WhiteFiber

IThe business

WhiteFiber rents out the things that AI work needs: buildings with power, and graphics chips. It does that in two ways. It owns data centers and leases space and power to other companies, which it calls colocation. And it runs its own GPUs, the chips behind AI models, and rents them out by the hour, month or year, which it calls cloud services.³ In the first half of 2026 colocation was $9.5M of revenue and cloud $40.6M R (reported).³

It is a young company with a short record. It was set up in October 2023 inside Bit Digital, a bitcoin miner, which launched GPU cloud services in January 2024 and bought a Montreal data center operator, Enovum, in October 2024.² It listed on Nasdaq on 7 August 2025 at $17.00 and sold 10,781,250 shares for $183.3M before costs C (calculated).¹¹³ Bit Digital kept 59.7% of the shares.⁶ So when I say history, I mean two audited years, 2024 and 2025, plus two quarterly reports this year. A ten-year record does not exist, and I do not pretend to one.

All figures are in US dollars, the company’s reporting currency, under US GAAP. The company is incorporated in the Cayman Islands and based in New York. Its fiscal year is the calendar year. It has not yet reported the third quarter; that report is due in mid-November. This page uses the latest filings: the 10-Q for 30 June 2026 and the 8-Ks and proxy through 21 October.

Colocation

The data center business is the larger idea. Four sites matter, and the table lists them.

The site that matters is NC-1, a former textile plant in Madison, North Carolina, bought in May 2025 for $45M with extra payments of up to $8M, and $200,000 a megawatt above 99 MW up to $5M, if Duke Energy delivers the power.³ Duke has completed the work to deliver 54 MW gross, and the company says the site may reach up to 200 MW over time.⁹ In November 2025 WhiteFiber signed a ten-year agreement with Nscale, a young UK AI infrastructure company, for 40 MW of IT load, about $865M in all, including yearly rent increases and one-off installation and build charges. Power and some other running costs are passed through to the tenant.³² Billing was first expected in June 2026, then in the third quarter, and the September presentation says initial billing has begun with the full ramp still subject to construction and customer acceptance.²³⁹ My model counts the delay.

Two more things shape this business. First, the contract is one tenant. At 30 June the company had $1,008M of future revenue it expects to recognise, of which $933M is colocation, and the Nscale contract and the Cerebras lease are most of that R (reported).³ Second, the company is already buying more land. On 16 August it agreed to buy two properties in Yadkin County, North Carolina, for $60.0M, each expected to have at least 30 MW of power, with up to 69 MW more each possible within seven years.¹³ It has not said what it will build there, who will rent it, or what it will cost.

Cloud

The cloud business began first and looks different. WhiteFiber buys or leases Nvidia GPU servers, houses them in leased space, mostly in Iceland, and rents them to AI developers. In 2024 one customer, which the filings call the Initial Customer, was 96.6% of revenue, and in 2025 it was 70.7%.² In the second quarter of 2026 that customer’s agreement ended. WhiteFiber booked a $12.3M termination fee as revenue, later raised to $15.7M and unpaid at the filing date, wrote off $2.2M of bad debt, and moved the freed GPUs to other customers.³ The company says it has seven customers now, with the largest at 63% of first-half revenue, and that new orders signed since January total $635.8M over periods of six months to three years.³

Large new contracts follow, though not yet in revenue. In May 2026 it agreed a five-year, $160M-plus deal for GPU capacity in the Paris region with a customer it calls an investment-grade technology company; service was due in July and is now expected in September, subject to delivery of equipment.³ The September presentation lists more than $550M of cloud contract value since May, including a roughly $108M deal for Nvidia’s next-generation Vera Rubin chips that is not expected to start until the second quarter of 2027.⁹ Contract value is what a customer agreed to pay if everything is delivered. It is not revenue, and the 10-Q’s own count of firm future cloud revenue at 30 June is $75.1M R (reported).³

What it costs to be this business

The balance sheet changed completely in a year. Net property and equipment went from $89M at the end of 2024 to $337M at the end of 2025 and $651M at 30 June 2026, most of it NC-1 under construction.²³ Debt went from nothing to $83M of loans and a $230M convertible note in six months, and then, in August, to a different shape: new $310M notes, $198M of old notes exchanged, and $6.3M new shares.³⁵ I come back to this in Financials. For now the point is simple. The company is not short of contracts. It is short of cash to build what the contracts need, and it is getting that cash on terms that carry real costs for shareholders.

IIThe moat

A durable edge keeps others from taking a business’s profit. I do not think WhiteFiber has one yet, and I would sooner say that and show what could become one.

What could protect it

• Power. Grid power is the scarce input for AI buildings. NC-1 has 54 MW delivered by Duke Energy, a utility agreement for 99 MW, and a study for another 200 MW, at about $0.05 a kWh.⁹ Power rights bought with a building are hard to copy quickly. That is the strongest asset here.
• Speed. The company says its retrofit of MTL-3 took six months against 18 or more for a typical new build, at a cost under C$9M a megawatt.⁹ Its general claim is $8M to $10M of build cost per gross MW, up to 40% below new construction.⁹ These are the company’s numbers and I have not seen them tested on a large site. NC-1 is the test: 54 MW was due in 2026, and billing slipped by at least a quarter.
• A signed anchor. Nscale’s ten-year contract gives about $865M of revenue if it is paid, which is long enough to finance the building against.³

Where it is thin

The first problem is that the anchor is a single tenant that is itself raising money. Nscale is private. On 28 September it announced $3.36B of convertible financing led by Third Point, with $2.36B at closing and $1B from Nvidia expected in mid-November, and says it has more than $103B of contracted value.¹⁴ I could not find a credit rating for Nscale. WhiteFiber’s presentation calls the NC-1 tenant an investment-grade offtaker; I cannot verify that, and I treat Nscale as a young company with strong backers and large promises of its own.⁹ If Nscale cannot fill 40 MW with its own customers, WhiteFiber is the landlord of an unfilled building with a ten-year contract that is only as good as the tenant.

The second problem is that cloud is a price-taker. Renting GPUs is a market with larger rivals, and the 10-K names CoreWeave, Crusoe, Nebius and Lambda.² The company changed the useful life of its cloud equipment from three to five years from 1 January 2025, which lowers depreciation and raises reported profit.² In the same period it sold 126 H200 GPUs for $26.1M, a gain of $1.8M over book.³ I read that as a sign the older chips still have a buyer, but the useful-life change is a choice and not a fact about the chips.

The third problem is that the record is short. The company’s main customer left in 2026. Two other customers ended contracts in the same half. The company’s own filings list the discontinuations at about $5.1M of lost first-half revenue.³ I would not call a business with this churn durable until a second year of renewals shows otherwise. The table of firm future revenue at 30 June shows what is truly locked in: $57.7M for the rest of 2026, $136.7M in 2027, then about $95M a year to 2030 R (reported).³ That is all colocation after 2028. There is no cloud revenue under contract past 2028 in that table.

IIIThe competitive advantage

The question is what WhiteFiber can do that its rivals cannot, and for how long. The company’s own list of rivals shows who it is up against.

It names the largest data center landlords in the world and four GPU clouds that raise much more capital than it can. Against them it claims small, fast retrofits near cities with power already in place, and a team with about 15 years of experience each in the industry.⁹³ A retrofit is a clever way to win time. It is not a barrier, because a rival can also buy an old plant. What is hard to copy is a power contract, and WhiteFiber holds some. The question is whether it holds enough of them, and whether it can pay for them.

The financing edge, or the lack of one

A landlord’s cost of money is its advantage. Data centers last decades, and the best owners fund them with cheap, long debt secured on tenants with high credit ratings. WhiteFiber has done the opposite so far. It has used a bridge loan from its parent at 9.5%, with a payment that must total at least 1.1 times what it borrowed. At 30 June the effective rate was 50.6% because the loan was short.³⁶ It repaid $50M of that bridge in August and September, leaving $30M outstanding.⁶ It is also negotiating a project loan on NC-1 that it says would return a significant part of the capital it has put in.³ If that closes at a low rate, the picture improves a lot. If it does not, the cost of money is the most likely way this thesis fails.

The cloud side

In cloud the company says it is “capital-light” on its newest deals: the Paris contract is supported by customer prepayments and project-level financing, and the roughly $17M Hyperbolic and Modal Labs contract uses GPUs it already owns.⁹ That is the right instinct, because renting chips only pays if you do not have to keep buying new ones. But the same presentation says the VR200 deal means “WhiteFiber’s first Vera Rubin deployment”, so the company is still moving across chip generations with each deal.⁹ Whether a sixth or seventh deal looks as good as these is a question the next year answers.

IVManagement

Sam Tabar, 54, is WhiteFiber’s chief executive and has been since February 2025. He has also been chief executive of Bit Digital since March 2023, and under an agreement that runs to August 2027 he gives up to about 30% of his time to Bit Digital.⁶ Billy Krassakopoulos, 46, is president; he started the data center operator Enovum in 2023 and earlier founded a hosting company that was sold in 2016.⁶ Thomas Sanfilippo, 63, the chief technology officer, ran the GPU cloud at Paperspace before DigitalOcean bought it.⁶ On 1 August 2026 Justin Zhu, 45, a former EY manager who had been Bit Digital’s vice president of finance, took over as chief financial officer from Erke Huang, who moved to an adviser role and stays Bit Digital’s CFO.¹⁰⁶ That is two finance leaders in the first year of public life, which is worth watching.

How they are paid, and what that pays them to do

The proxy shows 2025 pay for the first time. Tabar received $4.28M in total: $0.42M of salary allocated to WhiteFiber and $3.86M of stock awards, of which $1.81M was WhiteFiber shares and $2.04M was Bit Digital shares.⁶ Krassakopoulos received $3.14M, of which $2.66M was Bit Digital stock.⁶ Most of the executives’ equity pay is in the parent company, not in WhiteFiber. A manager paid mainly in Bit Digital stock is rewarded when Bit Digital does well, and Bit Digital owns 59.7% of WhiteFiber. That can line up, and it can also leave them favouring decisions that help the parent: WhiteFiber pays Bit Digital for management services at cost, borrows from it, and gave it a loan with a minimum 1.1 times return.⁶

The company’s own pay is stock. Stock pay was $16.9M in 2025 and $11.0M in the first half of 2026, which is 22% of the half-year’s revenue C (calculated).²³ The adjusted EBITDA the company reports adds all of it back.³ I do not, because stock handed to employees is a cost to the owner of the remaining shares.

How much of their own money is in

Not much. Directors and officers as a group own 354,334 shares, 0.8% of the company, worth about $3.7M at the close C (calculated).⁶¹ Tabar holds 283,303 of them, about $2.9M C (calculated).⁶ Nobody on the board other than Ichi Shih, who also sits on Bit Digital’s board, owns shares that the proxy counts.⁶ For the executives, the real stake is in Bit Digital, which I have not studied, and not here. That is a plain difference from founders who own a large piece of what they run.

Related dealings are worth reading closely. In May a Bit Digital subsidiary lent WhiteFiber up to $100M, at 9.5%, with at least 1.1 times the amount borrowed due at maturity. Each board set up an independent committee and got a fairness opinion, from Seaport Global for WhiteFiber and Needham for Bit Digital.¹⁵⁶ That is proper process, but the lender and the borrower share a chief executive. In August the company bought $1.0M of shares in SAIHEAT, a small listed company, and the proxy notes that four Bit Digital managers, including its CFO, invested $0.5M each personally in the same offering.³ I read nothing sinister in it. It shows how closely the two companies’ people are connected.

What they did with capital in the first fourteen months

The record is short. Four decisions explain most of it.

Over a full cycle I cannot say. Over fourteen months I can say that they built what they said they would, later than they said, with money that has become more expensive, and that they used shares as currency at prices well below where the stock traded in June. The price of the stock at which they issued is my sharpest test for management, and I could not pass it yet.

VFinancials

This is a short history, and the table says so. There are two full years, two half-years and a twelve-month total. The business also changed in the middle of it, when it bought Enovum and when it listed. Every line below is from the filings, and the bottom lines are mine.

Three facts jump out. Revenue is up 98% since 2024 in the last twelve months, to $94.5M, but $12.3M of the last quarter was a fee that will not come again.⁹³ Operating income went from a profit of $0.6M in 2024 to a loss of $26.8M in 2025 and a loss of $20.3M in the first half of 2026, mainly because administration costs rose from $10.3M to $52.5M, and $16.9M of the 2025 amount was stock pay C (calculated).² And the cash flow looks stronger than the business is. First-half operating cash was $89.1M, but $63.6M of that was customers paying in advance, and without it the number was $25.5M C (calculated).³

Owner cash

The cash an owner could take out is what is left after the company has paid for the equipment it needs. Over the last twelve months that was minus $313M C (calculated). Operating cash was $142M, of which $131M was growth in deferred revenue, mostly prepayments by customers for service still to come. The company spent $482M on equipment and deposits and sold $27M.²³⁴ The gap between adjusted EBITDA of $17M and owner cash of minus $313M is the size of the build. It is not waste, if the buildings are filled and the rent is paid. But it is paid for with debt now, so the first test is whether the new rent comes in.

The claims ahead of the owners

At 30 June the company had $56.1M of cash, $83.2M of term debt and loans, and a $230M convertible note carried at $222.6M. Its fair value was $407M, because the stock was then $38.85 and the notes were worth far more than their face.³ Then the August deals changed the picture. I estimate cash of about $223M and debt of $416M, so net debt of about $193M, before the third quarter’s building spend E (estimate). The company holds $143M of contract liabilities, mostly customer prepayments, which unwind as service is delivered without bringing cash in.³

There is also a cost I would not skip. The 2032 notes carry 5.00% and the loans cost between 7.9% and 9.5%, before the minimum-return clause on the bridge loan.³⁵ At roughly 7% on $416M of debt that is about $29M a year of interest, against adjusted EBITDA of $17M for the last twelve months and about $79M in my base for 2027 E (estimate). The company also has put and repayment dates. The remaining 2031 notes can be put in February 2029. The 2032 notes can be put in September 2030. The Iceland loan is due in two years, the RBC loan in three, and the July draws on the bridge loan, with 180-day terms, around January 2027 unless extended.³⁵⁶ A business that must refinance each year pays what the market asks that year. In October the market was not generous.

The share count

Shares outstanding grew from 38.85M to 45.28M in ten weeks, up 16.6%, because of the note exchange.³⁶ The table below shows what could still come, and what comes back.

The zero-strike call needs a word. In January the company paid $120M to a bank, Barclays, for the right to receive 5,905,511 shares for nothing starting in February 2031, or sooner if the bank asks. It is a prepaid buyback.³¹⁶ I count those shares as coming back, so I use 39.5M shares and not 45.3M. This relies on the bank delivering, and the company keeps no cash claim if the stock is worth less. The September presentation does the same bridge and ends at 39.2M shares before awards.⁹

VIValuation

First, how this goes badly wrong

Before the numbers, I asked what would most likely ruin this investment, and worked back. Three things stand out.

• The money runs out before the rent starts. The company had put $556M into colocation buildings and equipment by 30 June, mostly NC-1, on loans it has to roll. The July draws on the Bit Digital bridge have 180-day terms, so $30M of it falls due around January 2027 unless extended. A project loan on NC-1 is under negotiation with no assurance, the company says, that it will close on good terms or at all.³⁶ If lenders wait for Nscale to pay for a few months, the company has to raise money at the price of the day, from a share price that has fallen 77%. That is how a small landlord with good contracts becomes a dilution story.
• The tenant is the weak link. About 86% of firm future revenue is one contract with one young company that is itself borrowing to grow.³¹⁴ If Nscale cannot fill the 40 MW with its own customers, the contract may be renegotiated rather than broken, and a landlord with $400M of debt cannot afford to be difficult.
• The treadmill. GPU cloud has to replace its chips every few years, customers can leave on short notice, and the main customer already did. Cloud revenue can look large and leave nothing for the owners.³²

None of these needs a new event. They set my bear case. They also explain why I start from the cash that could reach an owner, and not from EBITDA.

What an owner can really take out

Reported adjusted EBITDA was $17M over the last twelve months. Take out stock pay of $21M and it is negative. Take out the equipment bought and it is minus $313M C (calculated).³² Looking forward, my base has $79M of EBITDA after head office costs in 2027, but only about $20M of owner cash if nothing new is built: I take off $16M of stock pay, $25M to $40M a year of cloud chips to replace, and tax and upkeep E (estimate). That is a quarter of the EBITDA. The gap is the number I care about. A buyer who values the company on EBITDA is valuing something that does not reach the owner.

How I built it

A company this young cannot be valued from its past, so I build the contracts forward. It starts from the 30 June balance sheet, adjusted for the financings the company has filed since: the new notes, the exchange, the bridge repayments. Cash flows count from 1 January 2027, each year’s at the end of the year; the second half of 2026 is a stub with no discount. I discount at 10.5% a year and continue after 2036 at 1.0% to 2.5%, depending on the case.

Data centers: NC-1 phase 1 is 40 MW paid for ten years. The company says $865M including yearly increases and one-off charges; the contract does not say how much is rent. I assume about $120M is build and installation charges the company has already collected, so rent starts at $66.5M a year and rises 2.5% a year E (estimate). Power is passed through and I leave it out. Margin on rent is 78%, taken from the colocation cost lines in the 10-Q. Montreal adds $19M of revenue at a 63% margin, and MTL-2 rents at $1.7M a megawatt. New megawatts cost $12.0M each, built the year before they are rented.³⁹ Cloud: revenue follows the signed Paris and other contracts, with a 50% margin after chip leases, and replacement chips cost 1.4 times a year of revenue over five years. Head office costs are $25M plus $0.1M a megawatt, and stock pay is $16M, growing 3%. Tax is 20% of operating profit after depreciation. I count customer prepayments of $25M, unwound, as debt E (estimate).

Shares: 45.28M on 21 October, less the 5.91M coming back under the zero-strike call, plus 0.15M for awards, makes 39.5M. If my value per share is above $25.91 the 2031 notes turn into 1.23M shares, and above $33.84 the 2032 notes turn into 9.16M shares, and I take the debt off.⁹⁵ Net debt is $193M, as in the Financials section.

What the price implies

The market value of the shares, net of the call option, is $409M. Add net debt and the prepayments and the business is priced at $624M, which is about what the property and equipment cost to build: $651M at book on 30 June C (calculated).³ In other words, the market is paying roughly what was spent, and my base says what was spent earns about its cost of money.

Two things follow. First, my base value if nothing more is ever built is $3.44 a share, and with 150 more megawatts it is $3.99. Growth adds only $0.55, because a new megawatt earns about 11% on its cost, close to the 10.5% I charge for money. It makes a difference if rents or build costs are better, and the bull case is exactly that. Second, the price equals my base value at a 9.5% rate, one point less than mine (the 9.7% a year in the table of 5 and 10 year values is a different measure; a note there explains). That is not a large gap in judgement. It is a large gap in dollars, because the equity is a thin slice on top of $416M of debt.

There is also a pure numbers check. Forward EBITDA of $79M in my 2027 base puts the business at 7.9 times, which looks low for signed ten-year rent. I assume that is part of the sell-side case: MarketBeat shows 11 analysts with an average target of $38.00 at 30 October, and I have not read their notes.¹⁷ My answer is the previous section. EBITDA here comes before $16M of stock pay and $25M to $40M a year of chip replacement, and the data centers sit under a head office that costs about $30M. I would pay a high multiple for a landlord with a rated tenant and cheap debt. I do not know that this is that.

What I can’t judge

Three inputs decide this page, and I cannot know any of them. The first is the cost of money. The company has just borrowed at 5.00%, with the right to be repaid in shares, and at 9.5% plus a minimum return, and I cannot say what its next dollar costs. The second is what Nscale can pay for ten years. The third is whether new megawatts earn more than they cost. The table shows the first.

Read that second table as the honest answer. A move of one point in the discount rate changes the base by $5 to $7 a share, which is between half and all of today’s price. Nobody can know the right rate to a point. That is why I give a range of $0 to $36 and not a number.

The discount rate

I use 10.5%. A normal check does not work here. The ten-year Treasury yields 4.94% and the implied equity risk premium is 4.20%.¹⁸¹⁹ The stock’s measured beta against SPY is above 4 on 61 weekly returns, which would give a cost of equity above 20%, and I do not trust a beta from 14 months of a stock that fell 77%.⁸ I use a blend instead. The company’s debt costs 5% to 9.5% a year; I put required return on its shares near 15%; with debt about half the capital, that is 10% to 11%. The contracted data centers deserve less and the cloud business more. Using one rate keeps the page comparable with the others.

Other uses of the same dollar

The ten-year Treasury pays 4.94% for no risk to the money.¹⁸ My base here returns about 9.5% a year on its cash flows at this price, a little below the 10.5% I require and above the Treasury. The bear loses almost everything, and the bull returns 27% a year E (estimate). On this site the nearest in kind are Oracle, an AI builder financing data centers with debt, whose price is 7% above my base, and Vertiv, which sells power and cooling to the same builders and trades 89% above mine. WhiteFiber is 159% above mine. Those pages use their own rates and dates. A reader weighing the same dollar will find larger and better-financed ways to own the AI build here, and none of them is cheap.

The margin of safety

• Price$10.36Close, Q4 2026 [1]
• Base value$4My estimate E (estimate)
• Margin of safety below$2half my base value C (calculated)

On the other pages I set the margin of safety from the bear case. Here the bear equity is zero, so that rule gives nothing, and I have to choose differently. I use half my base value, $2, because the base is a thin slice of equity that one point of rate can erase. The stock has never closed below $9.52, its low of 26 October 2026, since it listed.¹ So I do not expect $2 to be offered, and there is no margin of safety at $10.36.

Two different questions

How good is the business, and how good is the price? On the first: there are real assets, grid power at a site that took years to find, a ten-year contract that runs to 2036, and a team that built MTL-3 in six months. It is also a business that owns one anchor tenant, funds itself at a high cost, has shown a net loss every half-year since the IPO, and has been public for fourteen months. On the second: at $10.36, the price is 159% above my base, and it needs the bull case to have about a one in four chance. This looks like a business with a real asset at a price that still needs a good deal to go right. The 77% fall does not change that on its own, because the fall came from a price that was far higher than any reasonable value, and part of it was dilution.

Where this could be in 5 and 10 years

For my base to stand, by 2031 NC-1 has paid on time, about 100 MW of new capacity is rented at roughly $1.7M a megawatt, cloud sits near $100M, and net debt has peaked near $2.0B E (estimate). By 2036 building has stopped, EBITDA is $333M, owner cash is about $224M a year, and the debt is paid down toward 5 times EBITDA. That produces $10.77 in 2031 and $26.20 in 2036.

The cases differ on three numbers: how many new megawatts earn more than they cost, the rent and build cost per megawatt, and the cost of money. The bear builds nothing, and its equity is worth about nothing in 2031 and $0.35 in 2036. The bull builds 300 MW at higher rents, which is why it reaches $62 in 2031 and $96 in 2036, helped by both convertible notes turning into shares.

The single most likely way my base is wrong is the share count. Building 150 MW needs more money than lenders give, and if the company raises it by selling shares near $10, each $100M adds about 10M shares, a quarter of today’s count. That would cut my 2036 value per share a lot. The table shows how much, in my base only.

So the base is flattered, but by less than the headline suggests if shares go out near today’s price: a $162M raise at $10.36 takes about $2.60 off the 2036 value and 1.1 points off the return. At $5 the same raise takes $8. The risk is selling shares low after the price falls, which is what the 77% drop already shows can happen.

One more check, because two numbers on this page look alike. In the reverse valuation the price equals my base value at a 9.5% rate. The table above gives 9.7% a year to 2036. They measure different things. The 9.5% is the one rate applied to every cash flow, including the terminal value, that makes my base worth $10.36 today. The 9.7% is what an owner earns by paying $10.36 and holding to 2036, when my method values the shares at $26.20. Two effects nearly cancel: my debt costs 7.5%, below the rate I discount at, which adds about 3 points (13.0% instead of 9.8%), and my 2036 value still discounts later years at 10.5%, where 9.5% would give $35.20, which takes about 3 points back C (calculated).

A reader’s guide: 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 WhiteFiber’s favour and get less attention than the share price. First, the contracts are large against the price. The 10-Q counts $1,008M of future revenue the company has signed and expects to recognise, against a business I put at $624M at today’s price.³ Second, the hardest part of NC-1 is done: Duke Energy has finished the work to deliver 54 MW to the site, and a building that has power is rarer than a building.⁹ Third, customers are paying part of the bill: $72.6M of prepayments came in during the first half, and the Paris contract is described as supported by prepayments and project-level financing.³⁹ Those prepayments also create a duty to deliver service without new cash, which I count as debt.

The bull case

The strongest case in plain words: AI builders are short of power and buildings for years, WhiteFiber owns scarce power at a site with eight hyperscale data centers within 100 miles, it has a ten-year tenant for the first 40 MW, and the retrofit model lets it add more at about $8M to $10M a gross megawatt.⁹ If it finances NC-1 cheaply and rents the next phases at the same rent, the shares are a thin slice of a large, contracted asset, and a thin slice moves a lot.

What the company says. In its 12 August 2026 results release, Sam Tabar said NC-1 had “moved from construction into active customer deployment”, that lenders had entered exclusivity on a secured NC-1 financing with no assurance it would close, that cloud contracts of more than $540M had been signed since the May call, and that demand for scarce power available in 2027 was strong.²⁰ I checked these against the 10-Q. The contracts and the financing talks are in the filing. The 10-Q filed the same day says management expects to begin delivering capacity to Nscale in the third quarter, and the September presentation says initial billing has begun with the full ramp subject to acceptance, so the word “active” runs ahead of the filings.³⁹ I did not find a transcript of his public interviews. A video titled with the words “Markets ‘Incorrect’ on A.I. Sell-Off, Demand ‘Very Real’” attributes such remarks to the WhiteFiber CEO, but I could not confirm its date, channel or speaker, so I do not use it.

The wider debate, on All-In. The 17 October 2026 episode of All-In, “CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer”, spent about thirteen minutes on the neocloud model at roughly 6:00 to 19:30, by the automatic transcript’s approximate timestamps; the speaker labels in that transcript are patchy, and I cross-checked who said what with a dated report of the episode.²¹²² Nobody in it mentions WhiteFiber. Friedberg was not on the episode, and Ben Horowitz sat in. What they said, in my words, and what I make of it against the filings:

• Chamath Palihapitiya argued that the risk in a neocloud is its borrowing and the cost of rolling it, not weak demand, called default odds implied by credit prices “a coin flip”, and pointed to three customers making up about three quarters of CoreWeave’s revenue.²¹ That fits WhiteFiber better than the demand worry does: one customer was 63% of first-half revenue, one tenant is most of firm rent, and the bridge loan costs 9.5% plus a minimum return.³ The stock fell 42% in October, from $18.01 to $10.36, in the same weeks the episode describes, though I cannot show the cause.
• David Sacks called the sell-off “just volatility”, said demand for compute stays strong with older Nvidia chips still renting at high prices, said chip lives are lengthening, and called payback period the number to watch.²¹ The filings give partial support: WhiteFiber sold 126 H200 GPUs for $1.8M above book, and the Paris and other new deals are signed, but it also lengthened its own chip life from three to five years in 2025, which is the accounting version of the same point.³² Payback is what my 1.4 times revenue figure for chips tries to capture.
• Ben Horowitz said the world faces “a massive shortage, not an overage of supply”, and also that he has avoided investing in neoclouds because he saw overleveraged infrastructure break when prices fell.²¹ Both halves apply. The shortage is why Duke’s power and Nscale’s contract exist. The leverage warning is why I treat financing as the main risk.
• Jason Calacanis showed a prediction-market price for a CoreWeave default and the group thought it too high.²¹²² That is about a different company with far more debt. I take from it only that markets are pricing the sector’s credit, not its demand.

Others I looked for. I found no statement about WhiteFiber from Gavin Baker or Brad Gerstner. An 11 June 2026 BG2 episode and a 14 July 2026 Baker conversation on the a16z podcast are listed as touching AI compute and data centers, but I could not read a transcript, so I attribute nothing to them. The one named sell-side action I could verify is that H.C. Wainwright reiterated a Buy on 25 September 2026, per MarketBeat, which also shows an average target of $38.00 across 11 analysts. I have not read those notes and I give them no weight beyond saying the sell-side is far above my base.¹⁷ Among holders, Situational Awareness LP reported 1,757,600 shares ($68.3M, or $38.85 a share) at 30 June 2026 in a 13F filed on 14 August. That is a dated fact; the 13F does not show shorts or non-US holdings, and I do not infer what the fund thinks now.⁷

Against reality

The first half of the case is supported by the filings: power is delivered, a ten-year contract is signed and $1.0B is on the books. What is missing is the second half, which is cheap money, a second tenant and an anchor tenant whose own credit I can check. If NC-1 bills in full, if the project loan closes at a rate near 8%, if the next 150 MW rents at the same $1.7M a megawatt, and if the company stops issuing shares, then the price is possible. My base says so only at a rate of 9.5%. Today the score is: contract yes, power yes, financing not yet, second tenant not yet.

Multiply it out. Nscale pays on schedule and fills its 40 MW: 80%, a guess. NC-1 finances at near 8%: 50%. A new tenant takes 100 MW or more by 2031 at the rent I use: 35%. Cloud earns its cost of chips: 50%. Together that is about 7% E (estimate). They overlap, because a strong AI market helps all four, so call it 12%. The 25% I give the bull in my weighted $11 is, if anything, generous, and it is almost exactly what the price asks. With a 12% bull and a 60% base, the weighted value would be about $7.

What a 10x would take. Ten times $10.36 is $103.60 a share, about $4.7B of market value on 45.3M shares. With the net debt my base needs, the business would have to be worth $6B or more. At 12 times EBITDA that is about $500M a year, roughly 1.5 times my base for 2036 and below my bull for 2034 C (calculated). My bull reaches $96 a share by 2036, 9.3 times today’s price, so a tenfold needs everything in the bull, on time, with no dilution at today’s price. That is 26% a year for ten years. Tenfold in a decade from half a billion dollars of market value has happened to some small companies in booms, and to few that were financed this way. I put it under 5%, and I would build nothing on it.

Where I come down

The signed contracts and delivered power moved me: without them I would not give a bull case. The company’s own account of progress did not, because the filings describe a slower ramp than the release does. A financed NC-1 at a low rate, and a second tenant, would raise my values most: a 6% rise in rent on every megawatt adds about $2.5 a share to my base E (estimate).

VIIIRisks, and what would change my view

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

• Control. Bit Digital owns 59.7% and lends the company money, shares its chief executive for up to 30% of his time through August 2027, and can outvote everyone else.⁶
• Customer concentration and churn. Seven customers, the largest at 63% of first-half revenue; three orders discontinued in six months; $15.7M of termination fees unpaid and $5.1M of lost revenue.³
• Delivery. NC-1 billing slipped from June to the third quarter. The Paris service date slipped from July to September. Both depend on equipment arriving and customers accepting it.²³
• Power. Duke’s agreement for 99 MW is “commercially reasonable efforts” by 2029, and earn-out payments to the seller of up to $8M, or $5M, rise with delivery.³
• Refinancing and covenants. The RBC and Iceland loans carry leverage and coverage tests; the 2031 and 2032 notes can be put in 2029 and 2030.³⁵
• Dilution. The share count rose 16.6% in ten weeks, and up to 10.4M more could come from the notes.⁶⁹
• Chips. The lives of GPUs, the price of renting them and the supply of new ones are outside the company’s control, and it has already changed its useful-life estimate once.²
• Governance and reporting. An emerging growth company need not have its internal controls audited, and it has changed chief financial officer once in fourteen months.³¹⁰
• Currency and tariffs. Montreal revenue and costs are mostly in Canadian dollars, and the company lists tariffs on Canada among its risks.³
• Counterparty on the call option. The 5.9M shares come from Barclays in 2031. I count them as retired.¹⁶

Geopolitical risk, and what it means for the value

WhiteFiber earns its money in one country and builds in another. In 2025, 87% of revenue came from services provided in Iceland and 11% from Canada, while the long-lived assets sat about a third each in Iceland, Canada and the United States C (calculated).² North Carolina is where growth is meant to come from.

Chips are the first chokepoint. The chips it uses are made in Taiwan, and the filing names US-China and China-Taiwan tension, tariffs and Commerce Department export rules as risks.² Commerce withdrew its AI diffusion rule, which would have capped chip exports worldwide, and says a replacement is coming.² Iceland and France are allies, so I would expect licences before bans. A rule that did bite would land on the cloud business, where the Iceland and Paris revenue sits.

Power is the second. Iceland puts homes and some other users ahead of data centers if water runs short, and volcanoes cut geothermal output several times in 2023.² Duke Energy agreed to use “commercially reasonable efforts” at NC-1, not to guarantee supply.² The parent is the third. Bit Digital owns 59.7%, lends the money and shares the chief executive. When it bought the Montreal operator, a Canadian national security review ended with promises to keep a Canadian on the board, keep security controls and report the client list each year. That costs attention, not dollars.²

The table reruns my model with each case changed alone. Odds are my judgement, based on the filings’ own warnings and on how often utilities and builders run late; they are not data. Tariffs on electrical equipment and a late utility are the likelier ones. None is in my base, and the $10.36 price is already 2.6 times it, so I do not see a discount for them in the price.