Bitcoin miners are now AI landlords - and the stocks are 40% off their highs
13 August 2026 · 10 min read
The short version: the companies that used to mine bitcoin are now the landlords of American AI compute. Seven listed converters have signed more than $140 billion of contracted AI revenue between them, on their own disclosures. TeraWulf already takes 71% of its revenue from AI leases rather than mining. Cipher has changed its name. Core Scientific is self-mining at two sites and losing money doing it.
And yet the four converters on our 34-name board are four of the seven names furthest below their 52-week highs. Every one of them is 38% to 43% down from its high while the board median sits at 74% of its own. The contracts arrived, the re-rating happened, and then it went into reverse.
All board figures are as at the close on 12 August 2026, the last complete session.
Which bitcoin miners are pivoting to AI?
Seven listed companies have converted, or are converting, mining campuses into AI data centres under long-dated contracts. Four of them sit on our board. The scale of what has been signed is the part most coverage understates:
Every figure there is a base-term number from the company's own filings or releases, and the terms are long: 15 years at Applied Digital and Hut 8, 20 years at TeraWulf and Riot, 12 years at Core Scientific. Riot's largest lease runs to June 2048.
The individual deals are worth stating precisely, because the aggregate is easy to dismiss as a round number:
- TeraWulf signed a 20-year lease with Anthropic for approximately 401 MW of critical IT load at its Hawesville, Kentucky campus - roughly $19 billion over the initial term, rising to about $33 billion if both five-year extensions are exercised. Initial capacity is expected in service in the second half of 2027.
- IREN signed a $9.7 billion, five-year AI cloud contract with Microsoft covering 76,000 NVIDIA GB300 GPUs across 200 MW at Childress, Texas, with a 20% prepayment - then added a $3.4 billion NVIDIA contract, under which NVIDIA took a five-year right to buy 30 million shares at $70, worth up to $2.1 billion.
- Cipher signed a 15-year, 300 MW, $5.5 billion lease with Amazon Web Services at Barber Lake, alongside roughly $3.8 billion of 10-year Fluidstack agreements in which Google backstopped $1.4 billion of the lease obligations and took warrants over about 5.4% of the company.
- Hut 8 leased both phases of its 1 GW Beacon Point campus in Texas on 15-year terms, taking campus-level base-term value to $19.6 billion and portfolio contracted capacity to 949 MW. 100% of that capacity is leased to or backstopped by investment-grade counterparties.
- Core Scientific added a 530 MW AMD agreement across five sites worth more than $14 billion, on top of roughly 590 MW contracted to CoreWeave.
Why a megawatt is worth more to AI than to bitcoin
This is the whole mechanism, and it is arithmetic rather than narrative. The same energised megawatt produces wildly different revenue depending on what you point it at.
Cambridge figures put AI cloud revenue between $1,600 and $4,000 per megawatt-hour, against $80 to $151 for bitcoin mining. Gross margins run 95-98% for AI cloud and anywhere from 13% to 83% for mining, depending on the bitcoin price, network difficulty and the cost of power. That is a gap of roughly ten to forty times on revenue, not a few percentage points.
Mining economics have also been compressing. Hashprice - revenue per unit of hashrate - has been running near $29 per petahash per second per day through mid-2026, a level at which only fleets with efficient machines and cheap power clear their costs comfortably. Meanwhile the network's hashrate keeps climbing, which mechanically shrinks each miner's share of a fixed block subsidy.
Put those together and the decision facing a miner with an energised 300 MW site is not close. One tenant signs a 15-year take-or-pay lease at investment-grade credit. The other pays a variable commodity price that halves by protocol every four years.
The revenue mix has already flipped
The pivot is not a plan. It is showing up in reported revenue right now:
- TeraWulf, Q2 2026: total revenue $44.8 million, of which $31.9 million - 71% - was HPC lease revenue and $12.8 million was digital asset mining. The company ended the quarter with about $3.0 billion of cash and sold its 50.1% stake in the Abernathy joint venture for roughly $530 million.
- Core Scientific, Q2 2026: ended June with nearly 30% fewer miners online than at the end of Q1 and is now self-mining at just two sites. Self-mining produced $21.5 million of revenue against $33.7 million of cost - a $12.2 million gross loss. Billable data centre capacity reached 437 MW by mid-July, up nearly 200 MW in a quarter.
- Cipher, Q2 2026: revenue fell to $25 million from $35 million because it decommissioned mining at Black Pearl to make room for the data centre. It then delivered initial capacity there in early August, two months ahead of schedule at the tenant's request, and began charging rent.
- IREN, Q3 FY26: AI cloud revenue more than doubled sequentially to $33.6 million, against $111.2 million from mining. Management targets $3.7 billion of AI cloud annualised run-rate revenue by the end of calendar 2026, and 480 MW of capacity rising to 1,210 MW in 2027.
- Riot, Q2 2026: data centre revenue of $23.2 million against $113.7 million from mining - the smallest flip on this list, and the newest, with its 191 MW lease signed this quarter.
Cipher Mining renamed itself Cipher Digital in February 2026 and announced plans to sell its bitcoin treasury. When a company changes the word on the door, the transition has stopped being optional.
What our board shows: the converters are the furthest from their highs
Here is the measurement we have not seen made anywhere else, and it is the reason this article exists.
Of the 34 names we track, four of the seven furthest below their 52-week highs are the bitcoin-mining converters. Every one of them announced multi-billion-dollar contracts in the last twelve months.
Read the last column carefully, because it is the striking one. Applied Digital's contracted, take-or-pay, non-cancellable lease revenue is 4.1 times its entire market capitalisation. TeraWulf's Anthropic lease alone is 2.2 times its market cap. Cipher's book is 1.5 times.
Those are gross revenues over 15 and 20 year terms, and delivering them requires enormous construction spending that has not been made yet. That is precisely the point: the market is pricing the cost and risk of building, not doubting the demand. Nobody disputes that Amazon, Microsoft, Anthropic and CoreWeave will pay. The question priced into these shares is whether the companies can finance and build 1,410 MW without diluting shareholders to nothing along the way.
For contrast, the same board contains Dell at 99.8% of its 52-week high, Arista at 98.0% and NVIDIA at 94.7%. The picks-and-shovels names that sell into this build are priced for perfection. The companies that own the land and the interconnect are priced at a 40% discount to their own recent highs.
Sentiment and attention are sorting them by who has revenue today
Our market sentiment score reads price and volume only. It has no idea what a lease is. What it produced on 12 August is a clean split down the four:
- IREN - score 76, 10th of 34, up 9.9% on the session and 12.3% over five
- Applied Digital - score 68, 16th
- TeraWulf - score 40, 31st, still down 4.9% over five sessions
- Cipher Digital - score 36, 32nd, down 4.5% over five
Against a board median of 66, two of the four are comfortably above and two are near the very bottom - only Western Digital and Cadence scored lower. That is a 40-point spread inside a group of four companies doing structurally the same thing.
The dividing line is not size and it is not backlog. It is revenue recognised today. IREN is running GPUs and booking cloud revenue now. Applied Digital has a building operational and earning. TeraWulf's Anthropic capacity does not go into service until the second half of 2027 and does not reach full 401 MW until early 2028. Cipher only started charging rent at Black Pearl in the first week of August. The two names the market is paying for are the two with a revenue line; the two it is not are the two whose contracts start in 2027 and 2028.
Attention splits the same way. We count cashtag posts on X ourselves, one row per complete New York day. On 12 August the board carried 32,161 posts, and the four converters took 2,399 of them - 7.5% of the day's conversation. IREN alone took 1,722 of those, or 72% of the group's attention. Applied Digital drew 214 and TeraWulf 158, both inside the fifteen quietest tickers we track.
Attention is a measure of how much a name is being talked about and nothing else. We have measured its predictive value across the whole board and found a rank correlation of -0.00 with the direction of the next day's move. It tells you about crowding, not about value.
Our view: this is the future, and the reason is power
What follows is judgement. The score reads price and volume, and none of this comes out of it.
We think the conversion of bitcoin mining capacity into AI capacity is permanent, and that it is one of the largest and least understood asset transfers in the current build-out. The reason has nothing to do with crypto and everything to do with the grid.
For a decade, bitcoin miners were doing something that looked, to most observers, economically marginal: buying cheap land near substations, negotiating interconnection agreements, securing power purchase agreements, and energising hundreds of megawatts. They were treated as a nuisance by utilities and a curiosity by investors. What they were actually doing was accumulating the single scarcest input in AI infrastructure.
Consider what it now takes to replicate one of those sites from scratch. Lawrence Berkeley National Laboratory found roughly 8,200 projects actively seeking US grid interconnection at the end of 2025, and the median project that reached commercial operation in 2025 had spent 61 months in the queue - against 22 months in 2008. Large power transformers run three to four years of lead time. Heavy-duty gas turbine slots are largely spoken for into the 2030s. Data centre electricity consumption is projected by the IEA to more than double from 415 TWh in 2024 to around 945 TWh by 2030.
An energised, interconnected, fibre-served 300 MW campus is therefore a five-year asset that cannot be conjured with money. The miners have dozens of them. Hyperscalers and AI labs need them immediately and can pay ten to forty times per megawatt-hour what bitcoin can. There is no plausible bitcoin price at which mining outbids AI for the same megawatt, because the gap is an order of magnitude and the block subsidy halves again in 2028.
That is why we do not read this as a cycle that reverses. It is a repricing of an asset class that was mislabelled. These companies were never really bitcoin businesses; they were options on grid interconnection, written when nobody else wanted them, and AI has exercised the options.
The second half of our view is about how the market will eventually classify them. Look at the shape of the contracts being signed: 15 and 20 year terms, triple-net, take-or-pay, non-cancellable, with investment-grade counterparties or an explicit Google backstop. Those are not technology contracts. Those are real estate contracts, and the businesses signing them are converging on something closer to a data centre REIT than to a crypto stock. Applied Digital's lease revenue is described in its own filings as take-or-pay and non-cancellable over initial 15-year terms; Hut 8 reports average annual net operating income of more than $1.75 billion. That is landlord language.
The gap between that reality and the "crypto miner" label the market still applies is, in our view, the single most interesting mispricing in AI infrastructure right now - and it will close through the label changing rather than through the contracts failing.
Landlord or operator? The choice that decides who wins
There are two ways to convert a mining site, and they carry genuinely different risks. The distinction matters more than which company has the biggest headline number.
The landlord model - Applied Digital, Hut 8, Cipher, TeraWulf, Core Scientific, Riot - builds the shell, the power and the cooling, and leases it. Revenue is contractual, margins are predictable, and the tenant buys the GPUs and eats their depreciation. The risks are construction cost, financing, and concentration: one tenant, one campus, one 20-year credit exposure.
The operator model - IREN - buys the GPUs, runs the cloud, and sells compute. Revenue per megawatt is far higher, and so is the risk. You own accelerators with an uncertain useful life, you carry utilisation risk between contracts, and you are competing with CoreWeave and Nebius rather than renting to them.
Our view is that the landlord model is the better risk-adjusted structure and the operator model is where the upside sits - and that the market is currently paying for neither of those things. It is paying for whichever company can show a revenue line this quarter. IREN scores 76 because it has one. Cipher scores 36 because its rent started eleven days ago.
That is a rational way to price execution risk in a sector where construction delays have already produced litigation. It is not a view about which business model wins over fifteen years.
What to watch next
Each of these is a specific, dated, checkable event rather than a sentiment call:
- Cipher's Q3 2026 results. Black Pearl rent began in early August. This is the first quarter where Cipher reports actual HPC lease revenue rather than contracted revenue, and it will show whether the 700 MW book converts on schedule.
- IREN's calendar 2026 exit run-rate. Management has committed publicly to $3.7 billion of AI cloud ARR and 480 MW by year end. That is an unusually precise target for a company at this stage, and it is falsifiable within five months.
- Applied Digital's Polaris Forge 1 Building 3. The 150 MW CoreWeave lease is expected operational during calendar 2026. Applied Digital carries the largest backlog-to-market-cap ratio on our board, so its delivery cadence is the cleanest read on execution risk across the whole group.
- Core Scientific's billable capacity. Guided to surpass 450 MW by late summer and 590 MW by early 2027, from 437 MW in mid-July. A miss here would be the earliest sector-wide warning.
- Financing. Every company on the table above has to fund construction. Watch the mix of debt, converts and equity - dilution is the mechanism by which a large backlog fails to reach shareholders.
The neocloud segment page carries the live version of every score quoted here, and the X mentions board carries the current post counts.
Frequently asked questions
Why are bitcoin miners pivoting to AI data centers?
Because the same megawatt earns dramatically more from AI. Cambridge figures put AI cloud revenue between $1,600 and $4,000 per megawatt-hour against $80 to $151 for bitcoin mining, with gross margins of 95-98% versus 13-83%. Miners also spent a decade acquiring energised, interconnected sites, which is now the scarcest input in AI infrastructure: the median US project reaching commercial operation in 2025 spent 61 months in the grid interconnection queue.
Which bitcoin mining companies have signed AI contracts?
Seven listed converters have disclosed more than $140 billion of contracted AI revenue between them: Applied Digital ($36.2B across 1,410 MW), Hut 8 ($26.6B across 949 MW), Core Scientific (over $24B across 1.1 GW), TeraWulf ($19B for 401 MW leased to Anthropic), IREN ($13.1B from Microsoft and NVIDIA), Cipher Digital ($11.4B across 700 MW leased to AWS and Fluidstack), and Riot Platforms ($9.8B across 241 MW). Four of the seven - IREN, Cipher, Applied Digital and TeraWulf - are on the ChipSentiment board.
How much of TeraWulf's revenue now comes from AI rather than bitcoin?
In the second quarter of 2026 TeraWulf reported total revenue of $44.8 million, of which $31.9 million - approximately 71% - was HPC lease revenue and $12.8 million was digital asset mining. Its 20-year Anthropic lease at Hawesville, Kentucky covers roughly 401 MW and about $19 billion of contracted revenue, rising to approximately $33 billion if both five-year extension options are exercised.
Why are these stocks down if the contracts are so large?
The four converters on our board closed 12 August 2026 between 56.8% and 61.4% of their 52-week highs, against a board median of 74.4%, making them four of the seven names furthest below their highs of the 34 we track. The market is pricing construction cost, financing and dilution rather than demand. Applied Digital's contracted revenue is 4.1 times its market capitalisation, TeraWulf's 2.2 times and Cipher's 1.5 times, but those are gross revenues over 15 and 20 year terms that require large capital spending still to be made.
What is the difference between the landlord and operator models?
Landlords such as Applied Digital, Hut 8, Cipher, TeraWulf, Core Scientific and Riot build the shell, power and cooling and lease it on long take-or-pay terms, with the tenant buying the GPUs. Operators such as IREN buy the accelerators and sell compute directly, earning far more per megawatt while carrying utilisation risk and GPU depreciation. On 12 August 2026 the operator scored 76 on our board and two of the landlords scored 40 and 36, a split that tracks which companies are recognising revenue today rather than which model is structurally better.
Is the shift from bitcoin mining to AI permanent?
In our view yes, because the revenue gap is an order of magnitude rather than a few percentage points and the bitcoin block subsidy halves again in 2028. An energised, interconnected campus takes roughly five years to replicate given interconnection queues and three-to-four-year transformer lead times, so AI tenants can outbid mining for the same megawatt at any plausible bitcoin price. The contract structures being signed - 15 and 20 year triple-net, take-or-pay leases with investment-grade counterparties - are real estate contracts rather than technology contracts.
Sources
- ChipSentiment - live board - prices, 52-week ranges, market caps and sentiment scores for all 34 names, quoted as at the close on 12 August 2026
- ChipSentiment - neocloud segment - the six constituents including IREN, Cipher Digital, Applied Digital and TeraWulf, with live scores
- ChipSentiment - X mentions board - daily cashtag post counts, counted by this site against the X API; 32,161 posts across the board on 12 August 2026
- ChipSentiment - do X mentions predict stock moves? - the -0.00 rank correlation between post volume and next-day direction
- TeraWulf - second quarter 2026 results - revenue $44.8m, HPC lease revenue $31.9m (71%), digital asset revenue $12.8m, approximately $3.0bn cash
- TeraWulf - Anthropic lease at Justified Data campus - 20-year lease, approximately 401 MW, about $19bn contracted revenue rising to $33bn with extensions, Abernathy stake sold for approximately $530m
- IREN - $9.7bn AI cloud contract with Microsoft - five-year contract, 76,000 NVIDIA GB300 GPUs, 200 MW at Childress, 20% prepayment, $5.8bn Dell supply agreement
- IREN - business update and Q3 FY26 results - revenue $144.8m, AI cloud $33.6m, bitcoin mining $111.2m, $3.4bn NVIDIA contract and 30m share right at $70, $3.7bn ARR target and 480MW by end CY26
- Cipher Mining - 168 MW 10-year AI hosting agreement with Fluidstack - approximately $3bn contracted revenue, Google backstop of $1.4bn and warrants over approximately 5.4% of the company
- Cipher Digital - Q2 2026 results and contracted portfolio - revenue $25m, 700 MW contracted, $11.4bn contracted revenue, average annual NOI of $793m, 5.3 GW development portfolio, Black Pearl delivered two months early
- DataCenterDynamics - AWS signs 300MW hosting deal with Cipher Mining - 15-year, 300 MW, $5.5bn lease at Barber Lake, Texas
- Applied Digital - annual report for fiscal 2026 (Form 10-K) - 1,410 MW of contracted critical IT load across five campuses and approximately $36.2bn of take-or-pay, non-cancellable revenue over initial 15-year terms as at 31 May 2026
- Applied Digital - additional 150MW CoreWeave lease - Polaris Forge 1 total capacity of 400 MW and approximately $11bn of contracted revenue across three CoreWeave leases
- Hut 8 - second 352 MW Beacon Point lease - 949 MW contracted IT capacity, $26.6bn aggregate base-term contract value, average annual NOI above $1.75bn, 100% leased to or backstopped by investment-grade counterparties
- Riot Platforms - second quarter 2026 results - 20-year 191 MW lease worth approximately $9.1bn through June 2048, 241 MW and $9.8bn contracted with AMD included, mining revenue $113.7m against data centre revenue $23.2m
- Core Scientific - Q2 2026 earnings call transcript - 1.1 GW leased customer capacity and over $24bn contracted revenue, 530 MW AMD agreement worth over $14bn, billable capacity 437 MW in mid-July, self-mining gross loss of $12.2m
- CCN - mining bitcoin versus training AI in 2026 - Cambridge figures of $1,600-$4,000 per MWh for AI cloud against $80-$151 for bitcoin mining, and gross margins of 95-98% against 13-83%
- Blockspace - Cipher rebrands to Cipher Digital - February 2026 name change and the stated plan to sell the bitcoin treasury
- Lawrence Berkeley National Laboratory - Queued Up: 2026 edition - roughly 8,200 active US interconnection requests at end-2025 and a median 61 months from request to commercial operation for projects built in 2025
- IEA - AI and data centre electricity demand - 415 TWh in 2024 rising to around 945 TWh by 2030
Figures are taken from the public filings and the reporting linked above.