Core Scientific

IThe business

Core Scientific owns and runs data centres. At 30 June 2026 it had 11 of them, in Alabama, Georgia, Kentucky, North Carolina, North Dakota, Oklahoma and Texas.² Its customers rent space, power and cooling for racks of AI chips, and the company does not own the chips. It controls about 2.1 GW of grid power across the sites, of which it counts 1.275 GW as power a customer can use, and it has leased 590 MW of that to CoreWeave and, since 27 July, 529 MW more to AMD and one other tenant.²⁵ The year is the calendar year.

It began in 2017 as a bitcoin miner and a host for other miners. It went public by merging with a listed shell in January 2022, filed for Chapter 11 on 21 December 2022, and came out of it on 23 January 2024.³ On the way out it handed creditors new shares plus two sets of warrants, which matter below.³ The old business is now a wind-down. Hosted mining for the last customer is due to end by 31 December 2026, the remaining self-mining fleet runs for cash only, and in the first quarter the company wrote $266.5M off its mining equipment and buildings.²¹² In the second quarter it paid a $41.9M loss to cancel a contract with Block for mining machines.²

What the revenue looks like now

In the second quarter of 2026, colocation brought in $136.7M of the $164.2M total, 83%. In the second quarter of 2025 it was $10.6M, 13%.² A year ago the company was a miner. In 2025 colocation was $65.4M of $319.0M, 21%.³ The change happened in three quarters, which is why trailing figures say little. Colocation revenue was $31.3M in the fourth quarter of 2025, $77.5M in the first quarter of 2026 and $136.7M in the second.⁴

Billing capacity is the number to watch. The company bills for what it has handed over, and 395 MW was billing at 30 June, 437 MW by mid-July, against 590 MW leased to CoreWeave.²⁴ The rest is under construction, with the last of it due in early 2027.⁶ Power costs are passed to the customer without mark-up, and shown as revenue.² That inflates revenue and costs equally, and it is why I look at fixed lease income rather than revenue. In the second quarter, lease income was $98.8M fixed and $37.9M variable.²

The CoreWeave contract

This is the business as it stands. CoreWeave rents about 590 MW across five sites: Denton in Texas, 260 MW; Dalton in Georgia, 175 MW; Muskogee in Oklahoma, 70 MW; Marble in North Carolina, 65 MW; and Austin in Texas, 20 MW.⁶ The company describes it as a 12-year take-or-pay contract at a fixed price with an annual escalator, two five-year options to renew, no right for CoreWeave to end it early, and an expected profit margin of 80% to 85% on the fixed fee.⁶ CoreWeave pays for power, utilities and, up to $1.5M a megawatt, building work, which it gets back as a credit of no more than half of each month’s fees.⁶ The company puts the value at more than $10B over the term, about $850M a year of colocation revenue on average.⁶ The 10-Q gives the cash side: $6.09B of fixed payments still to be received on the leases already running, and another $3.1B on leases not yet started.²

The fixed payments do not arrive evenly. On the leases that are running, the 10-Q expects $132M in the second half of 2026, $271M in 2027, $369M in 2028, $528M in 2029 and $546M in 2030, then $4.24B after that.² That is the step-up from credits ending. It matters because the $3.3B of bonds that financed the build do not start to amortize until after rent has started and the credits have ended, which the company models as November 2029.²

AMD, and a second tenant

On 27 July the company signed leases for 377 MW with AMD itself, at Pecos and Hunt County in Texas and Muskogee in Oklahoma, and for 152 MW with an unnamed “Neocloud”, at Auburn in Alabama and Dalton in Georgia.⁵ Each runs 15 years with three five-year options. For the Neocloud sites AMD signed a credit support agreement: it can cure the tenant’s defaults and has rights over its own equipment in the building, but is not obliged to pay.⁵ The company says the five sites carry more than $14B of base contracted revenue.⁴ Delivery runs from the first half of 2027, at Pecos and Auburn, to the first half of 2028, at Hunt, with all 530 MW due by the end of 2028.⁶

AMD also has the right to lease up to 1,925 MW more through 28 December 2028, “at certain times and under certain circumstances”, which is where the “up to 2.5 GW” comes from.⁵ In return it received a warrant for 30M Core Scientific shares at $23.47, expiring 27 July 2031, which vests at 12,222 shares for each megawatt leased. About 6.5M vested on signing.⁵ The warrant is a discount on rent paid in shares that only pay if the stock does well. It also means AMD is a shareholder with an interest in the price, which I think is a point in the lease’s favour.

What it costs to build

Management puts the average cost at $11M to $12M a megawatt for the building and its power and cooling equipment, before the tenant’s own fit-out, which adds $1M to $2M.⁶ At $11.5M, 530 MW is $6.1B C (calculated). The company has not said how it will pay for it, and I take that as the main open question on the page. Its 10-Q commits about $1.0B of spending in the next 12 months, of which $264M passes through to the customer.²

Power, and where it comes from

Power is the scarce thing, so the 10 September 8-K matters. Texas’s grid operator, ERCOT, is allocating large new loads in a process called Batch Zero. Denton’s existing 297 MW is treated as long-standing load and is not in the process. Pecos’s existing 300 MW is conditionally approved, and another 300 MW is approved as studied load, but the company has not yet received its final allocation or ramp. Hunt’s planned 431 MW is conditionally approved and the company has posted collateral and begun construction.¹³ Conditionally approved is not delivered. Pecos holds 185 of the 530 AMD megawatts.⁶¹³

Since March the company has also bought power. In May it paid about $233M for land and power in Hunt County, expected to support about 430 MW.² On 13 August it paid about $444.3M for Polaris DS, next to its Muskogee site, with up to 440 MW of utility capacity, and it will pay $40M more if 40 MW of firm power arrives by the end of 2026.¹⁴ Neither purchase is leased yet. They are what lets it say it has over 2 GW of new sites under review and 1.5 GW or so of leasable power that is not yet contracted.⁶

IIThe moat

I ask two questions of any business like this. Can a rival take the customer at the same price, and would the customer want to move? For a landlord of data-centre space the answer is mostly about three things: power, time and the lease.

Power and time

A new data centre needs a grid connection, and in the main US markets those now take years. A site that already has approved power and a building shell, as many of Core Scientific’s old mining sites do, can deliver capacity sooner than a new build. That is the whole of its case. The company says it energized as many megawatts in 2025 as the rest of its public peers combined, and the deck names TeraWulf, Cipher, Galaxy, Applied Digital, Hut 8 and IREN.⁶ I could not check the claim, and the footnote does not show the arithmetic. The point it makes is fair: speed is the product.

It is not a permanent edge. Every name in that list also has mining sites to convert. Time is an advantage that shrinks when rivals catch up, and the big developers that are not bitcoin miners, the data-centre REITs, compete for the same tenants and the same grid slots.²

The lease

Once a tenant has put chips in a building, it does not leave. The company says the CoreWeave contract cannot be ended early by the customer, and CoreWeave has liens on the infrastructure, not the chips.⁶ That makes the revenue behave like debt. It also means the company cannot walk away either. If CoreWeave is late or the company is late, the contract does not simply end.

I do not count the lease as a moat in the usual sense, because it protects a single customer’s revenue and not the company’s ability to win the next one. The company’s own 10-Q says its colocation revenue comes from one customer, that its results depend on that customer’s finances, and that its practical ability to enforce recourse would depend on the customer’s financial condition at the time of any default.² In July, CoreWeave assigned some of the leases to a special-purpose subsidiary, with CoreWeave remaining a primary obligor.²

The customer

CoreWeave’s second-quarter revenue was $2,575M and its operating loss $49M. It held $5.5B of cash and carried $35.6B of debt, and it reported $103.7B of remaining contracted revenue.⁷ It is a fast-growing company with a huge order book, and it pays for that with debt. I do not think CoreWeave will fail to pay Core Scientific on a normal path. I do think the lease is only as good as the borrower behind it, and that is a risk I cannot size from outside. The new AMD leases are better on this count: AMD held $5.1B of cash and $2.4B of long-term debt at its June quarter end.¹⁵ But the 152 MW for the Neocloud are only backed by AMD’s right to step in, not its promise to pay.⁵

IIIThe competitive advantage

What Core Scientific has, other than power, is a record of building the same thing at speed. Between January and June 2026 it took billing capacity from 120 MW to 395 MW, which is the number a lender looks at.⁴ It has run high-density sites for CoreWeave since 2024 and has a customer that came back to expand the contract more than once.³ In the 10-K the company says it signed 16 MW in February 2024, 200 MW in June 2024 and exercised options to reach 590 MW by early 2025.³ That history is worth something to the next tenant.

The two customers come with different terms. The table puts them side by side, with the figures the company gave.

Where it is weaker

The AMD rent per megawatt looks about 30% higher than the CoreWeave rent C (calculated). I do not read that as a pricing win until I see the lease. A triple-net lease puts more of the cost on the tenant, and the AMD figure may include rent for years with escalators. The company also gives the AMD figure as “potential” base contracted revenue, which is a softer word than backlog.⁴ There is a second weakness. The company has so far funded the CoreWeave build largely with the customer’s money, up to $1.5M a megawatt, and the AMD build with its own. A business that grows by borrowing $6B against a tenant’s signature is a different business from one that builds with the tenant’s cash, and the track record is thinner.

IVManagement

Adam Sullivan, 34, has been chief executive since August 2023 and sat on the board since the company left bankruptcy.⁹ Before that he was a managing director at XMS Capital Partners, the advisory firm that represented the company’s predecessor in the 2021 deal that made it public.⁹ His chief financial officer, James Nygaard, 52, joined in March 2025 from the same firm and was the predecessor’s finance chief in 2021.⁹ The chief legal officer, Todd DuChene, has been with the company since 2019.⁹ The senior team is three people, and two of them come from the same adviser. It has done a great deal in two years, and I could find no one on it who has run a data-centre landlord through a downturn.

How they are paid

Sullivan’s total pay was $22.9M in 2025 and $41.9M in 2024, nearly all of it stock awards, $20.0M in 2025.⁹ Nygaard received $9.7M and DuChene $7.4M.⁹ The median employee earned $134,231.⁹ At today’s price those awards cost shareholders something real: stock pay across the company was $98M in 2025 and $94M over the last twelve months, which I subtract from owner cash C (calculated).³²

What the pay rewards is more interesting than its size. The 2025 bonus paid 200% of target to every eligible executive. It was judged on three equal parts: “ideal hashrate”, a bitcoin-mining measure, which came in above 98% against a top mark of 97%; controllable expenses at 92% of budget, against a top mark of 95%; and the committee’s view of individual contribution.⁹ Two of three measures belong to the business the company is leaving. The long-term awards in 2025 were 67% performance units, up from 25%, and they are judged on shareholder return against the Russell 2000, megawatts energized and 100 MW of new customers.⁹ Those are better targets: they pay for power and tenants. At the first measurement, as I read the table, Sullivan earned 741,545 shares.⁹ What none of the targets measures is return on the capital the company is about to commit, $6B or more. That is the number I would want the pay tied to.

Their own money

At 23 March 2026 Sullivan held 1,669,234 shares, plus 114,364 shares from Tranche 1 warrants and 741,545 shares coming within 60 days, in all under 1% of the company.⁹ His 1.67M shares are worth about $22M at $13.25 C (calculated), against about $59.5M of stock awards granted to him in 2024 and 2025 C (calculated). DuChene held 1.11M shares and 517,133 warrants, and Nygaard 475,205 shares.⁹ The directors’ holdings are small: Jeff Booth 68,241 shares and Jordan Levy 179,441.⁹ I would not call it a management with its savings in the stock. It is paid in stock, which is not the same thing.

What they did with capital

The record has to be read in order. In 2022 the company filed for Chapter 11.³ On the way out, in January 2024, creditors took new shares plus warrants and rights to payments.³ In 2024 it sold $460M of 3.00% convertible notes, convertible at $11.00, and $625M of 0% notes, convertible at $22.49.³ In July 2025 it agreed to be bought by CoreWeave in an all-stock deal, and on 30 October 2025 its own holders voted it down. It spent $21.6M on advisers for that deal.³¹¹ Since then it has raised and committed far more than any year in its life: a $1.0B bridge term loan in March, repaid in May from $3.3B of secured notes at 7.75%, a $100M revolver and $500M letter-of-credit line, $233M for Hunt County and $444M for Polaris DS.^2161714 All of it is paid for with borrowed money, and none with the company’s own earnings, because it has not had any.

Controls and the board

On 2 March 2026 the company said its accounts for 2024 and the first three quarters of 2025 could no longer be relied on. It had kept on its books the cost of buildings that it had already decided to demolish, so property was overstated.¹⁰ It restated the periods and disclosed a material weakness in internal control that it has not yet fixed.³ It had dismissed its auditor, Marcum, on 12 March 2025, and the board has proposed KPMG as auditor for 2026.⁹ The restated 2024 operating loss is $142.1M, against $19.2M first reported C (calculated).¹⁸³ I read a restatement of this kind as a warning about how fast the company has changed more than as a sign of fraud, but it is the sort of event that precedes more.

Holders have pushed. Two Seas Capital, then a 5.5% holder, agreed on 18 February 2026 to a cooperation deal: three new independent directors, a one-year standstill, and the chairman, Jordan Levy, not standing again in 2026.¹⁹⁹ Mark Adams, a former president of Micron, joined on 29 July.²⁰ The shareholders also decided the biggest question the company has faced, the CoreWeave sale, against the board’s recommendation. I count that as a sign the owners are paying attention.

VFinancials

A ten-year record would mislead here. The company in its current form began in 2024: it left Chapter 11 in January, and its filings before that describe a bitcoin miner and host with a different capital structure.³ The numbers below start in 2022 for that reason, they are restated where the company restated them, and the pre-2024 years are shown mainly so that you can see that nothing before 2025 predicts what comes next.¹⁰³ The figures for 2024 and 2025 are those in the 10-K filed on 2 March 2026, after the restatement.

Revenue fell from $511M in 2024 to $319M in 2025 as mining shrank, and over the last twelve months it is $440M C (calculated). Gross profit tells the shift better. It was $38M in 2025, a 12% margin, and $125M over the twelve months, 28%, as colocation at a 58% gross margin replaced mining at a negative one C (calculated).² In the first half of 2026 self-mining lost $29.2M at the gross line.² The operating loss over the twelve months of $561M includes the $266.5M impairment and a $41.9M termination loss, and neither will repeat C (calculated).²

Net loss is the wrong number

The net loss over the twelve months is $1.43B, and about $821M of it is the rise in the value of the warrants, which are carried as a liability and marked to the share price.²³ When the shares rise the company reports a loss. In 2024 the same effect was $1.37B on a $1.44B net loss, which leaves about $69M excluding it C (calculated).³ I ignore it, and look at cash.

Owner cash flow, operating cash less capital spending less stock pay, was −$549M in 2025 and −$1.05B over the twelve months C (calculated). Operating cash is flattered: it includes the customer’s prepayments, $98M of added deferred revenue in the first half of 2026, and a $208M sale of bitcoin the company counts as operating.² Capital spending of $954M in the first half does not include another $233M for land and power. Of the first-half capital spending, $181M was funded by CoreWeave.² The business has so far spent about $1.1B more than it has earned, in a year when revenue was $440M.

The balance sheet

At 30 June the company held $1,769.7M of cash and $49.7M of bitcoin, and a further $781.7M of restricted cash.² Restricted cash includes $344.8M in a reserve for the secured notes and $120M in escrow for Polaris, which has since closed.² Against that it owed $3.30B of secured notes, $460M and $625M of convertibles, and $654M of customer prepayments recorded as deferred revenue.² Stockholders’ deficit was $2.42B, which reflects the $1.98B warrant liability and an accumulated loss of $5.65B.²

The secured notes are the part to understand. They were issued by a subsidiary, Core Scientific Finance, and are secured on the five CoreWeave sites and what they earn.² The CoreWeave rent is paid into a lockbox tied to them.⁶ That is good for the lenders and, for the same reason, it means the parent does not see that cash freely. Interest on $3.3B at 7.75% is $256M a year C (calculated). The principal comes due as a balloon of $2.73B in 2031 after the amortization.² The company has $1.0B of building work committed over the next year, and no financing announced for the AMD sites.²

What rent is coming

For the leases already running, the 10-Q gives the cash that is contractually due. The table is the most reliable forecast on the page, because it is the company’s own count and is not an estimate of mine.

The warrants and the share count

This is the part most likely to be misread. At the end of June the company had 319.6M shares in issue.² But 96.4M Tranche 1 warrants, exercisable at $6.81 until 23 January 2027, and 6.6M Tranche 2 warrants at one cent were also outstanding.² The 10-K allows exercise without paying only for the Tranche 2 warrants, so exercising the Tranche 1 warrants would bring in about $656M of cash C (calculated).³ By 23 July the company counted about 508M shares on a fully diluted basis: 321M shares, 95M Tranche 1 warrants, 6M Tranche 2, 16M restricted and performance units, 42M from the 2029 notes and 28M from the 2031 notes.⁶ That count leaves out the AMD warrant for up to 30M shares at $23.47, of which 6.5M had vested.⁵ My own share count is in the Valuation.

VIValuation

First, how this goes badly wrong

I start with what would destroy the investment, and work back. There are three ways.

The money does not arrive. The leases are signed. The buildings are not. Building 530 MW at $11M to $12M a megawatt costs about $6.1B C (calculated), against $2.6B of cash and restricted cash and no financing announced.⁶² If lenders will not fund it at a sensible rate, the company either sells shares at a low price or delays, and a delay costs interest and rent at the same time. The customer stumbles. CoreWeave pays 77% of revenue and has $35.6B of debt, and a stress at CoreWeave would reach the secured notes first and the shares last.²⁷ The power does not come. Pecos holds 185 MW of the AMD capacity and its final grid allocation was not in hand on 10 September.¹³ The tenant pays only when the megawatts are delivered.

In my bear case the first two both happen a little. The leases are built, but a year late, at a cost of $12.0M a megawatt, with rent 8% under the average and a half-point higher cost of money. On those numbers the business is worth about $1.6B and the debt ahead of the shares is larger, so the shares are worth nothing. Equity cannot be negative, so the bear on this page is $0.

What an owner can take out

An owner here cannot take anything out for years. In the twelve months to June the company reported a net loss of $1.43B, and its owner cash flow was −$1.05B C (calculated). The first is mostly warrants, and the second is the real one: it spends more than it earns, and it will until the AMD sites are built and the CoreWeave credits end. On my base case the business generates cash after the capital it needs from 2030, and after that it is repaying debt, not paying owners. That is the sense in which reported earnings, which have swung from a $2.1B loss to a profit and back, say little. I value lease income after the cost of running the buildings, after head-office cost and stock pay, after maintenance, and after tax.

What I cannot judge

Four inputs are beyond my reach, and I show a range for each rather than a false precise figure. I cannot see the AMD lease: its rent per megawatt, its escalators and how much it costs the tenant, so I use the company’s total of $14B over 15 years, which is $1.76M a megawatt a year C (calculated). I cannot see how the build will be financed, so I price all capital at one rate and show what a different rate does. I cannot judge CoreWeave’s long-term credit. And I cannot tell how much of the 1,925 MW AMD may lease. The company’s own wording is “at certain times and under certain circumstances”.⁵ Because the last one is a choice and not a forecast, I use 340 MW in the base, none in the bear, and 1,500 MW in the bull.

How I value it

I count every lease dollar and every building dollar from July 2026, by site, and discount them. CoreWeave: the fixed rent in the 10-Q, plus the $3.1B of leases not yet started, at an 82% margin. AMD and the Neocloud: 530 MW at $1.76M of rent per megawatt, an 89% margin, $11.5M of cost per megawatt, with deliveries on the company’s schedule, and 15-year terms. Further leases: 340 MW at $1.70M, starting in 2029. After each lease ends I assume a renewal at 55% of the last year’s earnings, valued at a 10% yield. I take off head-office cost of $80M plus $0.085M per megawatt, which is $175M to $205M a year; $0.06M a megawatt of upkeep; $40M a year of start-up costs for two years; a 24% tax after bonus depreciation on 60% of new building cost and $728M of losses; and the $444M for Polaris. I add the cash held and subtract the debt. I discount CoreWeave’s rent at 8.5%, AMD’s at 7.5% and everything else at 8.5% to 10%, which I explain below.

The rates need a reason. The company pays 7.75% before tax on bonds secured on the CoreWeave leases.² After tax that is about 5.9%, and the equity behind those bonds costs more, so a blend near 8.5% is where I start for CoreWeave.² AMD’s credit is better, so I use 7.5% for its rent. I use 10% for the renewal value because it is furthest away. Nothing in this is observed in the market, and the whole answer moves with it, as the grid below shows.

The bridge to a share price

The value of the business at $13.25 is not the same as the value of the shares. The bridge below shows my base case. I start with the discounted cash, add the cash the company holds, subtract the debt, add the cash the warrants bring in, and divide by the shares that would exist if the value were what I say.

Two things in the bridge are easy to miss. The 2029 convertible notes only turn into shares if the value is above $11.00, and at $13.25 they would, so at the market price there are 41.8M more shares and $460M less debt. The AMD warrant, at $23.47, only counts in my bull case. The share count of 438M in my base is below the company’s 508M because the base value is under the conversion prices of both notes.

Three cases

I give the three cases 25%, 50% and 25% only to show a weighted value, $10. I do not use the odds elsewhere. The weighted value is above the base because the bull is further from it than the bear is.

The price against the range

At $13.25 the shares are 51% above my base value of $9 and 40% below my bull of $22. At that price the whole business, with debt and the warrants counted, is worth about $7.0B to the market C (calculated).¹ That is $4.0 for each dollar of the lease earnings I expect in 2030 C (calculated), before the $6B still to be spent on the AMD sites. The market is paying for something between my base and my bull. The next section says what.

The shares were $25.59 on the day the fund’s 13F counts them, 30 June 2026.⁸¹ They are 48% lower now. In that time the company signed the AMD leases, closed Polaris and opened a revolver. I do not know the cause of the fall and will not guess.

What the price needs

I turned the question round. Holding my other assumptions, what would make my base equal $13.25?

The point of the table is the third row. I expected more leased megawatts to be the main lever, because that is the pitch. At the economics the company has disclosed, and at my cost of money, extra megawatts add little: a lease that earns 13% on its cost, before head office and tax, earns about what the money costs. What moves the answer is the rent per megawatt and the cost of capital.

How sensitive it is

The margin of safety

I set it at $5, 60% of my base value. I cannot use the usual rule, which divides my bear value by 0.75, because the bear here is zero. I chose 60% so that a price at that level leaves room for the base to be 40% wrong, which on this page is a modest allowance, given the grid above. The shares last closed below $5 in June 2024 C (calculated).¹

Quality and price are two questions

The business that exists today, 590 MW of long, take-or-pay leases to one customer, is a decent one, with a margin I estimate near 80%, and the AMD leases look better. The financing is what makes it hard to value. Quality and price are two questions. The quality is real and still being proved; the price is 51% above my base value. A fair business bought above its value is a poor purchase, and I would put this one in that group on my numbers, though I would not call my base certain.

The next best use of the same dollar

On this site the nearest comparable is Oracle, which rents the same sort of capacity from landlords like this one, and Vertiv, which makes the power and cooling equipment inside such buildings. Both are far larger, with much stronger balance sheets, and neither depends on one tenant for most of its revenue. For those reasons alone I ask for a bigger gap between price and value here. A Treasury bill carries no building risk at all, and it is the yardstick for the part of the return that is only waiting.

Where this could be in 5 and 10 years

This is the value of the shares, in the page’s own model, at later dates: the value at that date of all the lease cash after it, less the net debt then, per share of today’s count, in the dollars of that year. It includes the AMD and further leases at the case assumptions. Net debt grows at 8.5% and every shortfall is borrowed.

For the base to hold at five years, the 530 AMD and Neocloud megawatts must be built by the end of 2028 at about $11.5M each and rent at about $1.76M a megawatt, and CoreWeave must be billing all 590 MW from 2027. At ten years it also needs 340 more megawatts leased by 2031, the first CoreWeave leases renewing at about 55% of the last rent from 2037, and money that still costs about 8.5%. The bull differs by leasing 1,500 MW and renting at $1.80M, which makes the 2036 value near $52. The bear differs on several things at once: no further leases, the build a year late at $12.0M a megawatt, rent 8% lower and money half a point dearer. On those the debt ahead of the shares is not covered.

The numbers that drive it are rent per megawatt, the cost of money, the build cost and the megawatts leased. The most likely way my base is wrong is the financing: I count today’s shares, and if the company sells stock at a low price to fund the build, each share is worth less than I show. These are estimates of value, not forecasts of the share price, and the market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three things cut in the company’s favour and get little attention. First, the dilution from the warrants comes with cash. If the 96.4M Tranche 1 warrants are exercised before they expire on 23 January 2027, the company receives about $656M C (calculated).² At $13.25, with the strike at $6.81, there is every reason to. That is a tenth of the build, arriving without a bond. Second, the CoreWeave rent goes into a lockbox that serves the $3.3B notes, and the notes do not begin to amortize until after rent starts and the early credits end.⁶² Lenders carry the build, and the first repayment is about three years away. Third, AMD has put its own interest into the lease. The warrant it holds is struck at $23.47, 77% above today’s price C (calculated), and pays only if the stock does well.⁵

The bull case

In plain words: AI buyers will pay for power and buildings faster than anyone can supply them, and Core Scientific has both, already approved, in a handful of places. It has a customer for 590 MW, a second for 530 MW and a right of AMD’s to lease up to 1,925 MW more, in a business where the land and the power are the constrained asset.⁵ The company says contracted revenue is already over $24B, and that AMD’s expansion could take customer-contracted capacity above 3 GW.⁴⁶

The people closest to the company say it in their own words. Adam Sullivan said at the AMD announcement on 28 July 2026 that the company was “proud to establish a strategic relationship with AMD”.²¹ Michael Intrator, the chief executive of CoreWeave, said on CNBC’s Mad Money, in an interview filed on 8 July 2025, the day after CoreWeave announced its offer to buy the company, that the two had worked together since 2018 and that the deal would “bring them under the same tent”. He put the contracted capacity at about 850 MW.²² That is higher than the 590 MW the company shows, and the company’s own deck says “~800MW gross”, so the gap is a matter of how the megawatts are counted.⁶ The holders did not agree on price: they voted the offer down on 30 October 2025.¹¹

The fund Situational Awareness LP held 26,008,473 shares at 30 June 2026, according to its 13F filed 14 August 2026, and I take no view from that.⁸ Among analysts I found three notes, all secondhand and not opened by me: Needham raised its target to $35 from $29 on 28 July; Canaccord went to $36 from $30; and KBW raised its rating with a $28 target.²³²⁴²⁵ All three followed the AMD announcement, when the shares were far higher than now.¹

What the podcasts and interviews say

I looked for what well-known investors and the chief executive have said in public interviews and podcasts about Core Scientific, its CoreWeave contract and the move from bitcoin mining to AI hosting. I found no one who has discussed the company in an episode I could open and check.

All-In. The episode that sits closest to the subject is “CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer”, published in mid-October 2026. A podcast-app recap names Chamath Palihapitiya, Jason Calacanis, David Sacks, David Friedberg and the guest Ben Horowitz as the speakers; that recap is dated 17 October 2026.²⁶²⁷ I searched the full transcript text for Core Scientific, bitcoin miners, hosting, colocation and landlord. None appears. What does bear on this page is two comments. At 00:38:19 the transcript has “power will be the moat”, and at 00:39:27 an argument for owning the land, power and shell because the asset is so constrained. The transcript has no speaker labels and I did not hear the audio, so who said each line is my inference from the recap’s list of speakers and the way the lines read: I infer the first two are Palihapitiya’s. At 00:23:56 the transcript has lending against a data-centre project “depends more on the credit of the tenant”, which I infer is Sacks.²⁶

I weigh these against the filings. On power, the filings agree: the company’s whole case is approved megawatts, and it paid $233M and $444M this year for land and power, mostly without a lease.²¹⁴ But “constrained” also means Pecos, with 185 of the AMD megawatts, had no final ERCOT allocation on 10 September, so the scarce thing is also the uncertain thing.¹³ On tenant credit, the remark helps the AMD leases, because AMD is the stronger name, and it hurts CoreWeave’s, which is 77% of first-half revenue behind $35.6B of debt.²⁷ The Neocloud leases get AMD’s right to cure, not its guarantee.⁵ It is also the best argument for why the $3.3B notes cost 7.75%.

Others. I found no remarks by Calacanis or Friedberg on Core Scientific, and none by Gavin Baker or Brad Gerstner; my searches returned nothing that mentions either with the company. Sullivan has given interviews, including CNBC on 13 September 2024, 11 minutes 22 seconds long, on why the company embraced AI compute, but I could not retrieve a transcript and will not paraphrase what I have not read.²⁸ I also could not open a transcript of the 28 July 2026 earnings call, so I rely on the company’s filed presentation.

Against reality

Today colocation revenue is $136.7M a quarter, about 83% of the total, and 437 MW of the 590 MW is billing.²⁴ The AMD sites have not been built and no financing for them is announced. If they are delivered on the company’s schedule, rent is near $1.76M a megawatt, the money costs about what I assume, and a few hundred megawatts more are leased, my base is possible: $9 a share. The price needs one of two things on top of that: rent of $2.22M a megawatt, 26% above the average the company implies, or money about 1.6 points cheaper across the board C (calculated). That is the distance between today’s numbers and the case.

What a 10x would take. Ten times $13.25 is $132.50 a share, about $67B of market value on 508M shares C (calculated). To reach that by 2036 the company would need all of AMD’s 1,925 MW and its 530 MW leased and built, about 3 GW in all, which is $28B more of building at $11.5M a megawatt and about $4.5B a year of lease earnings, with debt of roughly $30B ahead of the shares E (estimate). The market would then have to pay about 21 times those earnings for the business E (estimate). A tenfold return in ten years is 26% a year. It is also 2.6 times my bull value for 2036 of $52. Few companies of this size have done it, and the ones that have done it did not carry this much debt on the way. For Core Scientific from here, tenfold is not plausible on my numbers, and I would not build a case on it.

Where I come down

The lockbox and AMD’s warrant moved me, and so did the $656M of cash that dilution brings. The price, the commitment of $6B not yet financed, and a restatement did not. What would raise my values most is the AMD build financed at or below 7.75% without selling shares, with Pecos power confirmed, and CoreWeave’s debt rising slower than its contracted revenue. My values stay as they are.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• One customer. CoreWeave was about 77% of first-half revenue and 100% of colocation revenue.² Some leases now sit in a special-purpose subsidiary, with CoreWeave still a primary obligor.²
• Cash is ring-fenced. The notes are secured on the CoreWeave sites and their rent goes to a lockbox. Breach of the indenture or a customer default hits the secured lenders first. On a change of control the issuer must offer to buy the notes at 101%.²⁶
• Dilution. 508M shares fully diluted against 321M outstanding, and the AMD warrant for up to 30M more at $23.47.⁶⁵ The 96.4M Tranche 1 warrants expire 23 January 2027.²
• Borrowing conditions. The new revolver can be drawn only while the company’s market value is at least $3.0B, and liquidity must stay above $150M.¹⁷
• Reporting. The company restated 2024 and the first three quarters of 2025 and reports a material weakness in controls.¹⁰³
• The wind-down. Self-mining lost $29.2M at the gross line in the first half and the company took a $266.5M write-down; it holds bitcoin worth $49.7M.² Equipment held for sale is $13.8M.²
• Costs. Tariffs have raised equipment cost on the projects it funds itself, and some equipment has lead times over 12 months.²
• People. The senior team is small and drawn from one adviser; the chairman, Jordan Levy, said he would not stand again in 2026.⁹¹⁹

Geopolitical risk, and what it means for the value

Every site in the filings is in the United States: Texas, Oklahoma, Alabama and Georgia. There is no foreign customer, no variable-interest structure, no currency peg, no capital control and no state owner. The exposure is to rules and supply chains that cross borders. In Texas, 295 of the 530 AMD and Neocloud megawatts sit at Pecos and Hunt, and a state law passed in June 2025 makes sites above 75 MW pay toward transmission and be able to curtail load or run on backup power in a grid emergency.³ ERCOT’s Batch Zero decides the final power at Pecos, where 300 of 600 megawatts is still “studied load”.¹³

Equipment is imported. The 10-K says much of it is made outside the United States and that tariffs have, since 2019, added up to 25% to mining equipment.³ The 10-Q says the company carries tariff cost itself on the sites it funds.² CoreWeave’s 10-Q warns that US export controls on advanced chips could change what it can buy and deploy, which would reach this company as delayed conversions, not as lost sales abroad.⁷ Bitcoin rules matter little now: the self-mining business is closing and the company holds $49.7M of bitcoin, about $0.11 a share.²

Each case in the table is the base case rerun with one change. They are not forecasts, and I do not add them up. The worst single hit is CoreWeave, which is a counterparty risk more than a geopolitical one, but its chips, its debt and its sites all depend on the same trade policy. What is in the price: the shares at $13.25 are already 51% above my base, so the market is not paying for these risks at my numbers. If any happened, the gap between price and value would widen. What is not in my base: any further tariff, a cut at Pecos, or a CoreWeave stumble. What is: the build cost the company quotes today, the ERCOT approvals it holds, and AMD’s lease terms.

Core Scientific short report: Culper Research, March 2022

On 3 March 2022 Culper Research, a short seller, published “Core Scientific, Inc. (CORZ): Rigged Deals”. It says “We are short Core Scientific”, and its terms let it be long, short or neutral afterwards. It gains if the price falls.²⁹

The report says Core “wildly oversold both its mining and hosting businesses” (p. 2, ¶1).²⁹ Its centre is Blockcap, the miner Core bought on 30 July 2021.³⁰ Culper says Blockcap was assembled in December 2020 from five insider-managed entities, in shares it values at $58.5M, then sold to Core seven months later for $1.46B (pp. 2, 5 and 6).²⁹ I leave out its allegations about named people, their associates and other companies; no court or regulator finding about Core states them, as far as I read.

The company’s reply. I found none: “Culper” is in none of the 30 filings, letters and court papers I searched, and I found no press reply. I could not read the March 2022 call transcript.

What is established. Culper’s $58.5M reproduces from the prospectus as my sumC (calculated): 121.9M Blockcap shares times $0.48, the price it uses for the shares issued in the December 2020 deals.³⁰ The final prospectus gives $1.46B (shares at $19.33) in its summary and $1.13B (shares at $16.18, less $60.5M for old contracts) in the pro forma note; both used an independent valuation specialist, and the August 2021 S-4 had only the first.³⁵³⁰ The S-4 called a co-chair Blockcap’s “controlling shareholder”; the final text says about 12%, and on 4 October 2021 the company told SEC staff he held no “controlling financial interest”.³⁵³⁰³⁶

The bankruptcy. Core filed for Chapter 11 on 21 December 2022 (S.D. Tex., No. 22-90341) and emerged on 23 January 2024.³⁷³⁸ Its filing-day release named “the prolonged decrease in the price of bitcoin”, electricity costs and hosting customers that did not pay.³⁹ The sworn first-day declaration lists seven factors: bitcoin’s price and power costs, over $200M of construction commitments, about $275M of equipment debt, and three others, one a lender’s acceleration that put over $550M of convertible notes in default.³⁷ Some are sector events and some company choices; both documents are the company’s own account. Neither mentions the report, the lock-up or the Blockcap price, and no court has ruled on Culper’s claims, so I draw no link either way.

Lawsuits. Pang (W.D. Tex., No. 1:22-cv-01191, filed 14 November 2022) alleges Core failed to disclose that its decision to pass power costs to customers made it vulnerable to litigation and that some clients had breached contracts. Core’s account of it does not mention Blockcap or Culper.⁴⁰³⁴ I could not open the docket, so the rest is from Core’s filings. The plaintiff dropped its claims against Core. By April 2024 the court had dismissed the rest without prejudice, the fraud claim for failure to plead intent and loss causation; the plaintiff refiled in June 2024.³⁴⁴¹ That rules on pleading, not truth. On 5 November 2026 Core said the parties had agreed to resolve it, subject to court approval, with an immaterial Core payment toward indemnification.⁴¹

Ihle (Del. Ch., C.A. No. 2023-0759-LWW), brought by a SPAC shareholder against the SPAC’s directors, officers and sponsor, not Core, alleges the merger proxy was deficient on, among other things, how Legacy Core was valued; the defendants deny it. A court notice of 19 December 2025 describes a $14.75M settlement with a final-approval hearing on 17 March 2026; I could not find the outcome.⁴² Core paid $0.5M toward indemnification.⁴¹ The legal notes I read name no SEC action.

Where it touches my model. No allegation feeds an input. The 2021 price, the lock-up and 2022 mining set none of revenue, margin, capex or share count: the old stock was cancelled at emergence and exchanged for 21.0% of the new shares plus warrants, and my share count is of the new securities.³⁸ Bitcoin held is $49.7M; at zero my base is $8.6, not $8.75.² The one channel is trust, through the cost of money: half a point more takes my base from $8.75 to about $7.5E (estimate). I see no evidence to apply it.

So what. For someone valuing the company, the report adds no number to my model, so my values stay. Culper gains if the price falls, and management is paid mostly in stock.⁹ Culper is right if the 2021 Blockcap price reflected insiders’ interests more than value and the SPAC deck overstated mining and hosting. Core is right if the terms were arm’s length and the deck was fair when made. The filings support parts of each and no court or regulator has decided either. The business, leases and debt today can be checked; the fairness of the 2021 price cannot be from outside. Watch the Pang settlement approval, with no date I could find, and the 10-K for 2026, due by 1 March 2027, 60 days after year end for a large accelerated filer.³