IREN

IThe business

IREN rents out AI computing power. It owns the land, the electricity connection, the building and the graphics processors inside it, and it sells time on them to companies that train and run AI models.² The filing calls this three layers: data centers, compute and software. Until this year it was known for something else. IREN, formerly Iris Energy, was founded in 2018 as a bitcoin miner with data centers in British Columbia and later Texas, and bitcoin still made 82% of revenue in fiscal 2026 ($578.2M of $707.0M).² It plans to finish replacing the bitcoin machines with AI hardware by the end of December 2026.²

The fiscal year ends on 30 June. IREN lists on Nasdaq and is incorporated in Australia. It filed annual reports on Form 20-F under international accounting standards until fiscal 2024, then switched to the US domestic form, the 10-K, and US accounting rules from fiscal 2025.⁸⁹ That matters for this page in two ways. The audited history is short, five years, so I cannot show ten. And the figures before fiscal 2025 come from a different accounting rulebook, which I mark where it matters.

What it sells

Customers rent a cluster of graphics processors, called GPUs, in a data center IREN runs. Most rent them as bare metal, meaning the raw machines with no software on top, under contracts of three to five years that fix the capacity, the price and the date of delivery. Many pay part of the price in advance.² Revenue starts only after the machines are delivered, tested and accepted. That is the first thing to understand about the numbers: the revenue is signed long before it exists, and it only begins when a very large construction job is done.

How large the job is shows in the Microsoft contract. In November 2025 IREN agreed to give Microsoft GPU services in four stages at its Childress, Texas site, a total of about 200MW of computing load, for about $9.7B over a five-year average term.²¹⁰ A megawatt of IT load (MW) is the electricity the computers themselves draw, before cooling, and it is the unit the industry prices by. Dell is supplying the machines under a purchase agreement of about $5.8B.¹⁰ So the machines cost about $29M per MW, and the contract pays about $9.7M per MW each year C (calculated). At that price the machines pay for themselves in a little over three years, before the building. Horizon 1, the first 50MW, was delivered to and accepted by Microsoft in August 2026. Horizons 2 to 4 are due between the fourth quarter of 2026 and the second quarter of 2027, with grace periods in the contract.²⁶

Other contracts are priced higher. In May 2026 NVIDIA signed a five-year, $3.4B contract for its own AI work. In July 2026 three AI developers, Prometheus, Perplexity and Together AI, signed about $2.8B between them. In August a frontier AI lab signed a further contract whose size is not given.²⁵ IREN says recent three-year contracts bring more than $20M of revenue per MW a year, about a two-year payback on the machines, and that it is in talks at about $25M.⁵ Together the signed contracts and leases come to about $16.6B, of which $5.1B counts as unsatisfied revenue obligations and $11.4B as leases.² Management says $4B of annual run-rate revenue is contracted for 2026 capacity, and $1B was operating on 26 August.⁵ I divide the $4B by the 300MW of IT capacity IREN targets for the end of 2026 and get $13.3M per MW C (calculated), between the Microsoft price and the newer ones.

The size of the build

IREN had about 40MW of AI capacity operating at 30 June 2026. It targets 0.3GW of IT capacity delivered cumulatively in 2026 and 0.8GW in 2027, and says it holds grid connection agreements or equivalents for about 5GW across the United States, Canada, Spain and Australia.²⁵ The announced projects are Childress (750MW), Sweetwater 1 and 2 (2,000MW together) and Kiowa, Oklahoma (1,600MW) in the United States; Bundey in South Australia (800MW); Badajoz in Spain (300MW, bought with Nostrum in June 2026); and three sites in British Columbia (160MW together).² Gross megawatts include cooling and losses, so the IT figures above are smaller.

The company said it had $14B of cash, committed GPU financing and customer prepayments against its plans, and at 30 June it had commitments of $13.8B, up from $0.4B a year earlier.⁵² The cash at that date was $5.9B, plus $1.7B restricted for the Microsoft financing.² The market value of the whole company, at $23.95, is about $9.7B on the 404.6M shares outstanding, or $11.0B counting restricted units and notes that convert C (calculated).¹³ Against the $13.8B it has committed to spend, that is the first fact to keep in mind: the company is about to spend more than its market value, mostly on borrowed money.

Who owns it

The two co-founders and co-chief executives, Daniel and William Roberts, each hold 13.2M ordinary shares, 3.22% each, and one B Class share each.³ Each B Class share carries 15 votes for every ordinary share its holder owns, which gives each brother 19.76% of the votes and the two together 39.5% C (calculated). The B Class shares must be redeemed by 17 November 2033, or earlier if the holder leaves the board.³ Goldman Sachs and Bank of America each reported more than 5% in filings dated August 2026; bank filings like these can include client and trading positions.³ Of named funds, Situational Awareness LP, run by Leopold Aschenbrenner, held 9,474,099 IREN shares worth $433.3M at 30 June 2026, according to its Form 13F-HR filed 14 August 2026.¹¹ I give that as a dated fact only. The filing shows no shorts or non-US holdings, and it says nothing about what the fund has done since. The shares were $45.73 on that date, so the price has fallen 48% since C (calculated).¹

IIThe moat

A durable business keeps rivals from taking its profits. For IREN the test is what rivals cannot copy, and what they can. The evidence points two ways, and the answer differs by layer.

What is hard to copy

Electricity. A 100MW data center needs a utility to agree to deliver 100MW, and that takes years of studies, substations and permits. The 10-K says IREN holds executed grid connection agreements, letters of agreement or equivalents for about 5GW.² The phrase "or equivalents" is worth noticing: not all of that 5GW is a signed right to take power, and the filing does not say how much is. Even so, power at this scale is the slowest thing to build, and the one IREN started on earliest. British Columbia now limits new electrical capacity for data centers through a process run by BC Hydro, which raises the value of the 160MW IREN already holds there and shuts that door to newcomers.²

The second is speed of delivery. Horizon 1 was built, filled with machines, tested and accepted by Microsoft between the first Dell deliveries, scheduled from March 2026, and acceptance in August 2026.⁶¹⁰ NVIDIA tested the GB300 NVL72 cluster there and named IREN an Exemplar Cloud provider.⁶ That is a record few have, and it is what buyers pay for when computing is short.

What is not

The GPUs. IREN buys them from NVIDIA, AMD and server makers such as Dell, on the same terms as its rivals, and says in the 10-K it sources from a limited number of suppliers.² The software layer is thin: it bought Mirantis in August 2026, for 12.6M shares and about $40M, to add it.¹² A tenant can leave when a contract ends. Prices fall as new GPUs arrive; the 10-K warns that older machines may rent for less.² The competitors it names are the large clouds, Amazon, Google, Microsoft and Oracle, and the specialists, CoreWeave and Nebius among them.²

The price paid shows how thin the protection is. Microsoft pays about $9.7M per MW a year, and the newest contracts above $20M. Same company, same year, the same kind of machine. If scarcity made the price, Microsoft would not have it for half. The gap says that buyers with the most credit negotiate hard, and that the price IREN gets depends on who is buying, not on a right IREN holds E (estimate).

My view: the power is the protection, and the rest is execution. That is worth something, because power takes years, but the gain is the ability to build, not the right to earn a high price. In the filings I find no sign yet of the kind of pricing power, such as price increases with customers staying, that real protection would show. The pricing in the newest contracts is higher, and I give weight to that, but it comes in a market where computing is short, and shortages end.

IIIThe competitive advantage

The honest question is what IREN does better than other builders of AI data centers, and whether the difference shows in the numbers. There are three candidates.

Owning every layer. IREN designs and builds its own data centers rather than leasing space. The filing argues that ownership avoids termination rights, profit sharing and price changes that come with colocation.² That is real for a tenant that needs a ten-year plan. It also means IREN carries the construction risk, the equipment risk and the financing risk of all three layers. The data center layer is the one that lasts; the buildings are depreciated over 20 years, and the GPUs over five.²

Starting early. The company converted from bitcoin to AI while others were still debating it. By June 2026 it had $2.0B of GPU hardware on the balance sheet and $3.7B of construction in progress.² The early start explains why it had power and sites ready when Microsoft came. It also leaves it with $639M of impairments on stranded bitcoin equipment and data center parts in fiscal 2026, because the air-cooled halls built for mining have to be rebuilt for liquid-cooled GPUs.² The advantage of starting early was paid for in a write-off.

Financing. IREN is unusually inventive here. It borrowed $3.6B against the Microsoft contract at an average of 6.0%, which with customer prepayments covers 96% of the machines' cost, according to the company.⁵ For customers with weaker credit it raised $2.4B at a fixed 9.0% from Blue Owl and PIMCO-advised investors, covering 90% of the machines at Mackenzie.⁵ Prepayments from customers have been 45% to 55% of GPU spending in recent contracts.⁵ The rate difference shows how lenders rank the buyers: Microsoft at 6%, the others at 9%. I count this as a real skill, and a real cost. The skill is turning a signed contract into debt. The cost is that the debt falls on the project only so far: the parent has guaranteed the machines' cost if Microsoft validly terminates a tranche and no replacement customer is found, and has guaranteed hedge payments and the data center's performance.²

The evidence in the numbers

Revenue per MW is the clearest test. The older, mixed book earns about $13.3M per MW, and the newest contracts above $20M, according to the company.⁵ If the newest contracts hold, IREN's building skill is turning into price. If they do not, the advantage is only that IREN could build at all. The next several quarters of acceptance and billing will show which, and I list the indicators in Section VII.

NVIDIA appears in four roles at once: supplier of the machines, partner (IREN is an NVIDIA Cloud Partner and Preferred Partner), customer ($3.4B contract), and potential investor, with a right to invest up to $2.1B in IREN shares conditional on delivery of up to 600,000 GPUs.² A buyer that also supplies and invests deserves care. The contract may be good for both. It also means that part of IREN's demand is funded by the company whose machines it buys E (estimate).

IVManagement

Daniel and William Roberts, brothers, founded the company in 2018 and run it as co-chief executives. A separate independent chair, David Bartholomew, sits above them.³ The board is divided into three classes, except the founders, who hold their seats as founder directors.³ I judge managers by three things: how they are paid, how much of their own money is in, and what they did with other people's money.

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

Each brother's total pay in fiscal 2026 was about $50M ($49.9M and $50.0M), against $114,400 for the median employee, a ratio of 436 to 1.³ Nearly all of it is stock. In 2025 the board gave them time-based and share-price-based units. The price hurdles ran from $20 to $50 on a 30-day average, set against a 60-day average of $7.72 at grant. All six were reached by October 2025, and each brother earned 1,844,862 units on the way up.³ Another 1,844,862 time-based units each vest through June 2031.³

On 30 June 2026 the board went further. It granted each brother 9,099,328 time-based units, with an accounting value of $275.9M each, vesting in four yearly parts with a further two-year holding period after each. The grants replace any other long-term equity for them until fiscal 2031.³ The proxy says the awards, with the 2025 grants, come to about 3% of the company each, in line with a set of 28 founder-led companies the board's consultant reviewed.³ I read the structure plainly. It pays them for staying and for the share price rising. It does not pay for return on capital, for cash per share or for avoiding dilution, and the units themselves add to the share count. A fixed grant that cannot change for four years also means the board has no later lever. The two-year holds after vesting are the part that ties them to the share price after they have been paid, and I count that in their favour.

Their own money

Each brother holds 13,189,713 shares, 3.22% of the ordinary shares, which at $23.95 is about $316M each C (calculated).³¹ That is a large stake, and it is real, but a fair comparison is what they took out. In September 2025 each sold vested options through trusts to B. Riley for about $24.5M, and the company helped to arrange it.³ That is about 8% of the value of the stake each now holds. I do not read that as a signal either way; options have to be exercised and paid for.

What they did with capital

The record is short and it is entirely a growth record, so I describe it by facts.

• Shares. The count rose from 55.0M at 30 June 2022 to 404.6M at 28 September 2026.⁴³ Under the at-the-market plan IREN sold 47.2M shares for about $2.5B, an average of $53 a share, and it has up to $6.0B of such sales authorised.² In December 2025 it sold 39.7M shares at $41.12 in a registered direct offering.² Those prices are well above today's $23.95. Whatever one thinks of the dilution, the company sold into strength, and the cash it took in is in the bank or in the buildings.
• Debt. It issued $6.3B of convertible notes in the year, at coupons between 0% and 1%, with conversion prices from $51.40 to $85.63, and $525M of senior notes and $413M of bank loans have been drawn for the Microsoft GPUs.² The low coupon comes at a cost: when the share price is above the conversion price the holders share in the gain, and IREN paid $448.9M for capped call contracts to offset part of that.² In December 2025 it paid $1.63B to retire $544.3M of older 2029 and 2030 notes, an inducement that cost a $111.8M expense in the year.²
• Acquisitions. It bought Nostrum (Spain) in June 2026 for 837,424 shares and about EUR 82M, and Mirantis in August for 12.6M shares and about $40M, a total of about $544M.²
• Write-offs. It wrote off $638.8M in fiscal 2026, mostly bitcoin hardware, and a further $110.6M on the fall in value of machines held for sale.²

Overall the pattern is of an operator that moves fast and raises capital when it can, and a company in which the people running it are paid mostly in stock. That is a combination that has done well while the share price rises and is untested when it falls.

VFinancials

The audited record is five years: IREN listed on Nasdaq in November 2021.⁴ Ten years of results do not exist, so I show five and the model starts from the latest balance sheet instead. Figures for fiscal 2022 are under international standards and later years under US standards, which changes net income only a little for fiscal 2023 and 2024 ($171.9M and $28.9M losses against $171.8M and $28.9M).⁸⁹

Revenue went up 12 times in four years and the company lost money in four of the five years. The one profit, $86.9M in fiscal 2025, includes a $77.5M unrealized gain on financial instruments.⁹ Fiscal 2026 is easy to misread. The net loss of $702.6M fell mostly from non-cash lines: $638.8M of impairments, a $110.6M fall in the value of machines held for sale, $111.8M of inducement cost on the note repurchase, and a $558.5M gain on the value of capped calls that went the other way.² Adjusted EBITDA, the company's own measure, was $245.7M against $269.7M a year earlier, so on that measure profit fell as revenue rose 41%.⁵ The cause is cost running ahead of revenue: selling, general and administrative costs rose from $136.5M to $449.1M, of which $162.4M was extra stock pay and $52.8M extra payroll tax on stock awards.² AI revenue was $128.8M against $16.9M of direct cost, and the year's $417.7M of depreciation covers both businesses.²

Cash is not profit

Fiscal 2026 operating cash flow was $2,100.4M, and the net loss was $702.6M. The difference is $1,841.7M of customer prepayments, recorded as deferred revenue.² That is real cash, but it is a down payment on service not yet given. I do not treat it as earnings, and in my valuation I subtract it as a debt-like claim: the revenue it pays for is in my cash flows, so counting the cash too would count it twice.

What an owner can take out is operating cash flow, less the machines and buildings that must be bought, less the stock paid to staff. For fiscal 2026 that is −$2,438M counting the prepayments, and about −$4,280M without them C (calculated). In the three years from fiscal 2024 the total is about −$4.1B C (calculated). Reported earnings, in the best year, were $86.9M. The two have not been close, and for a business that is building, they are not expected to be: the capital is going in ahead of the revenue. What matters is when they meet, and my model, in Section VI, has them meeting in fiscal 2030.

What is owed

The balance sheet at 30 June 2026 was $15.8B of assets and $11.6B of liabilities.² Here is every claim ahead of the shares, from the filing.

The convertible notes need explaining. They are bonds that holders can swap into shares at a set price. Four series, $6.3B of the $6.7B, convert at $51.40 to $85.63, far above $23.95. The 2029 and 2030 series, about $446M now after the repurchase, convert at $13.64 and $16.81, so they are worth more as shares: that is about 29.7M shares, which I count as shares and not as debt C (calculated).² At my bull value all six series would convert, adding about 100M shares.

VIValuation

First, how this goes badly wrong

Before any number, I ask what would most likely destroy the investment, and work backward. For IREN there are three ways.

The price of computing falls, and the debt does not. Each megawatt costs about $46M to build and earns, in my base, about $15.7M a year in cash for three years and about $10.2M for the next two E (estimate). If the price a megawatt earns is 10% lower on new contracts, my model gives the shares a negative value, because the loans come due at the same amount C (calculated). GPU rental prices have a history of falling as new chips arrive, and the 10-K says so.²

The money runs out before the cash arrives. In my base IREN spends $19.6B in fiscal 2027 and $12.9B in fiscal 2028 on buildings and machines, and the cash it makes in those years is $2.7B and $6.7B before tax E (estimate). The gap must be borrowed or raised from shareholders. At today's price a $10B share sale would be about 420M new shares, more than doubling the count.

A buyer or a delivery fails. Revenue starts only on acceptance. Delay credits, service credits and termination rights are in the contracts, and some customers are early-stage private companies.² The Microsoft debt is guaranteed by IREN if Microsoft validly ends a tranche and no replacement is found.²

What an owner can really take out

In fiscal 2026 the company reported a $702.6M loss, which is not what an owner kept; it also reported adjusted EBITDA of $245.7M, which is not either. After stock pay and the capital spent, the owner's number was −$2,438M counting prepayments received C (calculated).² Stock pay is a real cost: $205.0M in the year, with $52.8M more in payroll tax. I deduct stock pay in every year of my model, at $350M in fiscal 2027 and rising 5% a year E (estimate). The distance between reported earnings and owner cash is large and it will stay large until the building slows.

What I cannot judge

Three inputs cannot be known: the price a megawatt will earn in 2029 and after; whether GPUs bought now keep their value for five years or three; and how much of this the capital markets will finance. I can say what each does to value, and I show the range instead of a single number. The range is wide: bear $0, base $21, bull $141. I do not think a narrower range would be honest.

The model in plain words

I model the business as a fleet of megawatts. A cohort of 300MW (IT load) is the 2026 capacity already contracted. After that I add megawatts each year. Every megawatt costs the machines, $36M, and the building and power connection, $10.5M E (estimate). Revenue per megawatt starts at $19.0M a year for the newer contracts, falls 4% for each later year's vintage as rivals catch up, and drops to 65% of that in the two years after a three-year contract ends. After five years the machines are replaced. Costs are 12% of revenue plus corporate cost of $400M a year, stock pay is deducted as a cost, and tax is 22% once the early losses are used. I discount the cash flows of the 2026 contracts at 8%, close to the 6.0% to 9.0% IREN pays on contract-backed debt, and everything else at 11%.⁵ Cash flows after fiscal 2036 are valued as a steady fleet growing 2% a year.

The reason the model is a megawatt model is that it matches the way the company contracts and spends. The next table shows what one megawatt does in my base.

Over five years one megawatt earns $68M and costs $46M, so the first cycle roughly repays the money and the gains come later, from the second and third cycles on the building already paid for C (calculated). That is a long wait with cash flowing out to build while the first cycle is still repaying. It is why the ten-year cash flow in my base totals −$17.5B and more than all of the value is in what the fleet earns after fiscal 2036.

What I get

The bear needs explaining. In it, IREN builds only what it has already contracted and a little more, new computing pays less and less, and the first set of machines is not replaced because the numbers no longer justify it. The model then values the shares at −$45, meaning the debts and the cost of finishing exceed the cash the assets would make. Shares cannot be worth less than nothing, so I use $0. It is not a forecast of failure. It says that in a world where prices fall by about a fifth from today's contracts, a company with this much borrowed money has little left for its owners.

The weighted figure, with odds of 30%, 50% and 20% for bear, base and bull, is $39 E (estimate). It is above the price only because of the bull case's tail, and I would not lean on it: it asks the reader to believe a one in five chance of the bull.

Reverse DCF: what the price needs

Working backward, $23.95 is what my model gives if each megawatt on new contracts earns about 1.2% more than my base, or equivalently if I discount at 10.8% instead of 11% C (calculated). Neither is a large move. That is the point about this stock: the price is close to my base, and my base is a thin amount left after large debts, so small changes in what a megawatt earns separate a zero from a price many times higher. A 5% change in revenue per megawatt is worth $14 a share C (calculated).

The price against the range

At $23.95 the price is 15% above my base value of $21, and 83% below my bull value of $141. It is not a case where the market has a clear mistake to correct. It is a case where the market and I agree that this is a bet on prices staying high enough, and disagree about by how much. I give no weight to a margin of safety: three quarters of a bear value of zero is zero, so there is no price at which the shares are safe against the bear I describe.

The next best use of the same dollar is another page on this site. Oracle, the closest listed company with the same exposure to AI building, has a base value of $122 against a price of $130.67 and a base return of 11.4% a year on my cash flows.¹⁴ Rocket Lab, another business that needs outside capital before it earns, returns −8.0% a year at its price.¹⁵ IREN's base return, 15.5% a year from $23.95 to year 10 in the next section, is higher than Oracle's, but it comes with a bear of zero and borrowed money. A higher expected return and a worse worst case are not the same bargain E (estimate).

Quality and price are separate questions. As a business, IREN has real assets: power rights, buildings, and contracts with some of the best-rated buyers in the world. The quality of the earnings is not yet visible, because they do not yet exist. As a price, $23.95 is about my base value, so the price asks the reader to accept that the base is right and offers no cushion if it is not.

The case needs many things to go right at once, so I multiply it out. Horizons 2 to 4 are delivered on time, 85%. Prices on new contracts hold near $19M per MW, 55%. Lenders keep funding at 9% or better, 75%. The machines last five years, 80%. The odds I give the base are these four together: 28% E (estimate). They are not independent, since a strong market helps all four, so I round up to 35%. That is lower than the 50% I use, and I use 50% because the base is also the case where small misses are survivable. The odds that the shares are worth something close to zero are of the same size as the odds that they are worth my base, and the page is best read that way.

Where this could be in 5 and 10 years

Value here is what the model says a share is worth at that date: the present value then of the cash that comes after, less the debt at that date, per today's share count. It assumes every shortfall is borrowed at 9% and none is raised from shareholders.

For the base to hold at five years, IREN must have delivered most of the 0.8GW it targets by the end of 2027 and be adding about 100MW a year from fiscal 2032. At ten years it must still earn about $19M per megawatt on 2027 contracts, less 4% a year for each later vintage, which means the market for computing has to hold up through two replacements of the machines. The bear and bull differ on those numbers: price per megawatt, megawatts built, and the discount rate. The most likely way the base is wrong is that prices fall faster than my 4% a year, because each generation of chips adds capacity faster than demand grows. Read the table with one fact in mind: the numbers are leveraged. If the shortfall were raised from shareholders at fair value instead of borrowed, the base would be $35 at five years and $59 at ten, a return of 9.4% a year E (estimate).

These are estimates of value, not forecasts of the share price, and the price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three things cut in IREN’s favour and are easy to overlook. First, the biggest contract is the cheapest one. Microsoft’s $9.7B works out to about $9.7M per MW a year, while IREN says recent three-year contracts pay more than $20M, so the book gets better as old contracts are replaced by new ones, if the new prices hold C (calculated).²⁵ Second, customers pay in advance. Prepayments have been 45% to 55% of machine cost in recent contracts, and the Microsoft debt plus prepayments cover 96% of its machines.⁵ The company is putting in less of its own money per contract than the headline capital figures suggest. Third, the company sold shares and notes at much higher prices than today’s. It sold 39.7M shares at $41.12 and notes convertible at $51.40 to $85.63.² The money is already raised; what is left to raise is about $2.7B of committed GPU financing and the $3.5B left on the share sale plan.²

The bull case

In plain words: computing is short, power is the shortage underneath it, and IREN has power. Co-CEO Daniel Roberts said on 27 August 2026 that the company’s “2026 capacity is largely sold out”.⁵ Management says it has about 5GW of power and sites, is in late-stage talks over a significant portion of 2027 capacity, and that discussions for 2028 are moving in parallel with financing.⁵ NVIDIA, which makes the machines, signed a $3.4B contract and holds a right to invest up to $2.1B.² Situational Awareness LP held 9,474,099 IREN shares at 30 June 2026, worth $433.3M, according to its 13F filed 14 August 2026.¹¹ That is a dated holding, not a statement of why, and not a view on the price now.

A note on named views. I looked for what the All-In hosts, Gavin Baker and Brad Gerstner have said about IREN and its Microsoft contract, and for interviews and podcasts with Daniel or William Roberts. I found no transcript I could open in which the All-In hosts, Baker or Gerstner discuss IREN. A podcast listing shows a 26 March 2026 All-In episode with four AI company chiefs that includes an IREN executive, but I found only its show notes, so I cannot say what was said, by whom, or at what time. I have left out anything I could not open. Daniel Roberts’s public remarks that I could verify are in IREN’s own releases and filings, which I cite above.

Against reality

Today’s numbers support the first half. IREN delivered Horizon 1 to Microsoft in August 2026 and had $1B of run-rate revenue operating on 26 August, against a $4B target for the end of 2026.⁶⁵ If it delivers Horizons 2 to 4 on time and the new contracts pay near $19M per MW, my base is possible, and the price is about where it sits E (estimate). The distance is large: AI revenue was $128.8M in the year just ended, so the $4B target is 31 times it, and the company must build nearly 260MW more in about six months.²⁵ Its contracts allow for grace periods that run to the second quarter of 2027, which shows the schedule has room built in.²

The bull needs more, and it helps to multiply it out. Delivery on schedule, 85%. New contracts keep pricing near $21M per MW, 40%. Lenders keep funding at 9% or less through 2030, 65%. Machines hold value for five years, 80%. GPU supply keeps up with the pace, 85%. Multiplied: 15% E (estimate). They are not independent, so I use 20%. Each step looks likely, but all five together do not.

What a 10x would take. Ten times $23.95 is $239.50, which on about 558M shares, counting every convertible note, is about $134B of equity value C (calculated). My bull case, at $141, already has 2.7GW of IT load and $47B of revenue a year by fiscal 2036. A tenfold move needs about 70% more value than that: more megawatts, or more than $21M per MW, or a lower discount rate, and a financing market willing to carry it. The stock has already risen from a $1.06 close to $76.41, and it has since fallen 69%.¹ From here, tenfold is possible only if the bull case is understated, and I would not build a case on it.

Where I come down

The contracts with Microsoft and NVIDIA moved me: without them there would be no base case. The $4B target did not, because it is a target for capacity not yet built, and the company’s own filing says revenue may be materially lower. A year in which acceptance, prices and financing all land on time would raise my values; a refinancing at 11% or above, or a fall in the price on new contracts, would take the base to zero.

VIIIRisks, and what would change my view

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

• Counterparties. The 10-K says Microsoft and NVIDIA together are a substantial majority of contracted revenue, and that some customers are early-stage private companies.² In a bankruptcy the claims for future payments may be capped.²
• Delivery and acceptance. Revenue starts only on acceptance and can be reduced by delay and service credits.²
• Refinancing. $6.7B of convertible notes mature between December 2029 and December 2033, and the non-recourse GPU debt amortizes through 2031.² The company issued $7.2B of debt in one year.
• Dilution. The share count is 7.4 times the 2022 level. The company can sell up to $3.5B more under the share plan, 26.2M restricted units and options are outstanding, and the 2027 founder grants add 18.2M units.²³
• Power and policy. The Microsoft contract has fixed-price power through its financing term, but IREN is exposed to prices for any extra consumption and to the gap between the ERCOT hub and its site.² British Columbia has capped new data center power.²
• Control. The two founders hold 39.5% of the votes through B Class shares until 2033.³
• Bitcoin runoff. Bitcoin gave $578.2M of FY2026 revenue and ends by December 2026, so total revenue may dip before AI revenue replaces it.²
• Accounting. The 10-K reports a $558.5M unrealized gain and a $638.8M impairment in one year. Earnings will move with the share price and with writedowns, which is why I use cash.²

The Culper short report on IREN

Who and what. Culper Research is a short seller. It says it holds short positions in the companies it writes about and may profit if the price falls.¹⁷ On 11 July 2024 it published “Iris Energy Ltd (IREN): A Prius at the Grand Prix”.¹⁶ The site blocks automated access, so I read the Internet Archive copy; page numbers are its.¹⁷

What it alleges. Culper calls the AI story “entirely contrived”: the sites were built for bitcoin and finished before April 2023 (pp. 2 and 4). It says Childress has “zero apparent backup power” and one 345 kV line (pp. 3 and 8), and that air cooling is insufficient for GPUs in Texas heat (pp. 3 and 10 to 12). It says IREN spent under $1M per MW, while an HPC-ready data center costs $10M to $20M, and several of its benchmarks exclude chips (pp. 2 and 7). It also argues the shares were worth 52% to 79% less (p. 15).¹⁷

The company’s response. I found no point-by-point reply from IREN, and no IREN filing I could search names Culper.²⁵ On 12 July 2024 IREN, then Iris Energy, told Benzinga it is limited in what it can say about such reports, encourages debate, and that “not all engagement meets these standards”; it stays focused, it added, on mining growth, AI cloud customers and its power and land.¹⁸¹⁹ On 23 July it filed an investor-day deck that does not name Culper but has pages on backup power, cooling and cost; the table below sets them out.²⁰ Bernstein, a sell-side research firm, said IREN “has not claimed it intends to retrofit its bitcoin mining site” at Childress.²¹

Legal. A securities class action, Williams-Israel v. Iris Energy Ltd (retitled In re Iris Energy Limited Securities Litigation), 1:24-cv-07046, was filed in the Eastern District of New York on 7 October 2024. Its complaint cites the Culper report.²⁸²⁴²⁶ The lead plaintiff, named on 7 January 2025, filed a notice of voluntary dismissal on 14 March, and the court dismissed the case without prejudice on 19 March.²⁷²³ That is not a ruling on the claims. I found no refiled case; the FY2026 10-K lists no regulator action, only a separate 2022 New Jersey class action over the IPO.²⁷²

Where it touches my model. Only the cost claim touches an input. I charge $10.5M per MW for building and power on the 1,500MW I add after the 300MW already contracted, inside Culper’s $10M to $20M and level with IREN’s $9M to $11M for infrastructure. Each $1M more on building lowers my base by about $2.40 a share; at $20M it goes below zero, which I read as $0 E (estimate). Culper’s backup estimate, $1.6M to $2.2M per MW C (calculated), would take the base to $15 to $17 if it were extra E (estimate).¹⁷ I charge $46.5M per MW for machines ($36M) and building; IREN’s figures for the Microsoft build imply $43M to $45M C (calculated), about $29M of machines and $14M to $16M of building.²²¹⁰ Machines are bought again every five years, so each $1M on them moves my base by about $6 a share, and on IREN’s split it would read about $48 to $53, not $21 E (estimate). That tests my own machine cost, not Culper’s claim; I leave my values as they are. The “contrived” claim touches no input; it bears on trust.

So what. Culper wrote about Childress in 2024; Horizons 1 to 4 are new, liquid-cooled halls, so IREN’s later builds neither confirm nor refute it.⁶ Cost can be checked now: the machine and building split behind IREN’s $43M to $45M a megawatt, against my $46.5M, is what I would test. Backup for the GPUs cannot be known from outside. For Culper’s point to matter to value today, the Horizon halls would have to lack it and customers would hold uptime rights; for IREN to be right they must meet the Tier III equivalent it describes, which Microsoft’s acceptance of Horizon 1 fits but does not prove.² The short seller profits if the price falls; IREN’s founders hold 39.5% of the votes and are paid mostly in stock, so they profit if it rises.³ Watch the Horizons 2 to 4 grace periods, which the 10-K says run to the start of the second quarter of 2027, and the 10-Q for the quarter to 31 December 2026, due by 9 February 2027.²

Geopolitical risk, and what it means for the value

IREN is incorporated in Australia, listed on Nasdaq and reports in dollars. It has no state owner, no capital controls and no China listing to worry about. The exposure is narrower: where the power comes from, where the machines come from, and one customer.

Power. The 200MW Microsoft contract is in Childress, Texas, and so are most of the announced US sites. On 3 August 2026 the governor of Texas told the utility regulator and ERCOT to audit every data center in the grid queue before more are approved, and the 10-K says this may delay energization at new and existing projects.² Machines. The GPUs come from NVIDIA and Dell. US customs claims the bitcoin miners imported from Southeast Asia in 2024 and 2025 came from China, which would add up to about $100M of tariff; IREN has protested.² The 10-K also says export controls could change which GPUs it may buy and which countries it may serve.² Customer. Microsoft and NVIDIA are a substantial majority of contracted revenue, and the Microsoft debt is guaranteed by IREN if Microsoft validly ends a tranche and no replacement is found.²

The rest is small. Bitcoin mining stops by December 2026, so a mining ban or tax is worth about $0.30 a share, the $160M of bitcoin cash in my base E (estimate). Australia and British Columbia carry no value in my base. Bundey (800MW) and the 160MW in British Columbia are extra to it, and Australia is writing rules that would make data centers pay network costs and offset their power use.²

The table reruns the base case with one change each. The Texas row delays every new megawatt by a year and leaves the 300MW contracted as planned. The machine row raises every future machine purchase 10%, with no pass-through to customers. The Microsoft row removes a year of revenue on 100MW. I do not predict any of these.

None of it is in my base, and the price leaves no room for it. The price of $23.95 needs only about 1% more per megawatt than my base, so it assumes delivery on time and machines at today’s cost. A one-year Texas delay would put the value 45% below it. The machine cost is the one to watch because it repeats at every replacement cycle. My model lets no cost pass to customers, so $0 is a harsh reading E (estimate).