 2026 — $70.5 million versus $66.7 million — marking a structural inflection point for the sector.
The pivot is driven by economics. HPC workloads generate approximately $1.5 million in annual profit per megawatt of capacity, compared with roughly $500,000 for Bitcoin mining at current hashprices. With hashprice closing Q1 2026 at $23.9 per PH/s/day — the lowest reading since 2018 — and the network hashrate sitting 20.6% below its October 2025 peak of 1,151.6 EH/s, the financial incentive to redirect power to AI tenants is unambiguous. The question facing the sector is no longer whether miners will transition, but how much Bitcoin hashrate the network permanently loses in the process.
The post-halving mining environment has compressed margins to levels that make pure-play Bitcoin mining unsustainable for most operators. The April 2024 halving cut block subsidies to 3.125 BTC, and hashprice — the revenue a miner earns per unit of computational power — has declined steadily since.
Hashprice peaked at approximately $63/PH/s/day in July 2025, fell to $35-37 by November 2025, collapsed to $28-30 by early March 2026, and bottomed near $23.9 in Q1 2026, according to CoinShares. A partial recovery to $39.63/PH/s/day was recorded as of September 6, 2026, driven primarily by Bitcoin's price appreciation rather than difficulty reduction.
For reference, Bitcoin mining at current rates generates roughly $500,000 in annual revenue per megawatt. AI/HPC hosting generates approximately $1.5 million per megawatt annually — a 3x premium. At scale, the differential is larger: Core Scientific's six-facility CoreWeave buildout is projected to generate $10 billion over 12 years from approximately 400 MW of capacity, implying roughly $2.1 million per MW per year.
Mining remains profitable only for operators running sub-15 J/TH ASIC hardware with power rates below $0.08/kWh. According to CoinShares, the average all-in cost for listed miners in Q1 2026 exceeded $90,000 per BTC, pricing out operators without access to the cheapest energy.
The scale of announced AI infrastructure deals among former Bitcoin miners is unprecedented in the sector's history. Major contracts include:
| Company | AI Client | Contract Value | Term | Capacity | |---------|-----------|---------------|------|----------| | TeraWulf | Anthropic | $19 billion | 20 years | 401 MW | | Riot Platforms | Anthropic | $9.1 billion | 20 years | 191 MW | | Core Scientific | CoreWeave | $10 billion | 12 years | ~400 MW | | Cipher Digital | Amazon | ~$2 billion+ | 15 years | 300 MW |
TeraWulf's deal, finalized July 6, 2026, will see Anthropic lease a purpose-built facility in Hawesville, Kentucky — a converted former aluminum smelting site. TeraWulf plans to invest $3-4 billion in construction, less than one-fifth the lease value. Initial capacity is expected online in H2 2027, with full 401 MW operational by 2028.
Riot Platforms signed its $9.1 billion agreement with Anthropic on August 11, 2026, leasing 191 MW at its Rockdale, Texas campus. The deal could extend to $16.1 billion with two five-year renewal options. Riot plans 96 MW online by December 2027, with full buildout by June 2028.
Cipher Digital (formerly Cipher Mining) rebranded in May 2026 to reflect its strategic shift. The company closed a $2 billion high-yield bond for its Black Pearl HPC buildout in Wink, Texas, where Amazon has signed a 15-year lease for approximately 300 MW. Cipher delivered initial capacity in early August 2026, two months ahead of schedule. The company also secured a 900 MW site option in Texas for future expansion.
The network consequences are measurable. As of early September 2026, Bitcoin's hashrate had been below its October 2025 peak of 1,151.6 EH/s for 316 consecutive days — the longest sustained period below peak in a decade, according to CryptoTimes.
Key hashrate data points:
IREN's installed mining capacity dropped from 50 EH/s to approximately 23.2 EH/s in one year as the company redirected infrastructure to AI workloads. Core Scientific paid $41.9 million to terminate its 15 EH/s Proto mining agreement, choosing to eat the exit cost rather than continue mining operations at those facilities.
The difficulty adjustment mechanism has partially compensated. Bitcoin's difficulty fell as much as 19% from peak levels during 2026, improving economics for remaining miners. In early September, difficulty rose 1.31% while hashprice jumped 22%, reflecting the interplay between departing miners and price recovery.
IREN: Q2 2026 total revenue was $137.2 million, down 26.7% year-over-year. Bitcoin revenue fell 63% to $67 million. AI cloud revenue doubled quarter-over-quarter to $71 million. This was the first quarter in which AI revenue exceeded mining revenue. The company took a $639 million impairment on Bitcoin mining assets and plans to complete its transition by December 31, 2026. It targets $4 billion in operating annual recurring revenue.
TeraWulf: HPC leases represented 71% of total revenue in Q2 2026. Crypto mining revenue dropped to $12.8 million from $47.6 million in the prior quarter. The company sold a majority interest in its Abernathy mining joint venture to FluidStack as part of the Anthropic deal restructuring.
Core Scientific: Q1 FY2026 total revenue was $115.2 million, up from $79.5 million a year earlier. CoreWeave accounted for 77% of H1 2026 revenue. The company reported $10 billion+ in AI contracts and a 3 GW development pipeline. Core Scientific sought a $3.3 billion bond sale in April 2026 to fund further AI data center expansion.
Riot Platforms: Transitioned from pure miner to "AI infrastructure landlord" following the $9.1 billion Anthropic deal. The agreement covers Riot's existing Rockdale campus, keeping infrastructure costs relatively low by repurposing existing grid connections and cooling systems.
Cipher Digital: Rebranded from Cipher Mining in May 2026. Delivered Black Pearl capacity to Amazon two months ahead of schedule. Bitcoin mining at the site is being wound down to redirect power to AI tenants. The company closed $2.2 billion in financing ($2 billion high-yield bonds plus $200 million revolving credit facility).
HIVE Digital: Revenue increased 158% to $297.8 million in fiscal 2026. Hashrate grew from 6.5 EH/s to 25.1 EH/s. Contracted HPC ARR reached $35 million. Unveiled a 320 MW AI facility in Greater Toronto targeting $360 million in ARR when fully online in H2 2027.
VanEck's mid-September 2026 Bitcoin ChainCheck report identifies energized, grid-connected power as the constraining factor in the AI buildout. According to the report, Nvidia, AMD, and Broadcom collectively require approximately 30 GW of US capacity through 2027, while annual grid additions run at 15-25 GW. Only 2% of projects in the US interconnection queue reached the grid in 2025.
This structural power shortage underpins the valuation premium that miners command. VanEck's base case implies approximately 83% average upside across its Bitcoin miner holdings, premised on the thesis that controlling energized power gives miners leverage as AI chip suppliers risk missing revenue targets due to customer power constraints.
The dynamic creates a paradox for Bitcoin: the same infrastructure scarcity that makes mining sites valuable for AI makes them too valuable to use for mining. At 3x the revenue per megawatt, the opportunity cost of running ASIC miners instead of GPU clusters becomes a fiduciary question for public company boards.
Bitdeer operates 1,797 MW across ten data centers globally as of June 30, 2026. Hut 8 manages 1,020 MW with 1,230 MW under development across 19 sites. The aggregate capacity controlled by listed miners represents a meaningful fraction of available US data center power — and an increasingly material share of AI training infrastructure.
The sustained hashrate decline raises questions about network security, though the implications are more nuanced than they appear. Bitcoin's difficulty adjustment mechanism is designed to handle miner departures — it recalibrates roughly every two weeks to maintain 10-minute block times.
However, the current transition differs from previous mining downturns in a critical respect: the power is not going offline temporarily due to unfavorable economics. It is being contractually committed to AI tenants on 12-20 year leases. When TeraWulf's Hawesville facility comes online for Anthropic, that 401 MW is structurally removed from Bitcoin mining capacity for two decades.
The network's hash rate could stabilize as difficulty drops make mining more profitable for remaining participants, and new-generation ASIC hardware (such as the Bitmain XP series) improves efficiency. PowerCompute, for example, expects 39% more hashrate per machine from its September 2026 miner refresh. But the structural reallocation of power-rich sites to AI suggests the network will not return to its 2025 peak hashrate on any foreseeable timeline unless Bitcoin's price rises sufficiently to overcome the AI revenue premium.
The Bitcoin mining industry is undergoing a structural transformation, not a cyclical one. Previous downturns saw miners go offline temporarily, waiting for better prices or difficulty adjustments. The current transition is different: power assets are being contractually reassigned to AI tenants on multi-decade leases. The economics — 3x revenue per megawatt for AI versus mining — make this transition rational at the firm level.
For Bitcoin's network, the implications remain manageable in the near term. The difficulty adjustment mechanism continues to function, and remaining miners with access to sub-$0.08/kWh power and modern ASIC hardware can operate profitably. But the structural departure of grid-connected capacity suggests that Bitcoin's hashrate will increasingly depend on new power development rather than the repurposing of existing data center infrastructure.
The irony is not lost on market observers: Bitcoin mining's most durable contribution to its operators may not be the coins they mined, but the power infrastructure they built to mine them.