Selling Duration:
Bitcoin Miners, AI Data Center Conversion,
and the $9 Billion Deal Their Shareholders Refused

On October 30, 2025, Core Scientific’s shareholders were asked to approve the sale of their company to CoreWeave for roughly $9 billion in stock. They voted it down.

Of 245,792,464 shares represented, about 80% of those outstanding, 20,752,327 voted for the merger and 203,451,498 voted against. That is 82.8% of votes cast, and it means fewer than seven percent of the company’s outstanding shares supported a deal its own board had unanimously recommended. On the companion vote to approve executive compensation payable on closing, 95.8% voted against.

Core Scientific terminated the merger agreement the same day.

Almost nothing written about the miner-to-AI pivot accounts for this. The transaction is still routinely described as the deal that set the market price for mining infrastructure, at roughly $6.9 million per megawatt. It set no price. The people who owned the megawatts looked at the offer and decided it was too low.

Everything in this document follows from taking that refusal seriously.

The standard account of the pivot goes like this. Bitcoin’s April 2024 halving cut mining revenue in half. Miners were sitting on cheap, permitted, interconnected power. AI needed power urgently. Converting a mine to a data center was fast and cheap. So the miners pivoted, and the market rewarded them.

Three of those five claims are wrong.

Conversion is not cheap. It costs $8 to $15 million per megawatt, which is greenfield money. The speed advantage is real but it comes from the interconnection queue, not from the buildings, and the buildings mostly have to be replaced. And the market did not reward a pivot so much as it repriced a specific financial structure: a long-dated lease to a creditworthy tenant, wrapped in third-party credit support, financed in the bond market.

What actually happened is narrower and more interesting than a pivot. A handful of companies discovered that the thing they owned was not a mine and not even a data center. It was a queue position with power attached, and the way to monetize it was to sell fifteen to twenty-five years of it to someone whose credit a rating agency would accept. Once that became clear, the mining business was no longer a business worth being in, and several of them left it outright.

The risk moved too. It did not disappear. It moved from commodity price risk, which miners understood, to construction and counterparty risk, which they are learning.

And the question that follows from all of it, which Part eight answers with Texas data, is what should happen to the megawatts that are left.

The halving is the right starting point. Most accounts get its arithmetic wrong.

The halving, and the number everyone quotes incorrectly

Bitcoin’s fourth halving occurred on April 20, 2024, at block height 840,000, cutting the block subsidy from 6.25 BTC to 3.125 BTC. That much is undisputed.

The commonly cited figure for hashprice on halving day is $0.12 per terahash per second per day. Hashrate Index, which publishes the series, records something different: $71.40 per petahash per day on the hour of the halving, which is $0.0714 per terahash. The $0.12 figure does not correspond to any measured value that week. What did happen is stranger and more instructive: hashprice spiked to roughly $183/PH the day after the halving, on a one-off fee frenzy driven by the Runes protocol launch, then fell below $50/PH within days.

So the widely repeated “75% collapse in twelve months” is arithmetic performed on a starting number that was never measured. Blockspace’s own analysis of the first post-halving year gives the real move: “USD and BTC hashprice have fallen 57% and 89%, respectively, since April 2024.”

The part almost nobody has updated

The more serious problem with the standard account is that it stops in April 2025. Hashprice kept falling for another fourteen months.

The bottom was not April 2025. It was the first half of 2026, roughly 60% below the halving-day level, and the recovery since August 2026 has retraced maybe a third of it.

Difficulty stopped only going up

Here is the fact that breaks the old framing entirely. Network hashrate peaked near 1,160 EH/s in early October 2025 and fell to approximately 850 EH/s by February 2026. CoinShares records it as the first sustained six-month decline since China’s 2021 mining ban. Difficulty peaked at 145.04T in March 2026 and has fallen about 12% since; the March 20, 2026 adjustment of −7.76% was the tenth-largest decrease of the modern ASIC era. As of late September 2026, hashrate sits near 977 EH/s and difficulty at 127.45T.

For fifteen years, “difficulty only rises” was a safe assumption for anyone modeling mining revenue. It is no longer true, and the reason it stopped being true is the subject of this document. Institutional megawatts left mining for AI, and there were enough of them to move the network.

Fees did not save anyone

Transaction fees were supposed to replace the subsidy eventually. They are going the wrong way. Coin Metrics put fees at 1.33% of total miner revenue in Q1 2025. CoinShares reports them “consistently below 1% of total block rewards” through Q1 2026. July 2026 came in at 0.69%.

The commonly cited “under 5%” understates how thin this is by a factor of five, and the thinness matters: it means block subsidy is effectively all of mining revenue, and the subsidy halves again in 2028.

Why this halving hurt differently

Each of the first three halvings was followed, with a lag, by a price cycle that restored miner margins. The 2024 halving arrived under different conditions, and two of them were structural rather than cyclical.

The first is the cost of capital. In 2020, ASIC hardware could be financed at effectively zero real rates. By 2024 it could not, and debt service on equipment financing became a permanent addition to break-even that no efficiency upgrade removes. A miner who bought machines at the top of the 2021 cycle was still servicing them through the 2026 trough.

The second is that difficulty entered the halving at all-time highs. A halving cuts revenue instantly; difficulty only falls when operators actually switch machines off, and sunk-cost hardware keeps running at negative contribution margin for a long time before anyone concedes. That is why the adjustment took two years rather than two quarters, and why when it finally came, in March 2026, it came as the tenth-largest single downward adjustment in the ASIC era.

The industry described 2025 as the harshest margin environment it had seen. It was, and then 2026 was harder.

What this did to the marginal operator

Bitcoin mining is a self-equilibrating commodity business by design. Difficulty adjusts every 2,016 blocks to hold a ten-minute block interval, so as capital enters, each unit of hashrate earns proportionally less. The only durable advantage is electricity below the industry’s marginal cost.

At $28 to $37 per petahash, with best-in-class fleets around 15 to 16 joules per terahash, the electricity price that clears is low enough to exist in a handful of geographies. Fleet efficiency has improved a great deal, from a network weighted average near 34 W/TH in 2023 to sub-16 J/TH machines at the frontier, but efficiency gains have never outrun difficulty for long. They compress the margin of the operator who has not upgraded and hand the rest to difficulty.

That is the pressure. It is real, and it is worse than the standard account says. What the standard account gets wrong is what the pressure bought.

The most repeated claim in this subject is that miners can convert sites to AI for $1 to $2 million per megawatt, against $6 to $8 million for greenfield construction. Both halves are wrong, and the error is specific: the two sides of the comparison have been transposed.

The real arithmetic

CoinShares, in its Q2 2026 Bitcoin Mining Report, states it plainly:

“Converting mining infrastructure to AI grade facilities costs an estimated US$8m to US$15m per MW against US$0.7m to US$1m for mining.”

The $1 million per megawatt figure that circulates as the conversion cost is approximately the cost of building a Bitcoin mine. Conversion to AI-grade costs eight to fifteen times that.

The most careful published work on this is Colin Harper’s What’s a Megawatt Worth? Analyzing the AI Opportunity for Bitcoin Miners, Blockspace Media, May 20, 2026, which breaks the stack into three tiers and reaches the same place from a different direction. A Bitcoin mine at “$1 million per MW (sans ASICs) could be considered expensive.” A powered shell runs “up to roughly $16 million per MW as an upperbound.” A fully owned AI data center reaches “upwards of $50 million per MW in extreme cases.”

And the operators confirm it in their own filings. Core Scientific’s Q2 FY2026 earnings deck, filed July 28, 2026, gives the company’s working assumption: “Estimating $11 million – $12 million / MW as an average capex range for the site buildouts,” with tenant fit-out of another $1 to $2 million per megawatt on top.

Greenfield is also more expensive than the standard account says. Turner & Townsend’s Data Center Construction Cost Index puts US benchmarks between $9.5 and $13.3 per watt depending on metro, which is $9.5 to $13.3 million per megawatt.

Put those together and the conclusion is uncomfortable for the popular version of this story: on construction cost alone, converting a mine is not meaningfully cheaper than building new. In some configurations it is more expensive, because you are paying to remove things before you pay to install things.

So what is the miner actually selling?

Not the building. A Bitcoin mining facility is built to a specification that AI cannot use, and the gap is not a detail.

The peer-reviewed assessment is blunt about this. Voica, Panait and Iacob, writing in Energies in January 2026, tabulate the differences and rate each barrier. Mining facilities run at Tier 0 or Tier 1, meaning no redundancy; AI data centers require Tier 3 or Tier 4 with 2N or N+1 redundancy, which the authors mark as a critical feasibility barrier. Mining uses evaporative or ambient air cooling; AI requires precision liquid or chilled water, rated a high barrier. Mining needs minimal connectivity; AI needs “massive fiber backbones (terabits/s),” also high.

Their conclusion is that these are obstacles “that most remote mining sites cannot satisfy.” That is a peer-reviewed paper arguing directly against the conversion thesis, and any honest treatment of this subject has to carry it.

What survives conversion is the land, the substation, the utility agreement, and the position in the interconnection queue. That is the asset.

The number that prices it

Lawrence Berkeley National Laboratory’s Queued Up: 2026 Edition gives the figure that makes the whole thing legible:

“The median project built in 2025 took 61 months from the interconnection request to commercial operations, compared to 36 months in 2015 and 22 months in 2008.”

Sixty-one months. Roughly 2,061 GW of capacity was actively seeking interconnection at the end of 2025. Of requests submitted between 2000 and 2020, only 13% reached commercial operation; about 75% were withdrawn.

Transformers compound it. Wood Mackenzie’s Q2 2025 survey puts power transformer lead times at 128 weeks and generator step-up units at 144 weeks, against roughly 50 weeks in 2021.

A Bitcoin miner who energized a site in 2021 holds a position that now takes five years and a coin flip to replicate. That is the entire arbitrage, and it has nothing to do with the cost of a building.

This is the correction that matters most for anyone underwriting a site. If you believe the advantage is cheap conversion, you will pay for concrete and steel that has to be demolished. If you understand the advantage is the queue position, you will pay for an interconnection agreement and budget $11 million a megawatt to build on top of it.

Now to the transaction everyone cites and nobody finishes.

What was offered

CoreWeave announced its acquisition of Core Scientific on July 7, 2025. The structure was all-stock, with a fixed exchange ratio of 0.1235 CoreWeave Class A shares per Core Scientific share and no collar. At CoreWeave’s July 3 close, that implied $20.40 per share and roughly $9.0 billion fully diluted, a premium of about 66% to Core Scientific’s unaffected price of $12.30 on June 25.

CoreWeave’s stated rationale was power and cost. It would acquire “approximately 1.3 GW of gross power,” of which about 840 gross MW already supported CoreWeave’s own contracts and about 500 gross MW was mining capacity available for conversion. And then, in the line that became the deal’s problem:

“Immediate elimination of over $10 billion of cumulative future lease overhead to be paid for existing contractual sites over the next 12 years.”

That $10 billion of “lease overhead” was Core Scientific’s revenue. The two companies had built a hosting relationship to roughly 590 MW of critical IT load across six sites, which Core Scientific described in February 2025 as “$10.2 Billion over 12-year Contract Terms.”

So CoreWeave was booking, as a cost saving, the extinguishment of its counterparty’s principal revenue line, and offering $9 billion for a company whose contracted book it valued at more than $10 billion. It is difficult to argue a company is worth $9 billion in the same document that quantifies one of its contracts at $10.2 billion, and the opposition did not have to work hard to find the number.

Why the sellers said no

Two Seas Capital, holding about 6.3%, announced opposition on August 7, 2025 and ran a contested solicitation. Its objections were about structure rather than strategy, and one of them did the work.

The board, Two Seas argued, “left Core Scientific stockholders fully exposed to volatile fluctuations and a decline in the value of CoreWeave stock.” No collar, in other words, on a currency whose annualized volatility over the period was 161%. By late October the fixed ratio implied roughly $16.40 per share against the $20.40 headline, which made the point for them.

Both proxy advisers agreed, and neither questioned the strategic logic. ISS:

“The board conducted an exclusive process on a short timeline, and it did not obtain downside protection against the volatility of the acquisition currency”

Glass Lewis followed a day later, criticizing “the board’s decision not to insist on price protection.”

CoreWeave’s rebuttal argued that Core Scientific had no alternative: CoreWeave was “~100% of Core Scientific’s HPC colocation revenue and more than 76% of total revenue for 2026E,” Core Scientific “has signed no other HPC customer since emerging from bankruptcy,” and “any acquirer would simply be buying the right to become CoreWeave’s landlord.” It headlined the response “There Will Be No Bump.”

It held that position to the vote and lost.

What the sellers got instead

Core Scientific went forward alone, and the record since is the cleanest available test of who was right.

The CoreWeave hosting contracts survived the failed takeover intact. Core Scientific’s FY2025 10-K confirms capacity “increased to approximately 590 MW” and that CoreWeave “currently accounts for 100% of our Colocation segment revenue.” By the Q2 FY2026 deck the buildout was nearing completion at roughly 430 MW billing, with anticipated margins revised up from 75–80% to 80–85%.

Then the argument about customer concentration was answered directly. In 2026 Core Scientific signed AMD: a partnership “with the potential to support up to 2.5 GW,” anchored by 15-year agreements for approximately 530 MW across five campuses, carrying “more than $14 billion of potential base contracted revenue.”

That single contract exceeds the entire $9 billion equity value CoreWeave offered for the whole company, and it is precisely the customer diversification CoreWeave argued was unattainable. Core Scientific’s Q2 FY2026 revenue was $164.2 million, up 109% year on year, with colocation at $136.7 million against $10.6 million a year earlier.

As for the shareholders who refused: the consideration was 0.1235 CoreWeave shares, unhedged, and CoreWeave has since roughly halved. At its price in late September 2026, that consideration would be worth about $10.35 a share. Core Scientific trades near $17.20. The critics were vindicated on the precise point they raised, which is rarer than it sounds, and the mechanism is the transferable part: a fixed-ratio, all-stock offer hands the acquirer’s volatility to the seller, and in a sector where the acquirer’s own multiple is the thing under debate, that is the whole negotiation.

What the refusal established

Mining infrastructure does not clear at $6.9 million per megawatt, because the holders would not sell there.

And the refusal changed the structure of every subsequent transaction. No acquisition of a Bitcoin miner by an AI or cloud company has closed since. What has been signed instead, at enormous scale, is leases. A lease needs no shareholder vote, no exchange ratio, no collar, and no proxy adviser. The counterparties stopped trying to buy the companies and started buying the contracted capacity.

Core Scientific itself now sits on the landlord side of that structure with AMD. The company that was nearly sold for its power is leasing it instead.

If nobody is buying miners, what are these multi-billion-dollar announcements actually made of? The answer is a specific financial structure, repeated across three companies, and it is the least-explained thing in the sector.

The intermediary

The tenant in the largest of these deals is usually not a hyperscaler. It is Fluidstack, a private company that owns very little.

Fluidstack is a London-founded, now New York-headquartered AI infrastructure firm that operates as an asset-light intermediary. It signs the leases, builds and operates to spec, and leaves the silicon to its customers. Reported raises take it to roughly $2.6 billion of capital at an $18 billion valuation as of September 2026, led most recently by Jane Street. Its financials are not public. Its founding year is reported three different ways by three different outlets, which is a fair indication of how much verified information exists about a counterparty now sitting on the other side of tens of billions in lease obligations.

Across the disclosed miner deals, Fluidstack is the named tenant on roughly 980 MW of critical IT load:

Each deal uses a separate special purpose entity: Fluidstack USA I, II, III.

The mechanism

An unrated private company cannot support a fifteen-year lease that a bond investor will lend against. So a third party stands behind it, and that third party is Google.

The TeraWulf sequence is the only one disclosed in full, and it is worth following precisely.

August 14, 2025. TeraWulf signs Fluidstack for more than 200 MW of critical IT at Lake Mariner, ten-year term, approximately $3.7 billion of contracted revenue. Google “will backstop $1.8 billion of Fluidstack’s lease obligations to support project-related debt financing,” and receives warrants for 41,011,803 shares at an exercise price of $0.01 per share, about 8% of TeraWulf pro forma.

August 18, 2025. An incremental 160 MW. Google’s backstop rises by $1.4 billion to approximately $3.2 billion; a further 32,500,000 warrants take Google to roughly 14%.

October 28, 2025. The Abernathy, Texas joint venture, 168 MW over 25 years, approximately $9.5 billion. Google backstops approximately $1.3 billion. This time, “no TeraWulf equity securities or warrants were issued.”

Cipher Mining ran the same playbook: 168 MW at Barber Lake in September 2025 with a $1.4 billion Google backstop and 24,178,576 warrants at $0.01 for about 5.4%, then an additional 39 MW in November 2025 with a further $333 million, taking the total backstop to $1.73 billion.

The detail that explains everything

Here is what the backstop actually is.

TeraWulf’s financing subsidiary, WULF Compute LLC, offered $3.2 billion of senior secured notes to fund the Lake Mariner expansion. Google’s Lake Mariner backstop totals $3.2 billion.

The guarantee is sized to the bond, not to the lease. The lease is worth $6.7 billion over its initial terms and up to roughly $16 billion with extensions. Google is not insuring the rent. It is collateralizing the debt.

And the offering document makes the collateral role literal. The notes are secured by first-priority liens on, among other things:

“prior to the completion of the Data Center Expansion, a pledge by Google LLC of warrants to purchase common stock of TeraWulf.”

The warrants Google received are pledged back as security for the construction period. They are not a fee for the guarantee. They are part of the collateral package.

The miners say the quiet part in their own filings. TeraWulf’s FY2025 10-K states that the arrangement “materially enhances the credit quality and bankability of our contracted lease revenues.” Its Q3 2025 investor deck puts it on a slide: “$3.2B Google lease backstop fully covers debt exposure.”

What it buys

Hut 8’s River Bend lease, signed December 17, 2025, is 245 MW of IT capacity on a 15-year base term with three five-year renewals, $7.0 billion of base-term contract value and up to $17.7 billion with renewals, triple-net with a 3.0% annual escalator and expected annual NOI of $454 million. The tenant is a Fluidstack subsidiary. Google provides a financial backstop covering lease payments and related pass-through obligations for the full base term.

Note what Hut 8’s investor presentation chose to display: not a rating for Fluidstack, but Alphabet’s. “Backstopped by Google, a subsidiary of Alphabet Inc. (S&P: AA+, Moody’s: Aa2).” The credit being marketed is not the tenant’s.

On April 30, 2026, Hut 8 DC LLC closed $3.25 billion of senior secured notes at a 6.192% coupon, maturing November 2042, rated BBB− by S&P with a positive outlook and BBB− by Fitch with a stable outlook. Non-recourse to the parent. Hut 8 later closed a further $4.25 billion for its Beacon Point campus.

That is the achievement, and it is not a pivot. A Bitcoin mining company issued investment-grade project debt at a coupon in the low sixes for sixteen and a half years.

The re-rating people describe as “mining multiple to data center multiple” is better understood as a change in what kind of instrument the business is. Mining revenue is an equity-risk cash flow: volatile, uncontracted, no floor. A backstopped fifteen-year triple-net lease is a bond-risk cash flow. The companies that executed did not find a better business. They converted the same megawatts from equity risk into bond risk, and the bond market paid for the difference.

By comparison, TeraWulf’s notes, issued earlier and structured with the backstop sized only to the debt rather than the full base term, priced at 7.75%. The spread between 7.75% and 6.192% is roughly what the structure is worth.

The tenant is learning to skip the intermediary

One more move, and it points somewhere the sector has not fully absorbed.

On July 6, 2026, TeraWulf announced a direct 20-year lease with Anthropic for approximately 401 MW at the Justified Data Campus in Hawesville, Kentucky. Approximately $19 billion over the initial term, up to roughly $33 billion with two five-year extensions. No Fluidstack. No Google backstop. No warrants.

In the same announcement, TeraWulf sold its entire 50.1% interest in the Abernathy joint venture to a Fluidstack-led investor group for approximately $530 million cash, monetizing an investment of about $450 million at a premium.

So the largest single lease in the sector was signed without the structure that the rest of the sector depends on, by the end user who had previously sat behind the intermediary. Anthropic is the disclosed end user at Hut 8’s River Bend as well, through Fluidstack, under a partnership contemplating “at least 245 megawatts (MW) and up to 2,295 MW.” At TeraWulf’s Abernathy the end user is described only as “a global hyperscale AI platform developing frontier-scale foundation models” and is never named.

Anyone modeling this market on the assumption that the intermediary layer is permanent should look hard at Hawesville.

Four structures, not one

The Fluidstack-and-Google arrangement is the most discussed but it is one of at least four distinct ways these deals have been put together, and the differences are the useful part.

The credit-enhanced lease. Landlord leases to an unrated intermediary; a rated third party guarantees an amount sized to the project debt; the landlord may give up warrants. TeraWulf at Lake Mariner and Abernathy, Cipher at Barber Lake, Hut 8 at River Bend. Confirmed at three public miners and nowhere else.

The direct lease to the end user. No intermediary, no guarantor, no equity given up. TeraWulf and Anthropic at Hawesville, 401 MW over twenty years at approximately $19 billion. Also Core Scientific and AMD, approximately 530 MW anchored on fifteen-year agreements with more than $14 billion of potential base contracted revenue. This structure is only available to a landlord the tenant will underwrite directly, which is a statement about the landlord’s balance sheet as much as its sites.

The prepaid services contract. Not a lease at all. IREN’s $9.7 billion, five-year Microsoft agreement is a compute services contract with a 20% prepayment, which funds the GPU purchase without a guarantor. The counterparty risk sits with Microsoft, so no credit enhancement is needed; the landlord instead takes technology risk on the hardware it now owns.

The project financing against a single tenant. Galaxy Digital closed a $1.4 billion project financing facility for its Helios campus in Texas with CoreWeave as tenant, with no Fluidstack and no Google backstop. Hut 8’s Beacon Point campus, contracted to 704 MW at $19.6 billion of base-term value, names its tenant only as “the high-investment-grade company that executed the Phase 1 lease,” and carries no disclosed Google involvement at all.

The pattern across all four is the same underlying trade: the landlord converts an uncontracted megawatt into a contracted one, and the contract is only worth what the tenant’s credit is worth. Everything else, the intermediary, the guarantee, the warrants, the prepayment, is machinery for solving the same problem when the tenant’s credit is not sufficient on its own.

How the market prices what is left

VanEck’s June 2026 framework is the most useful published attempt to value these businesses, and it explicitly rejects the multiple most commentary uses.

Its argument is that EV/EBITDA is the wrong lens, because these companies are largely pre-revenue during the construction phase that dominates their next three years. What it proposes instead is a multiple on energized power: roughly 2 to 6 times for uncontracted capacity, above 10 times for capacity under lease. Tenant credit quality drives the spread within that range.

That framework prices the thing this document has been circling. A megawatt is worth a low multiple when it is energized and idle, and a high multiple when it is energized and contracted to someone who will pay for twenty years. The entire corporate activity of the past eighteen months has been the work of moving megawatts from the first category to the second.

It also explains the Core Scientific vote in one line. CoreWeave offered a price for a company whose megawatts were, at that moment, already contracted. The holders declined to sell contracted power at an uncontracted multiple.

Public companies dominate the coverage because they file. The private mining sector faces identical economics with none of the equity re-rating, and its answers are instructive precisely because they are unsentimental.

Crusoe left entirely

Crusoe Energy built the most differentiated private mining business in the industry on digital flare mitigation, using stranded natural gas at wellheads to power mining and reduce flaring. Its marginal electricity cost was close to zero, which is the strongest possible position in a commodity business.

It sold the mining business anyway. On March 25, 2025, NYDIG acquired Crusoe’s Bitcoin mining and digital flare mitigation operations, more than 425 modular data centers representing over 250 MW across seven states and two countries. Terms were not disclosed.

What Crusoe did with the freed capacity is the point. It is the developer and operator of the Abilene, Texas campus, 1.2 GW across eight buildings, funded through a $15 billion joint venture with Blue Owl Capital and Primary Digital Infrastructure. The first phase went live in September 2025 running NVIDIA GB200 racks. The site is leased to Oracle, which serves OpenAI under the Stargate program, with Lancium as land and power partner.

Note the layering, because it is the same shape as the Fluidstack deals seen from the other side: Crusoe builds and operates, Oracle takes the space, OpenAI is the end user. The operator is three parties removed from the workload.

An operator with the lowest marginal power cost in the industry concluded mining was not the best use of its organization. That is a harder data point for the pure-play thesis than anything a public company has said.

Bitdeer is running both, and being honest about the timeline

Bitdeer operates a geographically diverse portfolio and has taken the dual-track approach: keep mining where it clears, convert where it does not. As of April 2026 it reported approximately 65.5 EH/s of self-mining hashrate, down from 68 EH/s in February 2026, alongside AI cloud annual recurring revenue of approximately $69 million, up about 60% month on month from roughly $43 million in March.

Its 570 MW Clarington, Ohio site is the more informative disclosure. It is worth stating what Bitdeer actually says, because this site is widely misreported as a paused mining build. Bitdeer lists its usage as colocation, its timing as “to be updated,” notes that “design and other preparation work continues,” and discloses that “timing of power availability and construction may be affected by ongoing legal proceedings filed by a neighboring company, American Heavy Plate Solutions, LLC.”

That is a site re-designated away from mining, with an execution timeline exposed to litigation. Both facts are more useful than the version in circulation, and the second is a reminder that the binding constraint on these projects is frequently neither capital nor demand but something local and slow.

What private operators can and cannot do

The structural difference is access to the capital stack, and it determines which part of the value chain a private operator can occupy.

Direct GPU ownership requires funding hardware outright, which without public equity means private capital, debt against an unproven asset, or a joint venture partner. Full AI-grade conversion requires $8 to $15 million per megawatt, which for a 50 MW site is $400 to $750 million.

What a private operator can realistically do is sell the queue position rather than build on it: lease the site, sell the site, or joint-venture it with a developer who brings the capital. The infrastructure arbitrage is identical regardless of ownership structure. The capital structure determines who captures it, and for most private operators the answer is that somebody else does, in exchange for a price at signing.

The concentration here is worth stating plainly, because the filings state each piece and almost nobody assembles them.

Roughly 980 MW across three separate public landlords is leased to one tenant, whose obligations are guaranteed by one company, ultimately serving a market dominated by one end user. Three of the four are named in the filings. The fourth mostly is not.

The backstop is narrower than the headline

The miners disclose this themselves, in risk-factor language that deserves more attention than it gets.

TeraWulf’s FY2025 10-K: backstop arrangements “are only triggered upon a payment or insolvency event of default… There are other events of default or termination events that may result in the termination of a lease without triggering the applicable backstop.”

Cipher’s 10-K: the guarantees “will only be effective after rent commencement under such leases and are subject to certain limitations.”

Read together with the structure, the shape of the gap is clear. The backstop protects against tenant credit risk once rent has commenced. It does not protect against construction risk before then. Quinn Emanuel’s March 2026 client alert on AI data center financing makes the point sharply: the backstop “activates only after construction is complete and Fluidstack’s lease commences; during the build-out phase, bondholders bear full project execution risk.” The same alert notes that if construction runs more than 180 days past target completion, the tenant can terminate, and the guarantee would never trigger at all.

Which is to say: the risk that has been guaranteed away is the one miners are least exposed to, and the risk left standing is the one they have least experience managing. Building a Tier 3 liquid-cooled facility to a frontier lab’s specification on a fixed schedule is not an extension of running an ASIC shed.

The disclosure is uneven

The three deals are not equally legible, and the asymmetry is visible in the filings:

Hut 8 obtained the best financing terms while disclosing the least about the guarantee behind them, and gave up no equity. Whether that reflects a stronger backstop covering the full base term rather than just the debt, a better-structured SPV, or simply a market that had learned the structure by April 2026, is not determinable from public filings. It is the first question to put to any of these companies.

One further limit on this analysis, stated openly: S&P and Fitch both rated the Hut 8 notes BBB−, and Fitch published a forecast minimum debt service coverage ratio of 1.31x. The full rating reports sit behind agency subscriptions and could not be obtained for this document. So how the agencies treated the Google backstop and the Fluidstack counterparty as rating factors is not documented in any public source. That the notes came in at BBB− while Alphabet is rated AA+ and Aa2 tells you the agencies did not grant full credit substitution. Why, precisely, is not public.

The circularity

Google backstops a tenant that leases from miners to serve an AI lab in which Google is an investor, and separately backstops roughly $35 billion of that same lab’s data center leases underpinning a chip deal. Quinn Emanuel’s comparison is the one that should give underwriters pause:

“These circular dynamics have striking similarities to the ‘vendor financing’ practices that characterized the dot-com bubble of the late 1990s.”

This does not make the structure unsound. Vendor financing is a normal way to seed a market with a genuine end demand, and the end demand for AI compute is not in doubt. It does mean the credit support and the demand are not independent of one another, and a document that presents the Google backstop as external validation is missing that.

The pivot is usually described as optionality: power that can mine Bitcoin today and host GPUs tomorrow, switching to whichever pays better. The filings describe something else.

IREN stopped being a miner

IREN was the sector’s proof that a miner could run its own GPU cloud. FY2025, reported August 28, 2025, showed revenue of $501.0 million, up 168%, and net income of $86.9 million.

Then, on November 3, 2025, IREN secured a $9.7 billion, five-year AI Cloud contract with Microsoft on NVIDIA GB300 hardware at its 750 MW Childress, Texas campus, with a 20% prepayment.

And in FY2026, reported August 27, 2026: revenue of $707.0 million and a net loss of $702.6 million, driven by $638.8 million of non-cash impairments from decommissioning Bitcoin mining hardware.

You do not write off $639 million of ASICs if you intend to switch back. IREN is exiting mining.

Nobody held the line

The pure-play thesis had two public champions. Both are gone.

CleanSpark was the most consistent articulator of efficiency-first pure-play mining, and its fleet is genuinely excellent, at a peak deployed efficiency of 16.07 J/TH as of July 2026. On January 14, 2026 it acquired 300 MW, expandable to 600, on 447 acres in Brazoria County, Texas, explicitly for AI and HPC. On August 5, 2026 it signed a 20-year lease with “a high investment-grade global technology company” for 175 MW of critical IT at Sandersville, Georgia: $6.6 billion over the initial term, $11.6 billion with extensions, deliveries from Q4 2027, expected average annual NOI of approximately $330 million.

MARA was the Bitcoin treasury play. It signed an investment agreement for a 64% stake in Exaion, EDF’s HPC subsidiary, in August 2025. More telling is the treasury itself: MARA held 52,850 BTC at September 30, 2025 and 35,577 BTC at June 30, 2026, having sold roughly $1.6 billion of Bitcoin in the first half of 2026 to repay debt. Q2 2026 revenue was $174.9 million, down 27% year on year, with a net loss of $611.3 million.

The company whose thesis was holding Bitcoin sold a third of it to fund the transition away from producing it.

The commitments outlast the cycle

Riot Platforms holds over 1,100 acres and 1.7 GW of power capacity across its two Texas facilities. On January 16, 2026 it leased 25 MW of critical IT at Rockdale to AMD, with options to 200 MW, on a ten-year term worth $311 million initially and roughly $1 billion if fully extended. On August 10, 2026 it signed “a leading frontier AI lab” for 191 MW of critical IT on a 20-year term running through June 2048, at $9.1 billion base and $16.1 billion with extensions. Riot does not name the counterparty; secondary reporting identifies Anthropic, and this document does not treat that as established.

A lease ending in 2048 spans five more Bitcoin halvings. Those megawatts are not coming back to mining under any hashprice scenario, because they are contractually unavailable.

Core Scientific’s AMD agreements run fifteen years. Hut 8’s River Bend notes amortize into 2042. CleanSpark’s Sandersville lease runs twenty years.

This is the correction to the optionality framing. Optionality is what a miner has before it signs. The whole point of the transaction is to sell the optionality, because the optionality is exactly what made the cash flow uninvestable. A tenant paying $9 billion over twenty years is buying the certainty that the landlord cannot switch the load to something more profitable. The premium the market pays is the price of giving up the choice.

What it did to the network

The aggregate effect shows up on-chain. Network hashrate posted its first quarterly decline in six years in Q1 2026, and a second consecutive drop in Q2, down roughly 12% from the December 2025 peak. Blockspace’s July 2026 analysis attributes it directly to capital shifting toward AI, and records the US losing 30 EH/s, about 8%, while retaining 36.7% of global hashrate.

The longer-term question, what happens to Bitcoin’s security budget as institutional operators leave and fee revenue sits below 1%, is genuinely under-researched. There is one SSRN preprint modeling miner exit and attack-cost leakage at the subsidy-to-fee transition, which this document is not prepared to lean on without verifying its provenance. The serious academic literature has not caught up. That is a real gap, and anyone presenting a confident answer about it is guessing.

Everything to this point is national. Texas is where it can be checked, because Texas is where the capacity is and because the record is public enough to count.

Type 3 maintains a facility-level dataset of Texas data centers and large industrial loads, assembled from company announcements, permits, local reporting and ERCOT materials. What follows is drawn from it, with its limits stated first.

What the dataset holds, and what it does not

The file carries 477 Texas facilities. Only 69 of them have a disclosed capacity figure. The other 408 are overwhelmingly retail and network colocation, the small carrier-hotel and edge sites that cluster in Dallas, Houston and Austin, and for most of them no operator has published a megawatt number. They are real facilities and they belong on a map. They are not large loads and they are not conversion candidates.

So the analytical set is 69 sites totaling 16,948 MW, and within that, 54 sites at or above 100 MW carry almost all of it. That concentration is itself the first finding: Texas large-load capacity is not distributed across hundreds of facilities. It sits in about fifty.

Second limit, and it matters more. 415 of the 477 entries are listed as Announced, against 45 Operating and 17 under Construction. Fifteen of the twenty-nine crypto-mining entries are announced with no capacity attached at all. The discipline established earlier in this document applies here with full force: announced is a pipeline figure, and VanEck’s finding that roughly a quarter of leased capacity has actually been delivered is the base rate to apply to it.

Third, a classification caution. The dataset’s category field records what a site was, not what it is. Galaxy Digital’s Helios campus in Dickens County is tagged crypto mining while it converts to a CoreWeave AI lease. Core Scientific Denton is tagged crypto mining while it runs CoreWeave HPC colocation. Riot’s Rockdale campus is tagged crypto mining at 700 MW while 191 MW of it sits under a twenty-year AI lease and another 50 MW is contracted to AMD. Any count that takes the category field at face value will misclassify the most important facilities in the state.

Fourth, and this is the honest boundary on what follows: the dataset contains no curtailment data and no power-source field. Not one of the 477 records describes what the site’s power is doing. It can tell you where load is, how large, who operates it and what stage it is at. It cannot rank sites by curtailment exposure. Doing that requires nodal price history joined to these coordinates, which is a separate build.

In Texas, the pivot is already most of the way done

Reclassifying by current use rather than origin produces the finding this whole document has been circling.

Of the fourteen Texas crypto-mining sites with a disclosed capacity figure, eight, totaling 3,629 MW, have converted or are converting to AI and HPC. That is 78% of Texas mining capacity.

What remains as pure mining is 1,040 MW across six sites: Hut 8 Vega in Oldham County at 205 MW, MARA Garden City in Glasscock at 200 MW, MARA Granbury in Hood at 200 MW, Genesis Pyote in Ward at 195 MW, Bitdeer Rockdale in Milam at 170 MW, and Genesis Rowdy in Wilbarger at 70 MW.

Set that against the national picture established earlier. IREN wrote off $639 million of mining hardware. CleanSpark signed $6.6 billion. MARA sold a third of its Bitcoin treasury. The Texas fleet-level count says the same thing from the ground: in the state with the most mining capacity in the country, roughly four megawatts in five have already left mining or are leaving.

The six that stayed skew west. 58% of the remaining pure-mining capacity sits west of the 100.5° meridian, against 45% of the converting capacity. That is the direction a cheap-and-congested-power thesis would predict, and it is consistent with where ERCOT’s wind sits. It is also six data points and a thirteen-point gap, which is not enough to call a pattern. It is a hypothesis this dataset is consistent with, and it is stated here as nothing more than that.

One site answers the firmness question

The single most informative record in the file is not a mining site at all.

Amazon’s Pecos County campus, listed at 960 MW and under construction, is described in the underlying notes as an 8,000-acre site north of Fort Stockton planned to run partly disconnected from ERCOT, via a dedicated 7.65 GW gas-turbine plant, 1.8 GW of battery storage and 750 MW of solar.

Read what that implies. A hyperscaler evaluated the question of whether firm power can be bought from the Texas grid for a campus of this size, and answered by building its own generation at roughly eight times the load. Note also that the 7.65 GW is generation capacity and not IT load, which is undisclosed; the 960 MW in the dataset is an estimate, and it should be read as one.

This is what the regulatory record in Case Study No. 1 predicted. Senate Bill 6 requires curtailment capability of transmission-voltage loads interconnected after December 31, 2025. NOGRR282 and NPRR1308 imposed ride-through obligations on large computational loads at or above 75 MW from August 1, 2026. Docket 59220 imposed uncompensated thirty-minute curtailment on a co-located load and barred it from demand response revenue entirely.

Firm grid power above 75 MW is no longer a thing ERCOT sells. Every large load in Texas is now curtailable by rule. The only remaining question is depth, notice, and who absorbs the cost. A tenant that needs genuine firmness has to manufacture it, and Amazon is manufacturing it.

The allocation question, answered properly

The intuitive rule is that interruptible power should go to Bitcoin mining and firm power should go to AI, with some number of curtailment days as the dividing line.

The direction is right. The dividing line does not exist, and the reason is that the two businesses fail in different ways.

Mining revenue is proportional to uptime. Lose a fifth of your hours, lose a fifth of your revenue. The business degrades gracefully and keeps running.

AI lease revenue is binary. These are triple-net leases; the tenant pays whether or not power flows. Curtailment does not shave the landlord’s revenue, it threatens the lease itself. Quinn Emanuel documents the specific mechanism: run construction more than 180 days past target and the tenant can terminate, at which point the credit support never triggers at all. There is no curtailment-days threshold at which an AI site gradually becomes a mining site. There is a cliff, and on the far side of it the asset is worth its land and its interconnection.

The capital asymmetry makes the asymmetry of outcomes worse. Conversion to AI-grade runs $8 to $15 million per megawatt, and Core Scientific’s own working figure is $11 to $12 million. A mining buildout, using the ASIC pricing Type 3 tracks and a $300 per kW infrastructure adder, runs between roughly $0.93 million per megawatt for a cost-optimised fleet and $3.43 million per megawatt for frontier hydro hardware at 9.5 joules per terahash. You are either betting eleven million a megawatt on a binary outcome or one million on a proportional one.

And the asset lives do not match. An ASIC is economically dead in three to four years. Hut 8’s River Bend notes amortise into 2042. These are not two uses of one asset. They are two different assets that happen to share a substation.

What a megawatt-hour is actually worth

The four largest disclosed leases, reduced to contracted revenue per megawatt-hour across the full base term:

The band is tight, between roughly $215 and $272, and it is net of power because these are triple-net structures where the tenant carries the energy cost.

Bitcoin mining, at the September 2026 hashprice of about $37.65 per petahash per day:

So a contracted AI megawatt-hour is worth roughly three times a mined one on a current best-in-class deployed fleet, and roughly twice one on hardware that is only now shipping. The efficiency frontier is closing the gap faster than most commentary registers, and an operator comparing the two on a 2024 fleet assumption will understate mining by a wide margin.

These figures move. Hashprice is the most volatile input in this document, and a return to the July 2026 trough of $28.09 per petahash would cut the mining column by a quarter.

The finding that inverts the intuition

Here is where the straightforward version of the allocation rule breaks, and the evidence is Type 3’s own internal curtailment model rather than anything published.

A mine that runs only on curtailed hours does not pay back.

Take the common framing of a site curtailed sixty days a year. That is 1,440 hours, or 16.4% utilization. At the September 2026 hashprice:

Against an ASIC that is obsolete in three to four years. Type 3’s internal model, run on different inputs, reaches 16.3 years and the same verdict. The hardware depreciates on a calendar, not on run-hours, and a fleet that sits idle 84% of the time is being destroyed by time while it waits.

Run the same fleets at 95% utilization on cheap power, self-curtailing when the grid pays more than the hash does, and the picture reverses completely: 1.7 years payback on the cost-optimised fleet and 3.1 years on the frontier fleet.

So the rule is the opposite of the intuitive one. Curtailment-exposed power does not make a mining site. Cheap baseload power with the right to curtail makes a mining site. The curtailment is not the business. It is an option written on top of the business, exercised when the grid outbids the hash, and it is worth something only because the mine was going to run anyway. That is precisely the model Riot ran at Rockdale when it earned $71.2 million in power credits in 2023, and it is why the credits showed up as a reduction in cost of revenue rather than as revenue.

What this means for a Texas site today

Three questions, in order, and the first one disqualifies most sites.

Can this site support a fifteen-to-twenty-year lease that a rating agency will accept? That requires a tenant whose credit stands on its own or a guarantor whose does, and firmness that in ERCOT above 75 MW now has to be manufactured rather than purchased. If yes, the $11 to $12 million per megawatt is justified and the site is an AI site. If no, no amount of curtailment arithmetic rescues it.

If not, can power be bought cheaply enough to run at high utilization? That, and not the number of curtailable hours, is what makes a mining site. The curtailment right is the second revenue line, not the first.

If neither, the asset is the interconnection. Sell it, lease it, or joint-venture it with someone who brings the capital. The 61-month median from interconnection request to commercial operation is what the buyer is paying for, and it does not depreciate.

What the Texas data cannot yet tell you is which specific sites sit on power cheap and congested enough for the second path. That requires nodal price history joined to these coordinates, and it is the next thing worth building.

A document that only makes the case is not worth citing. Here are the strongest arguments against the conversion thesis, with their sources.

Most mining sites cannot convert

Voica, Panait and Iacob’s January 2026 paper in Energies is the peer-reviewed counterweight, and it is unambiguous. The redundancy gap between Tier 0/1 and Tier 3/4 is rated a critical barrier. Cooling and fiber connectivity are rated high barriers. Their conclusion is that these are requirements “that most remote mining sites cannot satisfy.”

The sites that have converted are not a random sample of mining sites. They are the ones near fiber routes, with water, with transmission capacity to spare, on land that can hold a much larger building. The public companies announcing gigawatt deals hold the best sites in the industry. Generalizing from them to the sector is a selection error, and the Texas count in Part eight is subject to exactly that bias: fourteen sites with disclosed capacity are the sector’s most visible, not its most typical.

The same paper makes the other point worth carrying: “cryptocurrency mining presents a 3–5x greater load flexibility compared to AI inference workloads.” Mining functions as a virtual battery. AI inference is a stiff load. If the grid value of flexible computation is the thing you care about, the conversion destroys it.

The capital has not been raised and the buildings are not built

VanEck’s June 2026 framework for valuing miners as AI infrastructure is the sharpest institutional skepticism available, and it comes from a firm that is long the sector. Its findings: only about 25% of leased capacity has actually been delivered, against a near-term funding gap of roughly $50 billion and long-run capital needs near $221 billion if all announced development proceeds. Matthew Sigel’s framing is that “execution, not signing, becomes the next premium,” and that companies missing construction milestones “are likely to face structural de-ratings.”

The contract values in this document are announced, not earned. A twenty-year, $9.1 billion lease produces nothing until a building exists.

The financing has macro consequences

The Federal Reserve Bank of Dallas published an analysis in February 2026 finding that AI data center financing supplies roughly $360 billion in duration equivalents in 2026 alone, steepening the long end of the curve and displacing other borrowers. It is a market-structure argument rather than a credit-risk argument, which makes it harder to wave away, and it is the highest-authority source available on what this buildout costs beyond its own balance sheets.

Moody’s has separately warned that unprecedented AI capital spending threatens the credit quality of the hyperscalers themselves. If Google’s rating is the thing making these leases financeable, Google’s rating is a variable, not a constant.

The demand forecasts carry wide error bars

Lawrence Berkeley National Laboratory’s 2025 data center energy report, published June 2026, projects data centers at 11.8% of US electricity by 2030, with a range of 9.5% to 15.3% and a reference case of 649 TWh against bounds of 521 to 843 TWh. That is a 322 TWh spread, which is the honest answer to anyone quoting a point estimate.

The World Resources Institute has documented why interconnection-queue-derived forecasts systematically overstate demand, because the same project appears in multiple queues and most requests are withdrawn. LBNL’s own 13% completion rate is the evidence.

And Morgan Stanley’s own number has moved. The widely cited 49 GW US shortfall through 2028 is the February 2026 vintage. By August 2026 the firm had revised it to a 38 GW gap for 2026 to 2028, against roughly 68 GW of need. Anyone still citing 49 GW is citing a superseded estimate; anyone citing 44 GW is citing the November 2025 one.

The demand is real, and its size is not disclosed

It is worth being precise about the thing everyone is building for, because the numbers in circulation are not company disclosures.

Anthropic’s own published commitments are two. A $50 billion US infrastructure investment announced November 12, 2025, building data centers with Fluidstack in Texas and New York, “custom built for Anthropic.” And a multi-gigawatt Google and Broadcom TPU expansion announced April 6, 2026, described by Anthropic as “a major expansion of our November 2025 commitment,” with no dollar value attached.

The widely repeated aggregate, roughly $517 billion across about 14.8 GW, originates from a September 2026 report by The Information and is a tally of announced deals rather than a disclosed figure. The reporting itself flags the caveat, noting that actual outlays depend on how much capacity is delivered and used. Treat it as journalism.

The distinction matters commercially. $50 billion is contracted intent from the counterparty’s own mouth. $517 billion is a sum of headlines, several of which are options and extensions counted at full value. A landlord underwriting a twenty-year lease should know which of those two numbers its business case rests on.

The counter-argument to the counter-argument

Daniel Batten has argued publicly through 2026 that claims of AI displacing Bitcoin mining are overstated, and that AI data centers will increasingly need mining alongside them precisely because mining supplies the interruptibility that AI cannot. His position is the substantive contrarian take in this debate and it deserves engagement rather than dismissal. If he is right, the flexible megawatt reappears as a complement to the stiff one rather than being consumed by it. The evidence in this document does not settle that question, and the twenty-year leases will settle it before the argument does.

The temptation is to conclude that the miner-to-AI conversion generalizes. It does, but not in the form usually claimed.

What does not transfer is the idea that owning a mine is most of the way to owning a data center. The peer-reviewed assessment says it is not, the cost data says it is not, and the companies themselves are budgeting $11 to $12 million per megawatt on top of what they already own.

What transfers is the sequence, and it has four steps.

• Secure the interconnection before it is worth anything. The 61-month median is the moat. Everything else can be bought.
• Discover that the building is the cheap part and the wrong part. Conversion economics only work where the site was already suitable for something it was never built for.
• Find a counterparty whose credit a rating agency will accept, or find someone to stand behind one who isn’t. This is the step that separates the companies with $3 billion of investment-grade paper from the companies with press releases.
• Sell duration, and understand you are selling the optionality with it. The premium is compensation for giving up the ability to switch.

Any market absorbing large computational load will run this sequence. What differs by market is step one, because interconnection rules, queue reform and transmission cost allocation are jurisdictional. Case Study No. 1 in this series covers how ERCOT is rewriting exactly those terms, under a Governor’s audit, while these leases are being signed against them.

The two documents describe the same asset from opposite ends. The ERCOT study is about what a megawatt earns when it stays flexible and sells demand response and ancillary services. This one is about what a megawatt earns when it stops being flexible and sells duration instead. As of late 2026 the second number is roughly three times the first, and it is contracted. Whether it stays there depends on things neither document can settle: what the grid will pay for flexibility once large loads are the majority of load growth, where hashprice goes, and whether the buildings actually get built.

Drawn from the evidence above, not from a product pitch.

The story the industry tells about itself is that the halving made mining unprofitable and the miners cleverly pivoted to a better business using infrastructure they already had.

The record says something more specific and less flattering to the storytelling. The halving did compress mining margins, further and for longer than most accounts acknowledge, to a record-low hashprice in the first half of 2026 and the first sustained hashrate decline since 2021. But the infrastructure the miners held was not an AI data center in waiting. It was a queue position, and building the data center on top of it costs roughly what building one anywhere costs.

What the successful companies actually did was financial rather than industrial. They found tenants willing to commit for fifteen to twenty-five years, arranged for a AA-rated balance sheet to stand behind those tenants in an amount matching the debt they wanted to raise, and issued investment-grade bonds against the result. Hut 8 priced at BBB− and 6.192%. That is the transformation. Not mining to AI, but equity risk to bond risk.

And when one company tried to capture that value by selling itself outright, its own shareholders refused, at 82.8% of votes cast, and were proven right within ten months. The offer would be worth $10.35 a share today. The company is worth more than that standing alone, and it signed a single customer contract larger than the entire price it turned down.

In Texas, where the capacity is, roughly four mining megawatts in five have already gone through that door. What is left behind is not a smaller version of the same decision. It is a different one, and the intuitive version of it is wrong: curtailment-exposed power does not make a mining site, because a fleet that runs only when the grid lets it never earns back the hardware. Cheap power at high utilization makes a mining site, and the curtailment right is the option written on top.

What remains unresolved is whether the buildings get built. Roughly a quarter of leased capacity has been delivered. The credit support that made the financing possible does not extend to construction, the tenant can walk if construction runs 180 days late, and the companies carrying that risk are, by their own history, operators of sheds full of ASICs. The contracts are signed. The concrete mostly is not. That gap, not the halving, is the thing worth watching for the next two years.

Every figure is drawn from public filings, named research, or Type 3’s own Texas facility dataset and can be independently checked. Where sources disagree, the disagreement is stated rather than resolved silently.