Author: Mario Chow
Original link: https://mp.weixin.qq.com/s/T7-sK6lGPANb-SPY26EZAg
Disclaimer: This article is reproduced content. Readers can find more information via the original link. If the author has any objections to the way this article has been reproduced, please contact us, and we will make changes as requested. This reproduction is for information-sharing purposes only, does not constitute investment advice, and does not represent Wu Shuo’s views or position.
Stock prices are closing prices as of November 13, 2026; the Bitcoin price is $64,400 on November 16. MW refers to critical IT load, meaning the electricity actually delivered to servers. Gross MW refers to the electricity used by the entire campus, including cooling, and is typically 20% to 40% higher than the IT load.
TL;DR
Why convert? CoinShares values a completed, leased AI data center at about $27 million per MW, versus less than $3 million per MW for a miner’s powered capacity without a tenant. This 9-fold difference is not entirely the value of the grid connection, because the former includes $8 million to $15 million per MW in construction costs. Even after subtracting construction costs, the gap remains 4 to 6 times. On operating profit per MW, mining at the current hashprice nets about $230,000 a year, versus about $1.9 million for an AI lease. Bitcoin would have to rise to about $250,000 to close that gap. The underlying driver is the AI power shortage: in the hottest US markets, a new grid connection takes three to four years, while miners already have powered connections. The November 3 midterm elections made those connections scarcer, because Texas elected a governor who pledged to pause data center approvals.
How the conversion works. Grid connections, substations, land, and permits can all be reused; almost everything else is demolished and rebuilt, at $8 million to $15 million per MW—about the same as a new build. Miners are not selling a building, but the two to four years of waiting in line that tenants can skip.
Who is converting? Eight listed miners have signed leases for about 6.3 GW of IT capacity. On the morning after the election, Riot leased out the entire 910 MW Corsicana site; Cipher also has 700 MW of gross power. About 1 GW is actually online. CoreWeave alone accounts for 24%, with no guarantor behind the rent. Riot’s AI-lab tenant (reportedly Anthropic) now accounts for 18%, nearly as much as CoreWeave.
Revenue and costs. For leases that disclose NOI, figures range from $1.85 million to $2.15 million per MW per year after operating costs; Galaxy Helios Phase II is an exception, with NOI estimated from its margin at about $2.3 million. Annual return on building cost is 15% to 20%. Riot’s Corsicana lease breaks that range at about 33%, because the tenant builds the data center itself on Riot’s already-powered land. Differences in headline rent mainly reflect lease structure, not compensation for credit risk.
Impact. The miners still mining have sold about 40% of their bitcoin, and most of the remaining selling pressure is at MARA. Riot sold 20,264 coins this year and has only 2,418 left; MARA sold 13,597 after the quarter ended to repay bitcoin-backed loans. MARA, CleanSpark, and Riot now hold about 36,000 coins combined, down from 60,500 in August, with MARA’s roughly 20,400 the most likely to keep coming to market. Hashrate fell and recovered; miner exits have not weakened the network. Miner stocks are no longer bitcoin proxies either. After the election they diverged by state: Riot rose 25%, while bitcoin fell 12%; miners with projects still awaiting delivery in Texas and Georgia fell 15% to 18%.
IOSG view
CoreWeave is the uncompensated rent risk; Texas is the second risk. Galaxy charges CoreWeave the highest rent in the industry but has the lowest building return, at about 15% a year; CleanSpark’s Meta lease returns about 16%, and Hut 8’s investment-grade tenant lease returns 17% to 21%. The weaker the tenant’s credit, the more the landlord should earn; reality is the opposite. CoreWeave accounts for 100% of Galaxy’s AI business, about 53% of Core Scientific’s, and 28% of Applied Digital’s; Hut 8, CleanSpark, Riot, and TeraWulf have no exposure to it. CoreWeave reported better-than-expected revenue on November 10, and its stock surged 25% the next day, but recourse debt grew by about $10 billion during the quarter to $41.6 billion, and by November 13 the share price had returned to $53.
The election added a second test. Stocks with MW already approved, leased, and financed rose; those still awaiting Texas or Georgia approvals fell. Galaxy failed both tests: its business is entirely CoreWeave, and its Helios site is in Texas, with Helios II still unleased. Until CoreWeave credit stabilizes and the Texas pause has a clear end date, landlords with high exposure should trade at a discount to those with low exposure. Core Scientific is the cleanest example of CoreWeave exposure, and its stock stabilized after the election because 485 MW of the 590 MW it leases to CoreWeave is already collecting rent.
Together, the two tests divide landlords into three groups. First: already leased, outside Texas and Georgia, with financing in place—TeraWulf (New York and Kentucky, with Google guaranteeing project financing for its Anthropic lease), Hut 8’s River Bend (Louisiana, with non-recourse bonds), and Applied Digital (North Dakota, with guaranteed notes). Second: already leased, but in Texas or Georgia, with later phases still dependent on state permits or loans not yet finalized—Riot’s Rockdale and Corsicana, Hut 8’s Beacon Point, Galaxy’s Helios Phase I, and CleanSpark’s Sandersville. Third: not yet leased or not yet approved—Galaxy’s Helios Phase II, MARA’s Matagorda site, and Cipher’s Texas reserve projects. The first group should receive the smallest discount and the third the largest; Galaxy falls into both the second and third groups, with CoreWeave exposure on top.
As expected, the main sellers were the miners still mining, with MARA accounting for most of what remains. In August, converted landlords had almost no bitcoin, while MARA, CleanSpark, and Riot held about 60,500 coins combined, and all three had AI projects to fund. Two of the three have since sold. Riot sold 20,264 coins in the first nine months, worth about $1.5 billion, mainly to build data centers, and had only 2,418 left at the end of September. After the quarter ended, MARA sold 13,597 coins at an average price of about $81,000, repaid $750 million in bitcoin-backed loans, and said it would continue using bitcoin to fund its AI business. The three now hold about 36,000 coins combined (about $2.3 billion), meaning they have sold around 40% and have 60% left. Whether the rest is sold depends on who still needs cash. Riot’s 2,418 coins are immaterial; its Corsicana spending will be covered by cash and project financing. CleanSpark still has all 13,530 coins, and its Sandersville project is funded by guaranteed notes, so it has the least need to sell; however, it stopped issuing monthly production reports in October. MARA’s roughly 20,400 coins ($1.3 billion) are the selling pressure to watch: it says it will continue selling, it could owe up to $600 million for its Texas site if power is approved, and it is negotiating two leases that would require it to build the facilities itself. The sale of holdings has passed its peak, but is not over.
Miner stocks are now an AI-credit trade, with Texas permitting added as a second layer after the election. Converted landlords track CoreWeave and Nebius, not bitcoin: in October, while bitcoin rose, they fell alongside CoreWeave bonds. From November 3 to 13, bitcoin fell 12%, while Riot rose 25% on a 910 MW Texas lease with power already approved; Galaxy, Cipher, and CleanSpark fell 15% to 18% because they still have projects awaiting delivery in Texas and Georgia. Buying this basket of stocks is a bet on AI data-center financing and state-level permits, with only a little bitcoin exposure. Why convert?
Bitcoin rose, but mining economics got worse
Hashprice is the expected daily revenue per PH/s of hashrate before costs. It determines whether a miner can survive, and it has been falling steadily.
Miner
Before the 2024 halving, 1 PH/s could earn about $105 a day. Now it earns about $30, even though bitcoin’s price has returned to its level then. The halving cut the block subsidy in half, while network hashrate still grew about 70%, squeezing each machine’s share twice over. This is close to the low point reached this summer: hashprice briefly fell below $30, CoinShares estimated listed miners’ average cash cost at $75,500 per coin, and bitcoin was only around $60,000. The whole industry was mining at a loss.
Bitcoin rose to $84,000 in October, then fell back to $64,400 this month. Neither move changed much, because the returns from the other path are far higher.
The same 1 MW, two different calculations
The chart below uses a hashprice of about $35 from late October; at today’s $30, the mining figure would be even lower.
On operating profit per MW, 1 MW of new-generation miners nets about $230,000 a year after electricity costs at the current hashprice, and the next generation of miners would have to be purchased separately. Leasing the same 1 MW to an AI tenant can generate $1.85 million to $2.15 million in NOI (net operating income), on a 15- to 20-year contract with annual rent escalators. Even if bitcoin rose to $150,000 and difficulty stayed constant, mining would earn only about $1 million a year, with no contractual guarantee.
This comparison has one advantage for leases: $1.9 million in NOI requires $8 million to $15 million per MW of construction, while mining requires only about $800,000 per MW in miners. CoinShares’ valuation comparison is also not conclusive on its own. It values completed, fully leased AI data centers at about $27 million per MW, and miners’ powered capacity without tenants at less than $3 million per MW. That compares two types of asset, not the same site before and after leasing. Subtract construction costs from the valuation of a leased site and $12 million to $19 million per MW remains, still 4 to 6 times the value without a tenant—but that does not account for construction completion risk or the two years before rent begins. The direction is clear; the exact multiple is not.
The most convincing evidence is how much miners are willing to pay to exit. Core Scientific spent about $42 million to cancel a new miner order for 15 EH/s. Keel (formerly Bitfarms) stopped mining in June. Galaxy completely converted its Helios site and recorded an impairment.
How a mining site becomes a data center
Keep the power connection; rebuild everything else
After conversion, only what lies outside the meter, on the grid side, can stay.
Everything outside the meter can be reused; almost everything inside it is obsolete. ASIC miners are useless for AI, and the buildings usually have to be torn down. Miners proactively curtail when electricity gets expensive, so they have never installed backup generation or redundant power distribution, and cannot offer the uptime guarantees AI tenants require. An NVIDIA GB300 rack draws 130 to 140 kW, compared with about 9 kW for a standard server rack, and the heat has to be carried away by liquid-cooling pipes connected to the chips.
That puts conversion costs at $8 million to $15 million per MW: TeraWulf guides to $8 million to $10 million, Hut 8 to $9 million to $11 million, CleanSpark is at about $11.9 million, and Galaxy’s Helios Phase II is about $15.4 million—essentially the cost of a new build. What you save is not money; it is time.
Three ways to sell a megawatt
A landlord can lease a bare shell, rent out computing capacity, or sell the entire site outright.
Most of the industry sells a “powered shell”: land, power, buildings, and cooling, while tenants bring their own GPUs, networking, and operations staff. This division of responsibilities lets converted miners be valued like real estate companies. Riot’s Corsicana lease goes a step further, selling “powered land”: Riot delivers only the land, power, and rack platform, while the tenant builds and operates the data center itself.
IREN has chosen to rent out computing capacity. The clearest example is its 5-year, $9.7 billion contract with Microsoft, which requires it to purchase $5.8 billion of Dell hardware. Its revenue per MW is about five times that of a lease, but most of the chips’ value is gone after five years, and the contract term is only a third as long as a lease. Comparing IREN with the landlords on revenue per MW is the industry’s most common mistake.
The third option is to sell the site. TeraWulf sold its stake in the Abernathy joint venture to Fluidstack for about $530 million, of which $250 million had arrived by the third quarter. In 2025, CoreWeave tried to acquire Core Scientific outright, but Core Scientific shareholders voted it down.
Where the money comes from
A 200 MW site costs $2 billion at $10 million per MW, and construction spending starts about two years before the lease begins.
The timing risk to watch most closely is still Riot. Its $573 million Morgan Stanley bridge loan matures on December 31, 2026, but the first 96 MW at Rockdale will not begin leasing until December 2027, and the investment-grade guarantee for this $9.1 billion lease is still being finalized. At the end of September, Riot had $950 million in cash and bitcoin. It says it does not plan to issue shares and is arranging project financing for Rockdale and Corsicana. That financing has to be in place before the leases begin.
Who is doing this now
Landlords: signed does not mean built
The Corsicana lease makes Riot the second-largest landlord by contracted capacity, but almost none of that capacity is online yet.
Signed MW includes capacity that will not be delivered for several years; billable MW is the portion that has already started generating rent. Galaxy has leased 526 MW to CoreWeave, of which 133 MW is paying rent. The other 393 MW is not sitting vacant awaiting tenants; it has been leased and is still under construction. Core Scientific had 485 MW billable on September 30, up from 395 MW on June 30, and the remaining 100 MW at Dalton is expected online in early 2027. A grid connection solves only one bottleneck. Financing, construction, cooling, redundancy, and tenant acceptance still lie ahead, so low billable capacity alongside large future contracts does not indicate weak demand.
Cipher’s 700 MW is gross site power, not IT load, so we have not included it in the capacity share. However, its filings give Amazon’s IT load under the Black Pearl lease as 216 MW, equivalent to annual rent of about $1.7 million per MW. Black Pearl generated $5.9 million in lease revenue in the third quarter, Cipher’s first such revenue.
The industry’s largest lease was signed the morning after the election. On November 4, Riot leased its entire 910 MW Corsicana site to the same AI lab as at Rockdale, on a 15-year term for about $28 billion. Before that, there had been no new leases since September 18. Cipher extended the Barber Lake lease to 20 years, adding $5.2 billion to its contract value, but no new MW.
Tenants
With Corsicana, the share leased directly to AI labs more than doubled.
Neoclouds remain the largest tenant group, at about 37% of the roughly 6.3 GW. CoreWeave leases a total of 1,516 MW through Applied Digital, Core Scientific, and Galaxy, or 24% of the total; Fluidstack accounts for another 10%. Hyperscalers and tech companies account for about 30%, AMD-related arrangements for 9%, and direct leases to AI labs for 24%, up from just 11% before Corsicana. If Riot’s tenant is Anthropic, as reported, Anthropic’s direct leases total about 1.5 GW; add the 245 MW it leases at River Bend through Fluidstack, and end-user demand already exceeds CoreWeave’s.
Every landlord has a concentrated tenant base. Galaxy, CleanSpark, and Bitdeer each have only one tenant, while the same lab now accounts for 96% of Riot’s capacity. What really matters is CoreWeave concentration, because the credit picture has reversed since the summer. Back then, people worried about unrated AI labs. Since then, Anthropic raised $65 billion at a $965 billion valuation, reportedly told investors that its quarterly revenue in September was $16.8 billion, filed for an IPO on October 26, and is on a roadshow this month, reportedly targeting a valuation of about $1 trillion. What really scares the market is CoreWeave: rated B+, with $41.6 billion in recourse debt at the end of September, up about $10 billion in a quarter. In mid-October its bond yields approached 15% and 5-year CDS exceeded 1,000 basis points; they remained around those levels ahead of its November 10 earnings. The riskiest party in these leases is not the labs, but the middlemen subleasing capacity to them.
The few still mining
Only a handful are still mining at scale: Bitdeer at 82 EH/s, MARA at about 71, CleanSpark at 50, and Hut 8 subsidiary American Bitcoin at about 25. They are also keeping their options open. MARA’s 2 GW Matagorda site in Texas has been listed as a “studied load” in ERCOT’s Batch Zero process. MARA plans to sign two leases by year-end, but says they could slip into next year. CleanSpark’s Georgia mining rigs now sit next to its Meta lease. Riot is nominally still committed to mining: mining remains its largest source of revenue, but it has leased out 1,151 MW and has sold almost all its bitcoin.
Revenue and costs
How much rent can 1 MW earn?
Headline rent is contract value divided by the number of MW and years. NOI is what the landlord retains after property operating costs, before financing, depreciation, and taxes. Under a triple-net lease, the tenant bears operating costs, so the two figures are nearly identical; in a service-inclusive lease, the landlord collects higher rent but pays those costs itself, giving up 10% to 20%.
On an IT-load basis, headline rent ranges from about $1.44 million per MW per year (CoreWeave at Core Scientific) to about $2.6 million (CoreWeave at Galaxy’s Helios Phase II). The same tenant at both ends of the range shows that headline rent is clearly not the price of credit. At Core Scientific, CoreWeave pays for most of the data-center fit-out itself, so Core Scientific’s economic investment is about $1.5 million per MW. The lower rent therefore corresponds to much less landlord investment in the building. The four leases below disclose both rent and NOI, making the picture clear.
Riot’s lease with an AI lab at Rockdale (reportedly Anthropic) has headline rent 28% above Hut 8’s lease with an investment-grade triple-net tenant. After operating costs, the premium is only 9% using the midpoint of Riot’s disclosed range, and 4% at the low end. Most of the extra rent covers costs borne by the landlord, leaving almost nothing that could count as compensation for weaker credit.
Riot’s Corsicana lease is the newest, and a different kind of product. The 910 MW generates average annual rent of about $1.8 billion, or about $2 million per MW per year, including escalators, with a 99% NOI margin—within the Rockdale range. What changes is the cost. Riot invests only about $6 million per MW, compared with $8 million to $15 million for a full build, because the tenant builds and owns the data center. This is the first lease priced on “approved Texas power” rather than on a building. It was signed the morning after the election, which had just made approved Texas power scarcer.
Timing may matter more than credit. Amazon signed with Cipher at the end of 2025 for about $1.7 million per MW; nine months later, Meta signed with CleanSpark at $1.89 million. Both are triple-net leases with top-tier credit. Direct AI-lab leases signed in mid-2026 (TeraWulf, Riot Rockdale, Bitdeer) were priced at $2.37 million to $2.43 million. Corsicana, signed in November at about $2 million, breaks that pattern; on a base-term basis it would be lower still, because its $2 million is an average including escalators, while the other figures exclude them. It also covers far less construction. These are headline figures under different structures. We treat the pattern as a clue, not a conclusion.
Return on cost
Construction costs have risen almost as fast as rents, so building returns have stayed within a narrow range. Return on cost here is annual NOI divided by construction cost; it excludes financing and is not the return to shareholders.
A building with a 15-year tenant and a 15% to 20% annual return is an excellent real estate business. Miners can earn that because they hold a scarce resource; Corsicana shows how much that resource itself can earn when the tenant pays for the building. The 33% figure is Riot’s own, and assumes the project loan closes successfully; that loan is still being arranged. Galaxy, which has the weakest tenant credit, sits at the bottom of the range: highest rent, most expensive building, thinnest return.
How much is actually collecting rent?
Based on the latest disclosures, Core Scientific had 485 MW billable on September 30. Galaxy had 133 MW generating revenue on September 30, unchanged for the entire quarter; Phase II starts leasing in Q2 2027. TeraWulf had 102 MW online on September 30, with another 336 MW under construction. Since October 1, Applied Digital has had 250 MW available for service, and guides to 300 MW by year-end. Cipher received its first rent payment from Black Pearl in Q3 but did not disclose the MW figure; Amazon’s six-phase delivery schedule runs from September 30 through February 28, 2027. Riot’s 25 MW for AMD has been collecting rent since Q2.
Everything else is still in the delivery pipeline. Bitdeer’s two 60.5 MW phases are due in Q4 2026 and Q1 2027, respectively; Hut 8’s first River Bend building is due in Q2 2027; Riot’s Rockdale lab capacity is due in December 2027; Corsicana will be delivered in three phases from September 2027 to July 2028; and CleanSpark’s first rent payment is expected in Q4 2027. These figures use different dates and milestone definitions, so we do not combine them into an industry-wide percentage.
Impact
For bitcoin: the miners still mining are selling
First, the strongest bearish argument. Blockware estimates that publicly listed miners’ bitcoin holdings fell from about 127,000 at the start of the year to around 99,000 in mid-August, equivalent to about $1.8 billion in net sales to fund data-center construction. MARA is the largest holder and the loudest advocate of “HODL”—it changed its policy in March to allow sales from its reserves, and its holdings fell from 53,822 to 35,577 in six months. Blockware argues that this selling pressure is one reason the weakness in bitcoin in 2026 is being underestimated. That argument holds up. And it is continuing.
The converted landlords have long had little left to sell: Core Scientific and Cipher each hold fewer than 1,000 coins, about 1% of their market capitalization; TeraWulf, IREN, and Applied Digital hold essentially none. Selling pressure has therefore shifted to the few still mining and beginning to build their own AI projects. Riot sold 20,264 coins in the first nine months, mainly to build data centers, and had only 2,418 left at the end of September. After quarter-end, MARA sold 13,597 coins at an average price of about $81,000 and used $750 million of the proceeds to repay bitcoin-backed loans, releasing the pledged coins as well. Hut 8 sold about 3,100 coins to buy back shares. CleanSpark’s 13,530 coins are the only unchanged balance among the holdouts; it stopped issuing monthly production reports in October. MARA, CleanSpark, and Riot hold about 36,000 coins combined, worth roughly $2.3 billion at current prices, down from about 60,500 in August. That is a 40% reduction, not a complete sell-off. MARA still has about 20,400 coins and says it will continue selling to fund AI; CleanSpark’s construction is funded by notes, and Riot’s remaining 2,418 coins are immaterial.
Holdings are not the only source of supply. The network issues about 450 new bitcoins per day, and how many miners retain is their choice; conversion is pushing that choice toward selling. MARA changed its strategy in 2025 to allow sales of mined coins, then expanded it to its reserves in 2026; this year Riot sold more bitcoin than it mined. CryptoQuant data shows miners sent 21,140 coins to Binance on October 26, the sixth single-day transfer above 20,000 since February, with miner reserves near a five-year low. These flows also include private miners and mining pools, and a transfer to an exchange does not mean a sale: it shows coins were moved somewhere they could be sold, not that they have been sold. Bitcoin subsequently fell from about $84,000 in late October to $64,400, its lowest since August. Over the same period there were also outflows from spot ETFs, hawkish Federal Reserve comments, and an AI-stock selloff after NVIDIA announced on November 12 that Jensen Huang would step down as CEO. Miner selling added supply to this decline, but did not cause it.
The network absorbed the exits smoothly. Hashrate fell 19% from last November’s peak through July, then returned above 1,000 EH/s in October; based on mid-November difficulty, it is estimated at about 965 EH/s. Difficulty adjustment helped: after listed miners shut down, the remaining miners could mine more bitcoin per unit of hashrate, and new capacity came online elsewhere.
The longer-term issue is the security budget. This year, transaction fees accounted for just 0.63% of miner revenue, below 1% for the fifteenth consecutive month; revenue depends almost entirely on the block subsidy, which will be halved again in 2028. Adam Back’s view is that mining and AI are an arbitrage, and will eventually balance where returns in each are equal. This year’s hashrate recovery supports his argument. The 2028 halving is the real test: at current coin prices, after the subsidy is halved again, even new-generation miners will not cover their electricity bills at a power price of just $0.04/kWh, and more miners will shift power to AI. This year shows the network can absorb it: when some miners exit, others take their place.
For crypto markets: miner stocks are no longer bitcoin trades
Since August 3, bitcoin is up; every miner is down.
From August 3 to November 13, bitcoin rose 8%, while every miner fell, from Riot’s 18% decline to Cipher’s 72%. Miners still mining and those already converted fell just as hard, so this is not the market punishing conversion. There are five reasons. First, a timing mismatch: miner stocks peaked in June with the AI-infrastructure rally, while bitcoin bottomed near $63,000 in early August. Second, every miner, including the holdouts, now has AI projects to finance, and their stocks are being priced according to that financing. Third, the October CoreWeave credit scare affected all securities linked to AI data-center debt. Fourth, the benefit from higher coin prices is smaller than it looks because hashprice never rose much above $35. Fifth, the November 3 election discounted all projects still awaiting approvals in Texas or Georgia. Digital Realty and Equinix were between −6% and +2% over the same period, showing that the problem lies with the miners themselves: financing, permits, delivery, and tenant credit.
Based on 60-day rolling correlations, converted landlords were correlated 0.6 to 0.8 with CoreWeave and Nebius throughout the year, more closely than with either Digital Realty or bitcoin. What broke was the link to bitcoin, falling from about 0.55 in spring to 0.15 by mid-September. According to CryptoSlate, Hut 8’s bitcoin beta fell from 1.18 to 0.53, and Cipher’s from 1.13 to 0.58.
The election split the sector once again. From November 3 to 13, bitcoin fell 12%, while Riot rose 25%, Core Scientific gained 2%, and TeraWulf rose 1%; Galaxy fell 15%, Cipher 16%, and CleanSpark 18%. Bernstein cut its price targets for seven miners by 35% to 54% and now values AI hosting at 12 times next year’s EBITDA, citing tighter financing markets and uncertainty in Texas. Morgan Stanley favors TeraWulf, Hut 8, and Applied Digital because their sites are outside Texas, although Hut 8’s Beacon Point is actually in Texas.
This selloff began with companies borrowing heavily to lease out AI computing capacity, and has now spread across the entire AI sector. Since August 3, CoreWeave has fallen 38%, Nebius 46%, and NVIDIA 21%, with most of NVIDIA’s decline coming after its CEO announced he would step down. On October 14, CoreWeave bonds were sold off, and Cipher priced convertible notes the same day; Cipher fell 12%, while most of the sector fell 5% to 10%. The best-known miner ETF, WGMI, now has Cipher, Nebius, CoreWeave, and Hut 8 among its top holdings—two of which have never mined a single coin. For crypto investors, listed miners are no longer a basket of bitcoin exposure, but AI-infrastructure credit with a little bitcoin on the side.
For AI: real capacity, mostly for training
JLL estimates that about 66 GW of data-center capacity is under construction in North America, while BNEF estimates that the US grid can connect at most about 7.1 GW of new data-center load in a single year. Miners’ leases make up a substantial share of the capacity that can actually be powered before 2029, because the grid connections are already there. The midterm elections made this more apparent: the new governors of Texas, Georgia, and Ohio all campaigned on pausing new data-center construction, making already-leased grid connections more valuable. The limitations are discussed in the next section.
Risks
Texas. Riot, Galaxy, and Cipher’s AI sites are all in Texas, as are Hut 8’s Beacon Point and MARA’s new site. Riot, Hut 8 Beacon Point, and Galaxy alone account for 2.4 GW, or 38% of the total contracted capacity, not including Core Scientific’s Texas site or Cipher. Texas had already hit pause under a Republican governor. On August 3, Greg Abbott ordered an audit of the large-load interconnection queue. ERCOT immediately stopped Batch Zero studies and approvals for energizing new large computing loads; on September 21, Abbott also froze state environmental permits for data centers. The market had expected him to loosen restrictions after the election. Instead, on November 3 Democrat Gina Hinojosa won, pledging to pause all data-center approvals until the legislature passed a new law. The pause now has no end date. She takes office on January 19, a week after the legislature convenes. Her power has limits: the members of the Public Utility Commission are all Abbott appointees, her nominees require two-thirds approval in the state Senate, and Republicans still hold a 19-to-12 majority there. Our assessment is that the pause will last at least until mid-2027, with stricter conditions as the eventual outcome rather than a ban. The next milestones are ERCOT’s audit report on December 10 and the Public Utility Commission meeting on December 17. Contracted capacity on already-approved power is currently secure: when Riot signed Corsicana, its power had already been approved. The risk lies in capacity still requiring interconnection or permits: Galaxy’s additional 830 MW at Helios remains unleased; payment for MARA’s Matagorda site depends on power approval; and projects already under contract still need backup-generation permits. In Georgia, where CleanSpark’s Meta lease is located, the new governor wants a statewide temporary pause, and the Public Service Commission will also shift to a Democratic majority in 2027.
Demand is shifting toward inference. These sites are naturally suited to training: miners chose locations for cheap power, usually far from users. Gartner forecasts that inference will surpass training in AI-optimized cloud spending this year, at $23.3 billion versus $19 billion, and that inference will reach 59% in 2027. OpenAI has already scaled back some of its own construction plans, dropping an 800 MW expansion in Abilene in March and pausing Stargate UK in April. However, Microsoft took over the capacity in Abilene and Norway, so demand shifted rather than disappeared. The counterargument is valid: batch inference, agents, and reinforcement learning are not latency-sensitive, and the two types of load are not cleanly separated. But a 15- to 20-year lease is a long-term bet that the two will continue to overlap.
Intermediary credit. About 37% of capacity is leased to neoclouds, which then sublease it to labs. CoreWeave accounts for 24%; there is no guarantor behind the rent, it has $41.6 billion in recourse debt, and its debt was briefly priced as distressed in October. Management now says it will no longer issue unsecured bonds at current yields. Fluidstack accounts for 10% and is a private, unrated company; Google backs only specific obligations for TeraWulf and Cipher, not all of them. Bitdeer’s Volta depends on bank letters of credit that remain conditional. If an intermediary fails, the landlord is left with a half-built building and a single tenant, while the end-user lab never had a contract with it.
Financing. Most of these leases have not yet been paid for. Funding must come either from stocks that have fallen by more than a third since August or from debt priced against CoreWeave spreads. VanEck estimates the industry’s remaining funding gap at nearly $50 billion, and Bernstein now also points to tighter financing markets. Hut 8’s unrestricted cash fell to $157 million during the quarter; Bitdeer had $214 million at the end of September, while Tydal still has several hundred million dollars of capital spending remaining. Riot’s bridge loan, due December 31, is the nearest test.
Conclusion
The conversion is real, and the decision has largely been made. The grid connections held by miners are the scarcest resource in AI construction. Once completed and leased, a site is worth about $27 million per MW, compared with less than $3 million without a tenant; after construction costs, a leased connection is still worth four to six times as much. Bitcoin would have to rise to about $250,000 to change the operating economics, so even the holdouts are keeping their options open.
The election raised the price of this resource and split it in two. Adding 1 MW in the US is now more expensive and slower, making approved, leased, and financed MW more valuable, while MW still awaiting approval in Texas or Georgia are worth less. Signed the morning after the vote, Riot’s Corsicana lease is the clearest price signal for the former: when the tenant builds the data center, return on cost is about 33% a year.
What remains unfinished is the construction. About 6.3 GW is contracted, roughly 1 GW is online, and the rest must be financed over the next year or two. At returns of 15% to 20% a year on cost, these leases are good businesses, and rent differences between strong and weak tenants mostly reflect lease structure. That is reassuring, but it is also the problem: landlords with the greatest exposure to CoreWeave are not receiving additional compensation for it.
For bitcoin, the selling of holdings has passed its peak, but is not over. The converters sold first, followed by Riot and MARA. The three holdouts now have about 36,000 coins, with MARA’s roughly 20,400 the most likely to keep coming to market. Hashrate has recovered, and the network has absorbed the exits so far. For crypto investors, miner stocks are no longer a way to hold bitcoin. They carry AI-infrastructure credit risk and Texas permitting risk, with only a little bitcoin exposure left.
Methodology and notes
Capacity figures cover AI leases disclosed through November 13, 2026, excluding IREN’s GPU-cloud contracts. Contracted capacity includes future deliveries. Available-for-service, delivered, and billable figures are reported separately as each company describes them; reporting dates vary, and we do not calculate an industry-wide billable percentage. Capacity shares are shares of contracted IT MW, not revenue shares or credit ratings. Unnamed hyperscalers are not assumed to be the same company, and AI-lab demand subleased through a cloud provider is still attributed to the cloud provider. Riot has not disclosed the names of tenants for either of its AI-lab leases; Bloomberg reported that the Rockdale tenant is Anthropic, and Riot said Corsicana is leased to the same lab.
Cipher’s disclosed 700 MW is gross power, not IT load, and is therefore excluded from capacity shares. According to Cipher’s February 2026 investor materials, Black Pearl has 300 MW of gross power and 216 MW of critical IT load. Amazon’s 15-year, $5.5 billion contract therefore equates to about $1.7 million per IT MW per year ($1.22 million per gross MW). Cipher has been collecting rent at Black Pearl since August but has not disclosed billable MW.
Headline rent is the base-term contract value divided by critical IT MW and years, excluding escalators; Corsicana is the exception, since Riot disclosed only average rent including escalators. NOI figures are used only where companies disclosed them, namely Hut 8, CleanSpark, and Riot. NOI for Galaxy Phase II is estimated by applying its disclosed margin of more than 90% to its disclosed revenue. Return on cost uses each company’s construction-cost guidance; for Corsicana, it uses Riot’s $6 million per MW, including $4 million repaid to the tenant.
The Texas share includes Riot (Rockdale and Corsicana), Hut 8 Beacon Point, and Galaxy Helios. It excludes Core Scientific’s Texas site (not disclosed separately by site) and Cipher.
Mining economics assume miner efficiency of 19 J/TH (52.6 PH/s per MW), power cost of $0.04/kWh, continuous operation, and constant difficulty, and include only revenue less electricity costs. Hashprice estimates expected revenue using 144 blocks per day and network hashrate implied by difficulty; the current figure uses mid-November difficulty of 134.8T and a coin price of $64,400. Correlations use daily log returns on US trading days. Bitcoin holdings are each company’s latest disclosed balance as of the dates shown in the chart, valued at $64,400.
Lease details
IG means investment grade; n/d means not disclosed. Base value is the contract value over the base term. “Not disclosed” means no NOI figure could be found in the filings reviewed.
Sources
Mining economics and the network
CoinShares, Bitcoin mining report Q2 2026
Hashrate Index, September 2026 lookback
Blockware Intelligence, public miners sell about 28,000 BTC, and CoinDesk on the $1.78bn of selling pressure
CoinDesk, miners selling BTC to fund the AI transition
CryptoQuant, miners send 21,140 BTC to Binance
Charles Edwards and Adam Back on the AI pivot and security
Network data: blockchain.info and the mempool.space API
Markets and credit
CryptoSlate, miners no longer pure crypto proxies
WGMI holdings
CoreWeave: Q3 2026 results, Tech Times on recourse debt, CNBC preview on debt, 10 November, Q2 2026 results
CNBC, CoreWeave CDS at distress levels, 15 October, Bloomberg, CoreWeave risk gauge at a record, 15 October, Bloomberg, CoreWeave bonds tumble into distress, 14 October
Anthropic Series H; Reuters, 14 October 2026, "Anthropic's Q3 revenue surges 45% to nearly $17 billion"; Anthropic S-1, 26 October 2026; Reuters on the IPO valuation target, 10 November
Bitcoin and Nvidia in November: CoinDesk, 13 November, Yahoo Finance on the Nvidia CEO change
Analysts after the election: CoinDesk on miner target cuts, Yahoo Finance on Bernstein's cuts, Reuters via The Star on sector moves
Market data: Yahoo Finance daily closes to 13 November 2026
Power, AI demand, Texas, and the election
JLL North America datacentre midyear 2026 via Network World
BNEF via Bloomberg Law, US datacentre power demand
GB300 NVL72 rack power
Gartner, AI-optimised IaaS forecast, 10 August 2026
OpenAI's pullbacks: Bloomberg on Abilene, CNBC on Stargate UK, CNBC on Norway
Texas: Utility Dive on the interconnection pause and the ERCOT audit, Baker Botts on Batch Zero, Texas Tribune on the permitting halt, permits issued since and Hinojosa's plans, 4 November, Beveridge & Diamond, Texas Data Center Watch on what is paused
Election: CBS Texas on the Texas House flip, Georgia Recorder on the Georgia PSC, Scripps News on governors and data centers
The Block, Bernstein's "Follow the Gigawatts"
CoinDesk, VanEck on the $50bn funding gap
Company filings and announcements
Applied Digital: fiscal Q1 2027 results, fiscal Q4 2026 results
Core Scientific: Q3 2026 10-Q summary, Q3 2026 call, Q2 2026 results, Q2 2026 deck, CoreWeave HPC agreement, June 2024, AMD partnership, 529 MW
Hut 8: Q3 2026 results, Q2 2026 10-Q, $1.07bn revolver 8-K, River Bend lease, Beacon Point second lease
TeraWulf: Q3 2026 results, Q2 2026 investor presentation, Anthropic lease at Justified, Fluidstack and Google support (DatacenterDynamics)
Cipher: Q3 2026 business update, Black Pearl deck, February 2026, investor deck, 13 October 2026, Barber Lake amendment, $850m convertible notes
Galaxy: Q3 2026 results, Q3 2026 overview deck
Riot: Q3 2026 results and Corsicana lease, Q3 2026 10-Q, Q2 2026 results and Rockdale lease, Morgan Stanley facility 8-K, The Block on Bloomberg's report naming Anthropic
CleanSpark: Sandersville lease, investor presentation naming Meta, September 2026, notes indenture, $2.276bn senior secured notes, September 2026 operational update
Bitdeer: Q3 2026 results, Tydal lease with Volta
MARA: Q3 2026 shareholder letter, Q3 2026 10-Q, Texas
