Applied Digital
A landlord for artificial-intelligence computers: it builds data centres with the power and cooling they need, then rents each whole building to one tech giant for fifteen years.
The scores
Growth3/4 ↑
Revenue up 167%, but the record is under two years, one customer is 59% of it, and three-quarters of the quarter was construction pass-through.
Business quality3/4 →
Real scarce-input advantage - sites with power, qualification, on-time delivery - but untested by a cycle, and the newest leases priced worse than the first.
Financial position2/4 ↑
The plan needs about $10B not yet raised, and 79% of borrowings mature in FY2031, inside the horizon. Cash covers the near term.
Execution record3/4 ↑
Delivered 175 MW on time, cut build time from 24 months to under 12, and borrowing cost from 9.25% to 7.00%. One 167% restatement.
Per share capture2/4 ↓
Shares up 28% in a year, stock pay twice adjusted EBITDA, and 75% of the asset equity sold to a preferred paid first.
Potential3/4 ↑
A bounded upside that is genuinely committed: take-or-pay, non-cancellable, filed year by year. Not a 4 - financing for 810 MW is outside its control.
Risk profile2/4 ↑
More than one serious concentration: one tenant is 59% of revenue, one hyperscaler over half the contracted rent, and only three tenants exist.
Valuation3/4 ↑
12 times stabilised rental profit prices the plan working, but at a worse exit than management describes. Not a 4 - today's business earns $90.4M.
Evidence quality3/4 →
The rent, debt and preferred terms are filed facts. The build cost and loan-to-cost split are management claims graded, not verified.
If it works 2.1× · if it fails you keep about 0%.
The Verdict
We rate Applied Digital neutral. It builds data centres for AI and rents each one whole to a large tech tenant on long leases that must be paid whether the space is used or not. The rent is real, but most of it arrives years away and the building is paid for with expensive debt and a partner's preferred stake, so the business is mixed. The price is a wide discount to that rent, earned by what is paid first, so it is fair. As a bet it is fragile: a little under double if every building gets built and paid for, little or nothing left if it fails.
The long caseAI chips need more power and cooling than old data centres give, so tenants sign long leases for new buildings; that pays if Applied Digital can build and finance them.
The evidence under this - 9 claims
- 2026-10-08editedQ1 FY2027 release: Building 2 delivered in two 75 MW phases (1 July and 1 October), taking live capacity to 250 MW; leased total 1.41 GW unchanged.
- The 91% margin is one building's first partial year. An analyst put the mid-eighties to the CFO as the fairer number and got no denial; at 85% the figure is about $1.96bn
- The ladder excludes reimbursements and variable rent, so it understates gross receipts - but NOI is struck on base rent, so this cuts the right way
- It assumes every campus delivers. Two of the five are contracted for delivery in 2028 and have not broken ground on their full footprint
- The share count is the wrong one and known to be: it rose 28% last year and the model holds it flat, so on any realistic issuance the implied multiple at today's price is higher than twelve times
- It assumes Macquarie stops at the funded $1.8bn plus its share of the remaining equity. If the full $6.7bn commitment is drawn, the 1.80x floor alone is $12.06bn - more than today's entire market capitalisation
- It ignores the bitcoin hosting segment, ChronoScale, the Base Electron stake and the Babcock & Wilcox warrants, which together are real but small against a $16.9bn build
- Most importantly it assumes an exit multiple exists at all. No listed pure-play prices this asset (C40), so twelve times is a description of today's price rather than a valuation struck against an observable market
- 2026-09-27editedre-struck on the 2026-09-25 close of $26.25 from the refreshed pack after a 12.1% move since the price was struck; the stabilised NOI, debt, Macquarie floor and share count are unchanged
- 2026-10-21editedre-struck on the 2026-10-20 close of $19.33 (from $26.25, -26.4%, past the 12% drift trigger) after the Q1 FY2027 print left the stabilised NOI, debt, Macquarie floor and filed share count unchanged
- The company has repeatedly raised that ceiling and says so, and its delivery record supports it: the first Polaris Forge 1 building took about 24 months from start of construction to ready-for-service and the second took under twelve
- Against that, the admission is unprompted and specific - 'we're definitely exceeding that a little bit' - and it is the one growth constraint management named without being pushed
The company
Artificial-intelligence chips need far more power and cooling than older data centres can carry. Applied Digital builds the buildings that can: the shell, the substation, the power distribution and liquid cooling. It leases each one whole to a single tenant, which brings its own chips; Applied Digital owns none of the computing.
The tenant's alternative is to build for itself or to wait: only 1.4% of space in the main markets stands empty.
The evidence under this - 2 claims
How it makes money
The tenant pays rent for powered, cooled space under a lease it cannot cancel. On the one building running last year, 91 cents of every rent dollar was left as operating profit. For now most reported revenue is something else: construction work for tenants, passed through at roughly cost.
The evidence under this - 3 claims
- The split is given only on the call. The 10-K's income statement carries services revenue and data center rental revenue separately but does not decompose the HPC segment this way, so the $44.1m of base rent cannot be tied to a filed line
- It is, however, consistent with two filed figures: full-year HPC base rental revenue of $99.8m in the NOI reconciliation, and the fourth-quarter gross margin falling to 15.7% from 42.5% in the third quarter, which is what a quarter dominated by near-zero-margin construction pass-through looks like
- Rechecked 2026-10-21 against the first-quarter 10-Q: its segment note splits HPC revenue into $183.5M of services and $79.1M of rental revenue (base rent plus recoveries), matching the release for that quarter, but it does not restate the fourth quarter, so this split still rests on the July call.
- 2026-10-21editedvalid_through of 2026-10-08 passed; the figures are a closed quarter and do not go stale, so the claim now carries a refresh_by of the annual report instead, with the 10-Q recheck recorded. Not re-sourced to a filing: neither the release nor the 10-Q restates the fourth quarter.
Where the money comes from, stream by stream
| Stream | Share of sales (FY2026) | How it earns it |
|---|---|---|
| Tenant fit-out services | 44.3% | Buying and installing the tenant's equipment inside a new building, billed at cost plus a contractual markup while construction runs. |
| Blockchain hosting | 25.3% | Power and space for one bitcoin-mining customer at the older North Dakota sites; a shrinking legacy line. |
| HPC base rent | 16.3% | Fifteen-year take-or-pay rent on finished AI data-centre buildings, recognised straight-line; the line the whole case rests on. |
| ChronoScale cloud services | 11.7% | Renting GPU computing by the hour through the 96%-owned ChronoScale, consolidated but outside every adjusted measure. |
| Tenant recoveries | 2.4% | Operating costs such as power-related and property expenses recharged to the HPC tenants. |
Who runs it
Wes Cummins has run Applied Digital since 2021 and turned a bitcoin host into a landlord for the largest cloud companies. Insiders own under a tenth, mostly his, and much of his stake is pledged against a loan. His big award pays only if the share price climbs a long way; his deputies are paid on leases signed and rental profit. He and his president also run a power company the business has lent to.
The evidence under this - 3 claims
- The pledge means three-quarters of the chief executive's holding secures a personal loan, so a sharp fall could force sales that have nothing to do with his view of the business.
- Nineteen of the last year's twenty insider sales were discretionary (C36), so the stake has been shrinking by choice rather than on a schedule.
- The president's hurdles reward megawatts signed and delivered whatever the return earned on them, and no award pays on per-share value other than the chief executive's price hurdles, which more issuance does not hinder.
- On top of these, fully vested ChronoScale incentive units with no condition were granted to the same four officers in April 2026 (C38).
The company's expansion beyond its current campuses depends on new power (C05), and Base Electron exists to build it, so the stake and the loan may serve the company's own pipeline; separate related-party committees now review each transaction.
Why it wins, and who could beat it
Why customers come back
The scarce thing is not the building. It is land with a utility willing to supply hundreds of megawatts, secured before anyone asks. Management says it contracted 2.1 gigawatts of utility power in a single year, in regions other developers are not working.
Elsewhere the tenant joins a queue: 80.4% of what is being built in the main markets is spoken for, with about six months of demand left. The proof: three tech giants signed fifteen-year leases while it had one working building.
The evidence under this - 3 claims
- Every one of the fifteen competitors named in its own Competition section is pursuing the same scarce input, and several - IREN, TeraWulf, Cipher, Core Scientific, Hut 8 - inherited energised sites from bitcoin mining rather than having to queue for them
- The expansion beyond the existing campuses depends materially on Base Electron, a related party in which Applied Digital owns about 10%, whose first 1.2 GW of generation is not expected until calendar 2029 and 2030
The threat
The biggest rival is the customer. The tech giants build almost 60% of their own capacity, and Applied Digital's annual report lists them first among its competitors. Behind them come fifteen named companies, eleven of them bitcoin miners converting to chase the same leases.
The evidence under this - 3 claims
The rivals: MSFT, AMZN, GOOGL, META, WULF, CORZ, DLR.
Who else does this, rival by rival
| Ticker | Rival | How far along | Priced against it |
|---|---|---|---|
| MSFT · AMZN · GOOGL · META | Microsoft, Amazon, Google and Meta building their own capacityPriced against it: Different product | On sale | Different product |
| WULF | TeraWulf Inc.Priced against it: Same product | On sale | Same product |
| CORZ | Core Scientific, Inc.Priced against it: Same product | On sale | Same product |
| DLR | Digital Realty Trust, Inc.Priced against it: Same product | On sale | Same product |
How big it could get
How big this could get
Capacity under construction in North America's main data-centre markets rose 24.8% in the first half of 2026, to a record 7,481.1 MW. Roughly 40% of the tech giants' capacity is rented rather than built. They cannot honestly be turned into a market size: they count markets Applied Digital avoids, and nobody measures the ones it builds in.
The evidence under this - 3 claims
Who has signed
Three tenants have signed 1,410 megawatts across five campuses: CoreWeave for 400 MW, and two unnamed investment-grade tech giants for the rest. One of the unnamed pair holds roughly $20B of the $36.2B, more than half.
The evidence under this - 2 claims
The results
The numbers
Sales grew 5.5% in the year to May 2025, slower than the 147% in the year before.
Operating margin is −50%, up from −244% in the year to May 2022.
Running costs take 0¢ of each sales dollar, down from 233¢ in the year to May 2022.
Yearly sales growth has ranged from +5.5% to +548% over the last 3 years.
It used $797M of cash in the latest year, so the $44.0M it holds lasts about 0.1 years at that pace.
The share count rose 252% in 3.0 years, and pay in shares costs 16% of sales.
All the numbers
Filed figures from the company's 10-K and 10-Q, read 2026-10-08. Fiscal years end in May.
| May 2021 | May 2022 | May 2023 | May 2024 | May 2025 | |
|---|---|---|---|---|---|
| Size and growth | |||||
| Sales | $0 | $8.55M | $55.4M | $137M | $144M |
| Sales growth | - | - | +548% | +147% | +5.5% |
| Profit | |||||
| Gross margin | - | −11% | 20% | −8.6% | −50% |
| Operating margin | - | −244% | −77% | −72% | −50% |
| Net margin | - | −275% | −81% | −109% | - |
| What growth costs, per sales dollar | |||||
| Selling, general and admin | - | 233% | 97% | 72% | 75% |
| Pay in shares | - | 144% | 58% | 13% | 16% |
| Capital spending | - | 643% | 237% | 104% | 473% |
| Cash | |||||
| Cash from operations | −$83.0K | −$872K | $58.7M | $13.8M | −$115M |
| Free cash flow | −$103K | −$55.8M | −$72.5M | −$128M | −$797M |
| Balance sheet | |||||
| Cash and short-term investments | $11.8M | $38.8M | $29.0M | $3.34M | $44.0M |
| Debt | - | $7.19M | - | - | - |
| Net cash | - | $31.6M | - | - | - |
| Returns and shares | |||||
| Shares (diluted) | - | 57.1M | 94.0M | 114M | 201M |
What growth costs
Growth here costs buildings, not sales staff. Office overhead has shrunk fast as a share of sales, but pay in shares has ballooned and building spending runs far ahead of revenue. Much of today's revenue is construction work passed through to tenants, so the true test is the rent margin once buildings fill.
The evidence under this - 2 claims
- Fiscal 2026 sales include $270.6m of tenant fit-out passed through at cost plus a markup (C52), which flatters every per-sales-dollar ratio that year; on revenue without it, overhead excluding stock pay is about 33 cents.
- The filer tags no research and development line, and its stock pay figure is the cash-flow add-back of $220.1m, slightly above the $219.3m expensed (C17).
The last report (27 Jul)
Quarter ended May 2026, filed figures.
In the three months to 31 August 2026 sales were $341.9M. Rent, the line that earns, rose to $65.8M from $44.1M the quarter before, and live capacity reached 250 MW. But $183.5M was construction work passed through to the tenant.
The profit left on that rent was 89 cents on the dollar, below the 91 the case wanted but far above the 80 that would break it. On the call, management expects about 250 MW of new leases by year-end at rents over 15% higher; none is signed.
What it did to the thesisIt half answered the test we set: rent is taking over, but slowly, and the margin dipped as a second building started. The call added hopes of higher rents and cheaper financing, none signed, so the Neutral view stands.
The quarter in detail
Sales growth from the previous quarter, filed figures.
Earnings per share $-0.39 against $-0.11 expected.
The evidence under this - 4 claims
- 2026-10-21editedchecked against the filed 10-Q and the call, which the original entry had not seen: the figures agree, so only the sourcing changed.
- The split is given only on the call. The 10-K's income statement carries services revenue and data center rental revenue separately but does not decompose the HPC segment this way, so the $44.1m of base rent cannot be tied to a filed line
- It is, however, consistent with two filed figures: full-year HPC base rental revenue of $99.8m in the NOI reconciliation, and the fourth-quarter gross margin falling to 15.7% from 42.5% in the third quarter, which is what a quarter dominated by near-zero-margin construction pass-through looks like
- Rechecked 2026-10-21 against the first-quarter 10-Q: its segment note splits HPC revenue into $183.5M of services and $79.1M of rental revenue (base rent plus recoveries), matching the release for that quarter, but it does not restate the fourth quarter, so this split still rests on the July call.
- 2026-10-21editedvalid_through of 2026-10-08 passed; the figures are a closed quarter and do not go stale, so the claim now carries a refresh_by of the annual report instead, with the 10-Q recheck recorded. Not re-sourced to a filing: neither the release nor the 10-Q restates the fourth quarter.
- None of it is signed: the 10-Q and the release record no expansion lease, no financing for the three campuses and no Macquarie extension, so each is a statement of intent until the January print shows it.
- The CFO also said rates 'can move around 50 to 100 basis points' and that construction debt carries a 25 to 100 basis point premium, which is the cost the 7.00% placement does not yet show for the next three campuses.
What management did with the money
Shares outstanding rose 28.0% in the year as preferred stock converted and new shares were sold into the market. In April it also granted 9.3 million fully vested management units with no conditions attached, expensed at $47.9M on the day.
Insiders in 2026Q2: 6 purchases and 13 sales, from their Form 4 filings.
The evidence under this - 2 claims
The price, and where we disagree
In short. Fair. Cheap on the finished data-centre campuses, fair on the road to them. The discount pays for years of interest, Macquarie's preference and new shares before most of the rent arrives.
The shares are 61.9% below their high of the past year and down 38.3% over the last six months. In the six months before that the price rose only 2.1% while sales per share nearly doubled, so buyers had already been paying about half as much for each dollar of sales.
Meanwhile borrowing got cheaper: the latest $1.59B of project debt cost 7.00%, against 9.25% on the first placement in November 2025.
The evidence under this - 4 claims
- 2026-09-27editedre-struck on the refreshed pack's 2026-09-25 close: the 126-session window now runs from 26 March, so the price is up 2.1% rather than down 13.5% while the de-rating on revenue per share, 48.6%, still holds
- 2026-10-21editedre-struck on the 2026-10-20 close: the price is now 61.9% below the high and down 38.3% over 126 sessions; the revenue-per-share decomposition is kept for its original 26 March to 25 September window rather than recomputed.
- The split is given only on the call. The 10-K's income statement carries services revenue and data center rental revenue separately but does not decompose the HPC segment this way, so the $44.1m of base rent cannot be tied to a filed line
- It is, however, consistent with two filed figures: full-year HPC base rental revenue of $99.8m in the NOI reconciliation, and the fourth-quarter gross margin falling to 15.7% from 42.5% in the third quarter, which is what a quarter dominated by near-zero-margin construction pass-through looks like
- Rechecked 2026-10-21 against the first-quarter 10-Q: its segment note splits HPC revenue into $183.5M of services and $79.1M of rental revenue (base rent plus recoveries), matching the release for that quarter, but it does not restate the fourth quarter, so this split still rests on the July call.
- 2026-10-21editedvalid_through of 2026-10-08 passed; the figures are a closed quarter and do not go stale, so the claim now carries a refresh_by of the annual report instead, with the 10-Q recheck recorded. Not re-sourced to a filing: neither the release nor the 10-Q restates the fourth quarter.
How we read the multiple
Once every campus is built and let, the rent should leave about $2.09B a year of operating profit. The shares are worth about 11.3 times that: an 8.8% yearly return on rent tenants cannot cancel.
That looks cheap until you count what is paid first: about $486M a year of interest, due now, and Macquarie's preferred shares, both ahead of ordinary shareholders.
It is sensitive: each extra dollar buyers pay per dollar of that profit adds about $7.27 a share, roughly 38% of the price.
Against its own history. Its own past ratios describe a different company: most of the buildings that will pay the rent are still construction sites, so this comparison carries less weight than usual.
The evidence under this - 5 claims
- The 91% margin is one building's first partial year. An analyst put the mid-eighties to the CFO as the fairer number and got no denial; at 85% the figure is about $1.96bn
- The ladder excludes reimbursements and variable rent, so it understates gross receipts - but NOI is struck on base rent, so this cuts the right way
- It assumes every campus delivers. Two of the five are contracted for delivery in 2028 and have not broken ground on their full footprint
- The share count is the wrong one and known to be: it rose 28% last year and the model holds it flat, so on any realistic issuance the implied multiple at today's price is higher than twelve times
- It assumes Macquarie stops at the funded $1.8bn plus its share of the remaining equity. If the full $6.7bn commitment is drawn, the 1.80x floor alone is $12.06bn - more than today's entire market capitalisation
- It ignores the bitcoin hosting segment, ChronoScale, the Base Electron stake and the Babcock & Wilcox warrants, which together are real but small against a $16.9bn build
- Most importantly it assumes an exit multiple exists at all. No listed pure-play prices this asset (C40), so twelve times is a description of today's price rather than a valuation struck against an observable market
- 2026-09-27editedre-struck on the 2026-09-25 close of $26.25 from the refreshed pack after a 12.1% move since the price was struck; the stabilised NOI, debt, Macquarie floor and share count are unchanged
- 2026-10-21editedre-struck on the 2026-10-20 close of $19.33 (from $26.25, -26.4%, past the 12% drift trigger) after the Q1 FY2027 print left the stabilised NOI, debt, Macquarie floor and filed share count unchanged
- Cash interest is materially offset today: the company earned $55.7m of interest income in fiscal 2026 on a cash pile that ended the year at $3.97bn, and a large share of interest incurred is capitalised into construction rather than expensed ($39.6m in fiscal 2026)
- Contracted minimum rent is not the only receipt - tenant recoveries, variable rent and fit-out revenue all sit outside the Note 17 ladder, and fit-out alone was $152.4m in the fourth quarter
- The comparison is also a moment in time rather than a solvency test: the FY28 ladder step to $1.45bn more than triples the coverage, and the company held $1.59bn of unrestricted cash at year end
- The leverage runs both ways and the downside is not symmetric: at ten times the residual is $9.80 a share and at eight times it is negative, so the same arithmetic that makes the upside large puts a floor under the stock at a multiple only four turns below today's
- It holds the share count flat, which C19 says it will not be; every share issued dilutes the per-turn sensitivity
- 2026-09-27editedre-struck on the 2026-09-25 close of $26.25: the $7.27 a turn is unchanged and is now about 28% of the price, against about 30% of the $23.96 the thesis was struck at
- 2026-10-21editedre-struck on the 2026-10-20 close of $19.33: the $7.27 a turn is unchanged and is now about 38% of the price, against about 28% of the $26.25 it was last struck at
- Core Scientific is converging fast on the same product - roughly 920 MW of leasable high-density colocation across ten data centres - and within a year or two may be a genuine comparable
- The absence of a listed pure-play does not mean the asset is unpriced. Aligned changed hands at roughly $40bn for over 5 GW operational and planned (C31), which is a real transaction on a closely comparable platform, and Applied Digital's enterprise value against its contracted gigawatts is not far from it
- Digital Realty does write hyperscale build-to-suit leases, so the exclusion is a judgement about the balance of each business rather than a clean separation
Priced against its rivals
There is no group to price it against. No listed pure-play prices this asset, so the price faced no peer comparison. The nearest evidence is a private transaction on a comparable platform, not a cohort. The price is judged against its own history instead.
The bet: what it pays if it works
This is a bet that every building Applied Digital has leased gets built and pays rent. Even then, after the debt and the partner's preferred stake come out, the shares would be worth less than twice today's price at a fair valuation of that rent. More tenants could add to it, but none has signed.
If it fails, little is left: the lenders and the partner are paid first, and a late campus or a tenant walking away could leave nothing for shareholders.
| Revenue in two to three years | Contracted megawatts needed | At the stated multiple |
|---|---|---|
| $1.10B | about 675 at 100% kept | 1.0× |
| $2.30B (the big win) | about 1.4K at 100% kept | 2.1× |
| $3.44B | about 2.1K at 100% kept | 3.1× |
| $4.00B (market forecast) | about 2.5K at 100% kept | 3.6× |
JLL's outlook has about 17.5 GW built for lease in North America in 2029 and 2030; at a 6% share, under its 700 MW ceiling, Applied Digital adds about 1,050 MW to the 1,410 signed, billing $4.0bn. On that forecast, $4.00B pays 3.6× from today's value: a Fair bet on the market case.
To pay 2× (a Fair bet) it would need about $2.23B of revenue (about 1.4K contracted megawatts); to pay 5× (a Cheap bet), about $5.57B of revenue (about 3.4K contracted megawatts), at the same multiple.
The steps it has to take
- Done About 1,410 MW leased take-or-pay and non-cancellable to three tenants
- Done First 100 MW building operating and billing, at a 91% NOI margin
- Done Project notes and Macquarie preferred raised for the first two campuses, debt cost falling
- Slipped Base rent above Q4's $44.1m with NOI margin holding at 91%, as a second building loads due 2027-01-06
- Not started Macquarie's commitment extended to the last three campuses, new debt no wider than 9.25% due 2027-05
- In progress A $1bn NOI run rate, as management expects within a year due 2027-07
- In progress All five campuses ready for service, the last in the first half of 2028 due 2028-06
- Not started NOI margin at or above 80% across the fully loaded portfolio due 2028-12
The detail behind the label
The evidence under this - 7 claims
- The 91% margin is one building's first partial year. An analyst put the mid-eighties to the CFO as the fairer number and got no denial; at 85% the figure is about $1.96bn
- The ladder excludes reimbursements and variable rent, so it understates gross receipts - but NOI is struck on base rent, so this cuts the right way
- It assumes every campus delivers. Two of the five are contracted for delivery in 2028 and have not broken ground on their full footprint
- The company has repeatedly raised that ceiling and says so, and its delivery record supports it: the first Polaris Forge 1 building took about 24 months from start of construction to ready-for-service and the second took under twelve
- Against that, the admission is unprompted and specific - 'we're definitely exceeding that a little bit' - and it is the one growth constraint management named without being pushed
- The share count is the wrong one and known to be: it rose 28% last year and the model holds it flat, so on any realistic issuance the implied multiple at today's price is higher than twelve times
- It assumes Macquarie stops at the funded $1.8bn plus its share of the remaining equity. If the full $6.7bn commitment is drawn, the 1.80x floor alone is $12.06bn - more than today's entire market capitalisation
- It ignores the bitcoin hosting segment, ChronoScale, the Base Electron stake and the Babcock & Wilcox warrants, which together are real but small against a $16.9bn build
- Most importantly it assumes an exit multiple exists at all. No listed pure-play prices this asset (C40), so twelve times is a description of today's price rather than a valuation struck against an observable market
- 2026-09-27editedre-struck on the 2026-09-25 close of $26.25 from the refreshed pack after a 12.1% move since the price was struck; the stabilised NOI, debt, Macquarie floor and share count are unchanged
- 2026-10-21editedre-struck on the 2026-10-20 close of $19.33 (from $26.25, -26.4%, past the 12% drift trigger) after the Q1 FY2027 print left the stabilised NOI, debt, Macquarie floor and filed share count unchanged
Where we disagree with the market
A safe compounder: $36.2B of rent from top-rated tenants. Thirteen of fourteen rated analysts say buy.
The rent is real, but debt and Macquarie are paid first. The discount prices that funding path; it is not a mistake.
Bulls read the $36.2B as if it all reached the shareholder; bears read the accounting loss as if the rent were unsigned. Macquarie has funded $1.8B of preferred units that accrue 12.75% a year, compounding, and must be repaid at least 1.8 times over before the common shares see anything.
What would prove us too cautious: the remaining 810 MW financed near the 7.00% just achieved, with the first full campus holding its 91% margin.
What analysts say, and the evidence
The evidence under this - 5 claims
What could go wrong, and how we would know
The risks
- Rent arrives late. Next year's contracted rent covers 93% of the interest bill, before a dollar of operating cost.
- More to build than it can. Management cites a supply-chain ceiling near 700 MW a year, and has contracted to build about 1.5 GW over the next couple of years.
- Insiders sold, then two bought. After a $122.1M award in January the chief executive sold 165,000 shares. On 16 October he and the president bought 355,803 at $18 to $19.
- Figures restated before. Operating income for one 2024 quarter was restated by 166.7% from what was first reported.
The evidence under this - 5 claims
- Cash interest is materially offset today: the company earned $55.7m of interest income in fiscal 2026 on a cash pile that ended the year at $3.97bn, and a large share of interest incurred is capitalised into construction rather than expensed ($39.6m in fiscal 2026)
- Contracted minimum rent is not the only receipt - tenant recoveries, variable rent and fit-out revenue all sit outside the Note 17 ladder, and fit-out alone was $152.4m in the fourth quarter
- The comparison is also a moment in time rather than a solvency test: the FY28 ladder step to $1.45bn more than triples the coverage, and the company held $1.59bn of unrestricted cash at year end
- The company has repeatedly raised that ceiling and says so, and its delivery record supports it: the first Polaris Forge 1 building took about 24 months from start of construction to ready-for-service and the second took under twelve
- Against that, the admission is unprompted and specific - 'we're definitely exceeding that a little bit' - and it is the one growth constraint management named without being pushed
- Plan use tracks company maturity, and on the calibration this repo keeps - Amazon 128 of 128 sales under a plan, Meta 240 of 250, NVIDIA 346 of 375, against St. Joe and LightPath at zero - a company four years from its listing sitting at one of twenty is not obviously anomalous
- The bought-to-sold ratio for the last complete quarter was 0.46 against a twelve-quarter median of 0.49, so the level of selling is entirely ordinary for this company; it is the absence of a plan, not the volume, that stands out
- Most of the selling clustered in late October 2025 at $33 to $36, and the shares subsequently reached $50.73 - so these were not sales at the top
- The purchases are 356k shares, about $6.6M, against 23.4M shares the chief executive holds (17.6M pledged) and the January award; a purchase at a one-year low is a statement of confidence but small beside the stake and the pledge, and does not change any claim about leases or debt.
- 2026-10-23editedForm 4s of 16 October show two officers' first open-market purchases in the period; 'no purchases' was no longer true
What we still do not know
- What each building costs and how much of it is borrowed. The $11M to $13M a megawatt and the 80% debt share exist only on the call.
- What rent per megawatt the leases carry. Neither the company nor its unnamed tenants publish it.
The evidence under this - 2 claims
- No filing states a cost per megawatt. It is checkable only crudely against fiscal 2026 capital expenditure of $3.04bn, which straddles one completed 100 MW building, a partially complete 150 MW building and groundworks at three more campuses, so it neither confirms nor refutes the range
- It is a forward figure on a programme that has delivered 175 MW. Management's own framing is that it aims to 'set expectations that we can at least meet, if not beat', which is a stated bias in the number
- The company concedes the constraint that would push it up: it has cited a supply-chain limit near 700 MW a year and has contracted to build roughly 1.5 GW over the next couple of years
What would change our view
- Neutral→AttractiveThe campuses deliver and the financing is in hand
It needs all of:
- ○Each campus opens within two quarters of its promised year, and a fully loaded campus keeps at least 80 cents of each rent dollar.checked at the next results, due by 6 Jan
- ○Macquarie extends its funding to the three remaining campuses, and new project debt costs no more than the 9.25% first placement.watched as it happens
- Neutral→UnattractiveThe senior claims or the leases give way
It needs any of:
- ○New project debt costs more than the 9.25% first placement, or Macquarie declines to fund the remaining campuses.watched as it happens
- ○A tenant is let out of a lease without paying the full rent still owed.watched as it happens
What is coming
- 4 November 2026 Annual meeting and say-on-pay vote: Whether holders accept the five-year awards and the new bonus formula after only 52.6% backed pay last year; a second weak vote would mark a governance discount.
Every dated event ahead
- 4 November 2026 Annual meeting and say-on-pay vote: Whether holders accept the five-year awards and the new bonus formula after only 52.6% backed pay last year; a second weak vote would mark a governance discount.
- December 2026 Polaris Forge 2 reaches initial capacity: The first building for the second investment-grade tenant; on time keeps the delivery record intact, a slip starts the two-quarter invalidation clock.
- 6 January 2027 Q2 FY2027 results: Whether base rent keeps rising faster than fit-out pass-through and the net operating income margin recovers toward 91% as the third Polaris Forge 1 building is built.
- December 2027 Delta Forge 1 begins operations: The first southern campus for the hyperscaler holding about $20bn of the rent; whether the North Dakota build model travels to a new region and power market.
Since we wrote this
- 23 OctNewsChief executive and president bought shares at a one-year lowTheir first open-market purchases in the period, 356k shares near $18 to $19. A signal of confidence; leases and debt, which hold the view, did not change.
- 21 OctEarningsCall read: higher rents hoped for, none signed; price re-struckManagement expects about 250 MW of new leases at rents over 15% higher, but nothing is signed yet. After a 26% fall the shares price the rent at 11.3 times.
- 14 OctBet moveBet re-struck: Priced for success to SpeculationRe-priced at the 13 October close; the rating moved from Expensive bet to Fair bet.
- 8 OctEarningsBase rent up to $65.8M; margin 89%; view heldRent is rising and 250 MW is live, but fit-out still drives growth and the margin sits below 91%. No criterion triggered.
- 7 OctNewsApplied Digital's Finland power deal is an untenanted option; the case held. View heldApplied Digital secured an agreement for up to 1 GW of potential power capacity in Finland, its first site outside the United States.
- 5 OctThesis reviewThe bet now names what it ridesThe owner asked that every bet name the macro trend behind it, after BWXT's bet was missing its chief executive's case for nuclear over gas.
- 4 OctThesis reviewOwner's checklist: revenue streams, cost of growth, management, catalystsThe owner's checklist found the method silent on who runs the company, what growth costs, where the money comes from and what is dated ahead; this pass fills those four blocks from the proxy and filings.
- 3 OctNewsApplied Digital says Polaris Forge 1 is now 250 MW operating; the case held. View heldApplied Digital brought another 75 MW online at Polaris Forge 1, taking it to 250 MW operational.
- 27 SepPrice moveShares up about 10%; the rent yield priced in is 8.2%The price rose to $26.25 from the $23.96 the thesis quoted, with nothing new from the company. It now prices the finished portfolio at about 12.3 times rental profit; the case still waits on the 8 October results.
- 15 SepThesis writtenNeutral, medium convictionWritten from the fiscal 2026 annual report, whose lease table showed 94.7% of the signed rent arriving after fiscal 2028, and Macquarie's preferred stake paid ahead of the shareholder.
Also checked 5 times with nothing new to act on.
Sources 25
Sources - 25
Every claim above cites one of these, or states the reasoning that derived it. The quote is frozen at the date it was read: a link can rot, and an agent fetch returns a reading of a page rather than the page. Backs counts the whole claim ledger, so it can name a claim this page does not surface.
- S01primarycompany filing2026-07-29the contracted lease portfolio table, Note 17's minimum payment ladder, Note 5's customer concentration, Note 8's debt components and maturity ladder, Note 16's Macquarie preferred terms and waterfall, Note 11's TopCo 2 balance sheet, Note 15's stock compensation, the NOI reconciliation, the balance sheet and the cash flow summarybacksC01C02C03C04C05C07C09C12C14C16C17C19C20C22C23C32C35C37C38C41C42C43C44C45C50C52C56C57C58C59
- S02primarycompany release2026-07-27the quarterly non-GAAP summary, the post-quarter $1.59bn 7.00% placement, the CoreWeave credit restructuring and A3 refinancing, the Delta Forge 2 lease and the ChronoScale separationbacksC12C16C20C32C42C43
- S03primarytranscript2026-07-27Applied Digital fiscal fourth-quarter 2026 earnings call transcriptthe $11-13m per megawatt build cost, the 80% loan-to-cost and three-quarters Macquarie equity split, the $2bn contracted NOI and $1bn run-rate goal, the fourth-quarter fit-out and base rent decomposition, the 700 MW supply-chain ceiling, and thirteen analyst turns from seven analysts - the questions being a separate origin from the answers, and the first of them the lease-yield objection this thesis takes seriously
you can do the math and kind of walk through that we have about $2-plus billion of net operating income contracted on an annual basis at this point
backsC05C15C24C25C26C32C39 - S04primaryvendor data2026-09-27Research pack, frozen at content/APLD/research-pack.jsonprice and the return decomposition, the analyst rating series, FINRA consolidated short interest, the quarterly revenue and margin series, the restatement record and the segment instance readbacksC33C34C35C41
- S05primarygovernment data2026-09-15the underlying rate of change, measured by a statistical agency rather than the industry
Servers alone accounted for an estimated 7% of commercial sector electricity consumption in 2025. Data center server electricity use grows to 22%-33% of commercial building electricity use by 2050 across our cases.
backsC13 - S06secondaryindustry data2026-09-15vacancy at 1.4%, 80.4% of under-construction capacity pre-committed, under 1,500 MW available - the scarcity the moat rests on, measured outside the company
primary market vacancy fell to a record low of 1.4% in H1 … commitments made on 80.4% of all under-construction capacity … Total under-construction capacity across primary markets increased by 24.8% in H1 to a record high of 7,481.1 MW … Less than 1,500 MW of future capacity across all primary markets remains available, representing approximately six months of demand
backsC06C11 - S07secondaryindustry data2026-09-15Synergy Research Group - hyperscale operators to account for 67% of all data center capacity by 2031 Synergy Research Groupthe size of the substitute - how much of its own capacity the concentrated buyer builds rather than leases
Almost 60% of that hyperscale capacity is now in own-built, owned data centers with the balance being in leased facilities … they now account for 48% of the worldwide capacity of all data centers … by 2031 hyperscale operators will account for 67% of all capacity
backsC08 - S08primarycompany release2025-10-15AI Infrastructure Partnership, MGX and BlackRock's Global Infrastructure Partners to acquire all equity in Aligned Data Centers Aligned Data Centersan actual transaction price for a closely comparable platform - approximately $40bn of enterprise value - sold by funds managed by the same asset manager that holds Applied Digital's TopCo 2 preferred
Aligned's portfolio includes 50 campuses and more than 5 gigawatts of operational and planned capacity, including assets under development
backsC31 - S09primarycompany filing2026-02-27a named competitor's own revenue split - HPC lease revenue at 10% of total - which is what decides the cohort questionbacksC10
- S10primarycompany filing2026-08-27AI Cloud Services at $128.8m of $707.0m of revenue, and the fact that IREN sells GPU compute rather than leasing shellsread, but no claim rests on it
- S11primarycompany filing2026-03-02ten data centres, roughly 920 MW of leasable customer power capacity, and a revenue mix still shifting out of mining into high-density colocationread, but no claim rests on it
- S12primarycompany filing2026-02-25Compute revenue at $202.3m of $235.1m, confirming the whole-company multiple prices mining rather than AI leasingread, but no claim rests on it
- S13primarycompany filing2026-09-15SEC Form 4 filings, read for the Rule 10b5-1 flag via scripts/insider_fetch.pynineteen of twenty open-market sales discretionary, no open-market purchases, and the January 2026 grant-then-sale sequencebacksC36C37
- S14primaryindustry data2026-08-11a broker's North American absorption, the capacity under construction and the owner-occupied part of it, the tenant mix and rent growth; read through ir_fetch
North America data center demand reached an all-time high in the first half of 2026 with 25 GW of absorption, double the level from a year ago and five times that of two years prior. … North America now has 66 GW under construction, 95% of which is pre-committed. … Hyperscalers remain the dominant force, accounting for 59% of 2026 data center tenant demand while meeting capacity needs through both leasing and self-building, with 28 GW of owner-occupied capacity currently under construction.
backsC46 - S15primaryindustry data2026-01-05a broker's forecast of global capacity to 2030 (about 103 GW to 200 GW), the Americas' share and its 17% supply growth rate; read through ir_fetch
The Americas is the largest data center region , representing about 50% of global capacity. The Americas also has the fastest growth rate of the three global regions, with a projected 17% supply CAGR through to 2030, preserving its position as the dominant data center region. The U.S. drives most of the activity in the region, accounting for about 90% of capacity in the Americas.
backsC47 - S16secondaryindustry data2026-08-28Data Center Frontier - CBRE: Record Data Center Construction Fails to Ease Capacity Crunch (reporting CBRE North America Data Center Trends H1 2026) Data Center Frontier, reporting CBRECBRE's primary-market net absorption and supply for the half and Northern Virginia's large-deployment asking rent; cbre.com answers 403 to every route press_fetch tried, so CBRE's figures are read from this trade-press account, same origin as S06
Net absorption reached 1,456.2 MW, up 11.7%, as hyperscale cloud and AI infrastructure operators continued competing for increasingly scarce blocks of contiguous power and capacity. … For 10-MW-plus requirements, CBRE reports asking rates of roughly $160 to $185 per kW per month.
backsC48 - S17primarycompany filing2026-07-28590 MW leased to CoreWeave with 395 MW billing, and the 27 July 2026 leases of 377 MW to AMD and 152 MW to a neocloud, with AMD's right to reserve up to 1,925 MW morebacksC49
- S18primarycompany filing2026-08-10the Helios campus: 133 MW plus 393 MW of critical IT load leased to CoreWeave on 15-year termsbacksC49
- S19primarycompany filing2026-08-04700 MW of HPC data centres in development across three sites for hyperscaler tenants, including about 300 gross MW for an investment-grade hyperscaler at Black Pearl and the Fluidstack lease at Barber Lake backstopped by GooglebacksC49
- S20primarycompany filing2026-08-0581 critical IT MW of HPC leasing energised at Lake Mariner, the Fluidstack leases amended to 168 MW, and near-term gross capacity of about 500 MW therebacksC49
- S21primarycompany filing2026-09-25beneficial ownership at 16 September 2026 and the CEO's pledge, the executive officers and their tenure, the CD&A (FY2026 bonuses, the five-year RSU and PSU awards and their hurdles, the FY2027 bonus formula, the 2025 say-on-pay result) and the related-person transactions with Base Electron and ChronoScalebacksC53C54C55C57
- S22primarycompany filing2026-10-07quarter to 31 August 2026: revenue by line, base rent, tenant fit-out, net operating income and its margin, cash and debt, live capacity and the subsequent eventsbacksC04C60
- S23primarycompany filing2026-10-07Note 5's customer concentration for the quarter (56%, 21%, 11%), the HPC segment split into $183.5M services and $79.1M rental revenue, and the subsequent events: 128,750 Series G shares converted into 5.2 million common shares, the sale of the legacy Ekso business and the Base Electron power agreementbacksC60
- S24primarytranscript2026-10-07Applied Digital fiscal first quarter 2027 earnings call transcriptthe question section: eight analysts on North Dakota demand, financing options for the next three campuses, the Finland site, the 250 MW of expansion leases at higher rates, supply-chain commitments, the ChronoScale build-out and the pace of delivery to 3.5 to 4 GW by 2030
we fully expect to do some lease expansion and new campuses at materially higher rates versus what we signed our last contracts for in the first half of 26.
backsC61 - S25primarycompany filing2026-10-23SEC Form 4 filings of 16 October 2026 (Wes Cummins, Jason Zhang), read via scripts/insider_fetch.pytwo discretionary open-market purchases on 16 October 2026: 55,803 shares at $18.04 by the chief executive and 300,000 at about $18.86 by the president, neither under a Rule 10b5-1 planbacksC36