AI Cloud Finally Outgrows Mining, and the Street Sells the Capex Guide Instead

Results vs. Consensus

IREN reported the June quarter and full fiscal year 2026 on August 27 after the close, with the call at 5:00 p.m. ET. This is the fourth quarter of a deliberate teardown: revenue-producing Bitcoin mining hardware is coming out of energised buildings so that GPUs can go in, and for three quarters running the removal has outpaced the installation. The June quarter is the one where the two lines crossed.

Sequential comparison

The fiscal year, and the four quarters inside it

The annual columns are the more useful frame this quarter, because fiscal 2026 is the year the business changed shape. They are also, for the first time in this coverage, a clean comparison. Fiscal 2025 and 2024 are presented on a US GAAP basis in this year's Form 10-K, so the IFRS-to-GAAP break that made last quarter's year-over-year table unusable no longer applies to the annual figures.

Split into quarters, the year tells the story in one picture. Mining fell 71% from the September quarter to the June quarter while AI Cloud rose 9.5x, and the two lines met in June.

The December-quarter column is taken from the prior quarter's release; the September-quarter column is the residual of the audited annual figures less the three quarters that have been separately reported, and the four columns sum exactly to the audited totals.

Segment Performance

Two reportable segments, disaggregated for the first time during fiscal 2026 with prior periods recast. The chief operating decision maker is provided with segment revenue and segment cost of revenue only. No segment operating expenses, assets or liabilities are allocated, which is the limit on how far this table can be pushed.

AI Cloud Services

Revenue more than doubled sequentially to $70.5m and is running 9.5x the level of three quarters ago. The gross margin held at 87.0%, its third consecutive quarter above 86%, on cost of revenue of $9.2m. For the full year the segment's entire cost base was $16.9m, of which $4.5m was electricity: this is a business whose direct costs are close to trivial and whose economics live in depreciation and the cost of capital, neither of which is allocated to it.

The geography note in the annual report is worth reading alongside the growth rate. Every dollar of fiscal 2026 AI Cloud revenue was generated in Canada, which is to say at Prince George. The Childress campus, where the Microsoft Horizons and the NVIDIA contract sit, contributed nothing to fiscal 2026 revenue at all.

"For the June quarter, revenue was $137.2 million, including AI cloud revenue of $70.5 million. This was down $7.6 million compared to the prior quarter as we decommissioned mining hardware ahead of GPU installations, partially offset by AI cloud growth." — Anthony Lewis, CFO

Assessment: The segment is compounding at the pace the hardware allows and the unit economics at the gross line remain excellent. But $70.5m a quarter annualises to $282m against $4bn of contracted 2026 ARR, so the ratio of promise to delivery is still 14 to 1. What changed this quarter is that the promise now has an audited number behind it, and the first Childress capacity has been accepted by a customer.

Bitcoin Mining

Revenue fell 40.0% sequentially to $66.7m. The margin went the wrong way this time, down from 68.3% to 63.9%, which reverses last quarter's flattering "retire the worst hashrate first" arithmetic. Fixed site costs are now spread across a fleet that is shrinking faster than the cost base attached to it.

The annual report restores the operating statistics that the quarterly releases dropped, though only at annual granularity. Fiscal 2026 produced 6,075 Bitcoin against 5,499 in fiscal 2025 at an average operating hashrate of 36.5 EH/s against 25.7 EH/s. The number that matters more is the exit rate: installed mining capacity at June 30 was approximately 23.2 EH/s, roughly 36% below the year's average, occupying about 380MW of data-centre capacity.

Assessment: Mining is now a self-liquidating funding source with a stated expiry date of December 2026. At $66.7m of revenue and $42.6m of gross profit a quarter it still carries a meaningful share of the operating cost base, and its disappearance over the next two quarters is a headwind that the AI Cloud ramp has to absorb before it can show net growth. The 380MW it currently occupies is the more interesting asset: energised, connected capacity that converts to GPU revenue without a new grid connection.

Key Topics & Management Commentary

Overall Management Tone: Assured to the point of being unbothered by the quarter being reported, with the prepared remarks structured around customers, pricing and financing rather than results. The posture was less defensive than at the March quarter, because this time management had delivery to point to rather than an announcement. Where it was least convincing was on the arithmetic connecting the capital plan to the funding plan, which was presented as a set of mechanisms and precedents rather than as a bridge.

1. The crossover, and the replacement rate that finally moved

For three quarters the central number in this story has been the gap between how fast mining revenue was being switched off and how fast AI Cloud revenue was being switched on. In the March quarter the business replaced 29 cents of each dollar it gave up. In the June quarter it replaced 83 cents. The absolute revenue line still fell, but the second derivative turned, and on the current trajectory the September quarter is the last one that can plausibly print a sequential decline.

The composition milestone matters for how the equity gets valued. AI Cloud is now 51.4% of revenue, up from 3.1% three quarters ago, and it carries a gross margin 23 points higher than the business it displaced. Group gross profit was flat sequentially on a 5.2% smaller revenue base, which is what a mix shift of that violence looks like when it works.

Assessment: The teardown thesis has now been demonstrated rather than asserted. Anyone who underwrote the March quarter on the argument that the revenue decline was self-inflicted and reversible has been paid on that specific point. What remains unproven is the magnitude, not the direction.

2. The $25bn to $30bn capex guide, and why it was the entire story

The single most consequential disclosure of the day appears nowhere in the press release. It was delivered on the call, in one sentence, by a CFO on his second results presentation.

"For FY '27, we're guiding CapEx of approximately $25 billion to $30 billion." — Anthony Lewis, CFO

For scale: IREN's total investing outflow in fiscal 2026 was $4,723.0m. The midpoint of the new guide is 5.8x that, and roughly 1.9x the entire equity market capitalisation at the reaction close. The guide covers the capital for the 2026 deployments that produce the $4bn ARR target, the balance of the 2027 air-cooled programme, and a significant portion of the liquid-cooled builds landing in the second half of calendar 2027. It explicitly excludes the GPUs for those new liquid-cooled facilities, which fall into the following year's plan.

Management also flagged unit-cost inflation inside the number, expecting data-centre and GPU capital requirements to be up approximately 15% to 20% for ongoing and new deployments, with revenue increases expected to more than offset. That is a claim about pricing power rather than a measured result, and the deals that test it have not yet been signed.

Assessment: The number is defensible as a statement of ambition and indefensible as a communication decision. Putting a figure twice the market capitalisation into spoken remarks, with no accompanying slide arithmetic in the release, invited exactly the reaction it got. The substantive point is subtler: capex of this size is the mechanical consequence of a demand book the market has spent a year begging IREN to sign. You cannot applaud a sold-out 2026 and recoil from the bill for delivering it.

3. What the balance sheet already commits

Behind the guidance there is a hard, audited number that was not mentioned on the call. As of June 30, 2026, IREN carried $13,810.0m of commitments, of which $13,611.0m is payable within twelve months. A year earlier the figure was $368.8m.

Set that against the resources actually in hand. Cash, cash equivalents and restricted cash were $7,619.5m, of which $1,670.3m is restricted current and largely earmarked for Microsoft GPU capital expenditure, with a further $53.7m restricted non-current, leaving $5,895.6m unrestricted. Undrawn capacity under the investment-grade GPU facilities was $2,707.0m, being $3,645.0m of committed capacity less $938.0m drawn. That is roughly $8.6bn of identified resource against $13.6bn of contractual commitment due inside a year, before a dollar of the incremental programme.

The CFO's own framing of the funding plan is consistent with that gap, and honest about which parts are secured.

"We're targeting roughly an additional $8 billion of GPU financing and prepayments in support of GPU CapEx requirements, noting the healthy prepayments that we are seeing in recent contracting and the growing market for GPU financing that Dan has spoken to. The balance of the requirement we expect to meet through data center financing, operating cash flows and corporate sources." — Anthony Lewis, CFO

Assessment: This is the real risk in the equity and it is now quantified rather than gestured at. The $8bn target is a target, the data-centre financing market is one the company has not yet transacted in, and "corporate sources" is where equity lives. On the other hand, the commitments number is what a company looks like when it has ordered the hardware for a sold-out year, and IREN has raised roughly $19bn in twelve months with only about $3bn of it from equity. The gap is a schedule problem, not obviously a solvency problem.

4. The $1.8bn of operating cash flow is a customer prepayment

The June-quarter cash flow statement shows net cash from operating activities of $1,811.1m, against $75.3m in the March quarter. Read as a measure of the operating business that is a spectacular number, and it is not one. Within it sits a $1,722.2m increase in deferred revenue, which is customer money received in advance of service delivery. Strip it out and underlying operating cash flow was $88.9m. For the full year the same adjustment takes $2,100.4m down to $258.7m.

This is not an accounting criticism. Customer prepayments are real cash, they are non-dilutive, they carry no interest, and management is right that they are the single most encouraging thing in the funding picture. But they are financing that happens to be recorded above the investing line, and modelling them as recurring operating cash flow would be a serious error. The company's own presentation of the FY27 funding plan puts prepayments in the financing bucket, which is the correct place for them.

"And I think those prepayments are probably the most exciting part for us, when they're funding half the GPU CapEx upfront on top of the 90% financing we're getting already, these guys are sending a pretty clear signal. It's not just about contracting capacity. They're starting to finance our build-out for us." — Daniel Roberts, Co-Founder and Co-CEO

Assessment: Prepayments running at 45% to 55% of GPU capital expenditure are the most credible demand signal in the whole disclosure, better than any contract headline, because a customer that wires half the hardware cost before delivery has done its own diligence on IREN's ability to deliver. Investors should count it as evidence and not as earnings.

5. Lease accounting, and why the Microsoft contract has not touched revenue yet

The annual report contains a disclosure that materially changes how the next two years of reported revenue should be modelled, and it was not discussed on the call. During fiscal 2026 IREN determined that its dedicated GPU-capacity agreements contain leases of specified GPU equipment and dedicated data-centre space, and classified them as operating leases with IREN as lessor under ASC 842. The lease and non-lease components are combined, because the lease component is predominant. Leases commence on customer acceptance, with terms of approximately three to five years.

The consequence is stated flatly in the revenue note:

"No lease revenue was recognized during the periods presented" — Form 10-K for the fiscal year ended June 30, 2026, Note 4

So the $128.8m of fiscal 2026 AI Cloud Services revenue is entirely ASC 606 service revenue from Prince George. Of the $1,842.5m of deferred revenue on the balance sheet, $1,623.5m is deferred lease revenue for tranches that had not commenced at June 30. Horizon 1 was accepted on August 13, which is when the largest of those tranches begins to unwind into the income statement.

Assessment: Two implications. First, the reported revenue line has been understating the economics of the contracted book, and the September quarter is the first in which that reverses. Second, lease accounting means revenue recognition is tied to acceptance rather than usage, which makes the revenue ramp more mechanical and more predictable than a consumption model would be, and makes acceptance dates the single variable to track.

6. $16.6bn of contracted revenue, and the $0.9bn that lands in FY27

The most useful number in the entire disclosure package was published only in the annual report. As of June 30, 2026, unsatisfied remaining performance obligations under ASC 606 were $5.1bn, and the aggregate contracted value of lease arrangements under ASC 842 was $11.4bn. Together, approximately $16.6bn. That is an audited figure, unlike ARR, which the company is careful to describe as an operating metric that is not derived from and may be materially higher than recognised revenue.

The $0.9bn figure is a floor rather than a cap for the ASC 606 piece, since it covers only contracts in force at the balance-sheet date and excludes everything signed since, including the frontier-lab contract announced on this call. It is still the tightest constraint anyone has on the fiscal 2027 revenue line.

Assessment: This disclosure is worth more to an investor than the entire ARR presentation, and IREN spent the call on the ARR presentation. At the reaction close the equity was capitalised at roughly 0.85x the audited contracted book. That is a strikingly low number for infrastructure with these counterparties, and the reason it is low is that converting the book requires $25bn to $30bn of spending the company has not yet funded.

7. SG&A up 57%, and most of the increase is stock compensation

Selling, general and administrative expense of $128.3m was up 56.8% sequentially, and it is the line that turned a flat gross profit into a 68% decline in Adjusted EBITDA. It also rose before the Mirantis headcount arrived: that acquisition closed on August 3, after the balance-sheet date.

The annual bridge explains what is actually happening. Of the $312.7m fiscal-year increase in SG&A, $162.4m is higher stock-based compensation and a further $52.8m is accrued payroll tax on those same awards. That is $215.2m, or 69% of the increase, attributable to equity compensation. Cash employee benefits rose $23.0m. Total stock-based compensation for the year was $205.0m against $42.6m in fiscal 2025, and $42.9m of it landed in the June quarter alone.

Guidance for the September quarter is for a further step-up, and the CFO was specific about which measure.

"We expect first quarter cash SG&A to increase approximately $40 million to $50 million sequentially as we continue to invest for growth across sales and marketing, R&D, development, site and cloud operations and other functions ahead of significant revenue growth over the coming periods" — Anthony Lewis, CFO

June-quarter SG&A less stock-based compensation was $85.4m, so the guide implies cash SG&A of roughly $125m to $135m in the September quarter, before whatever the equity charge turns out to be. Headcount was 685 at June 30, with approximately 580 more arriving through Mirantis in August.

"Our headcount nearly tripled in FY '26, including hundreds of colleagues who joined through Mirantis and Nostrum, and we expect similar growth again in FY '27." — Daniel Roberts, Co-Founder and Co-CEO

Assessment: Building the organisation ahead of the revenue is the right call in a land-grab, and management said as much. The composition still deserves scrutiny: a cost line where two-thirds of the growth is equity compensation is diluting shareholders and depressing Adjusted EBITDA at the same time, since the measure adds the charge back but not the payroll tax on it. September-quarter Adjusted EBITDA will be negative unless AI Cloud revenue rises by more than the SG&A guide.

8. Pricing is up 125%, on a book that was priced before the repricing

The clearest good news in the update is the pricing trajectory, and management put numbers on it.

"So pricing has moved a lot, 3-year contract pricing is up about 125% since November, 5-year is up about 70%." — Daniel Roberts, Co-Founder and Co-CEO

Recent three-year contracts are being written at more than $20m of revenue per megawatt of IT load with a payback of roughly two years, and active discussions are at around $25m. Customer prepayments on recent deals run at 45% to 55% of GPU capital expenditure.

November is a pointed reference date, because November 2025 is when the Microsoft contract was signed. That agreement is disclosed as five years and approximately $9.7bn of total contract value across four 50MW deployments, which works out to $9.7m per megawatt of IT load per year. The anchor tenant that gives IREN its scale, its investment-grade financing and its delivery credential is contracted at pricing struck before the market repriced.

Assessment: This is the most underweighted fact in the story and it cuts both ways. It means the reported economics of the next two years, dominated by Microsoft, will understate the economics of the marginal contract, so margin should improve as the mix shifts. It also means anyone extrapolating current spot pricing across the whole 5GW pipeline is double-counting: a large slice of the near-term book is locked at half of it.

9. The customer book broadened, which was the specific test we set

In May we wrote that contracted ARR excluding NVIDIA and Microsoft had to move, and that another quarter of nothing would confirm the concentration read. It moved. Between the prints IREN announced $2.8bn of new customer contracts with AI developers in July and raised the year-end ARR target from $3.7bn to more than $4bn, with roughly 85% under contract. On this call it added a multi-year contract with an unnamed frontier AI lab, separate from the July signings.

"Together AI and Fireworks AI have both renewed and expanded." — Daniel Roberts, Co-Founder and Co-CEO

The named roster now runs to Microsoft, NVIDIA, Cohere, Prometheus, Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI, Hume AI and Higgsfield AI. Renewals matter more than logos at this stage, because they are the only available evidence that the delivered product works. Weighted-average contract term across the portfolio is approximately four years.

The counterweight sits in the risk factors, where the company states that the Microsoft and NVIDIA agreements together "represent a substantial majority of our contracted revenue." Breadth has arrived in the customer count well ahead of the dollars.

Assessment: The bear point that mattered most in May has been answered on the axis it was posed. Concentration has not been eliminated and will not be until the 2027 capacity is contracted, but the claim that IREN could only sell to counterparties with a strategic interest in its success is no longer supportable.

10. The impairment tail, now bounded

Impairments of $450.4m in the quarter took the fiscal-year total to $638.8m, against $7.2m in fiscal 2025. A further $102.1m of fair-value decrease was recorded on mining hardware reclassified as held for sale, and $25.1m of disposal losses on written-off damaged equipment.

Last quarter we complained that management guided to further charges without sizing them and that no disclosure existed to bound the tail. The annual report supplies the bound. Gross mining hardware fell from $1,135.6m to $597.0m over the year, accumulated impairment across property and equipment stands at $484.5m, and assets held for sale carry at $72.5m after the fair-value writedowns. Whatever remains to be charged against the mining fleet is a fraction of what has already been taken.

"We currently expect mining operations to be effectively decommissioned by the end of December 2026." — Anthony Lewis, CFO

Assessment: The charges are non-cash and the market has correctly looked through them, but the disclosure gap that made them unmodellable has closed. Investors can now size the remaining tail at the tens of millions rather than the hundreds, and the December quarter should be the last one carrying a material mining writedown.

11. Financing cleared at both ends of the credit spectrum, at rates the filing puts higher than the release

The most substantive operational achievement of the period is not a contract. It is that IREN financed GPUs against a non-investment-grade offtake at a single-digit fixed rate, which nobody had demonstrated a year ago.

"But let's just look at what happened 12 months ago, GPU financing barely existed as an asset class. And then in the last 3 months, we've raised $6.5 billion of it at both ends of the credit spectrum. So that's not us getting lucky with financing, that's a market forming." — Daniel Roberts, Co-Founder and Co-CEO

The two packages are a $3.6bn investment-grade facility for the Microsoft deployment, comprising an approximately $1.5bn delayed-draw term loan and $2.1bn of US private placement senior notes, and a $2.8bn non-investment-grade package including $2.4bn at a 9.0% fixed rate for the Mackenzie air-cooled expansion, led by Blue Owl and PIMCO-advised funds and signed on August 25.

One detail deserves care. The release describes the investment-grade package as priced at 6.0%, footnoted as a weighted-average rate across the private placement and the term loan excluding fees. The annual report's own effective-rate column for the same two instruments reads 7.12% for the term loan and 7.05% for the notes. Both statements are accurate and they measure different things, exactly as last quarter's "3% blended" figure did once prepayments were removed from the denominator.

Assessment: The achievement is real and it is the single best answer to the funding bear case, because it removes the requirement for a trillion-dollar counterparty on every deployment. The presentational habit of quoting the flattering version of a rate is now established across two consecutive quarters, and readers should assume a 100bp gap between the headline cost of capital and the filed one.

12. The dilution stack, and an 18.2m-share grant nobody mentioned

Shares outstanding rose from 258,103,209 to 380,710,559 over the fiscal year, and to 394,058,648 by August 14. Proceeds from issuing ordinary shares were $4,742.8m during the year, of which $1,632.4m went straight back out to repurchase $544.3m of principal on the 2029 and 2030 convertible notes, an induced conversion that also cost a $111.8m expense and issued 39,699,102 shares. The at-the-market programme has sold 47,165,838 shares for approximately $2.5bn gross, an average of roughly $53 a share.

Sitting on top of that, and disclosed only by an 8-K on July 1 and a subsequent-events note in the annual report, the board granted 9,099,328 restricted share units to each Co-Chief Executive Officer, 18,198,656 in aggregate, or about 4.8% of shares outstanding at June 30. The awards vest annually over four years with a two-year post-vest holding period on each tranche, and no expense was recognised in fiscal 2026. In exchange, neither Co-CEO receives a further equity incentive grant until the 2031 fiscal year, and the awards were approved unanimously by the independent directors after review by an independent compensation consultant. Neither the results release nor the call mentioned any of it.

Weighted anti-dilutive securities excluded from the fiscal-2026 diluted share count total 146,132,206, against 26,345,147 a year ago: 71.8m from convertible notes, 23.3m of restricted share units, 32.1m of capped calls, 8.9m of prepaid forwards, 5.5m of options and 4.5m of NVIDIA investment rights.

Assessment: The equity issuance has been well-timed rather than desperate, with the ATM struck well above the current price, and the capped calls do real work in limiting convertible dilution. The founder grant is better constructed than its size suggests: a six-year combined vest and hold, plus a five-year moratorium on further grants, is closer to a full-period arrangement than to an annual award. The objection is to the route, not the structure. A grant equal to 4.8% of the company, to the two people who control the vote, belonged in the annual results package where shareholders would see it discussed, not in a standalone filing between prints.

Guidance & Outlook

IREN does not guide to quarterly revenue or earnings. It guides to capacity, contracted run-rate revenue and, from this quarter, capital expenditure and cash operating expense. Four of the five items below moved, and the two new ones are the reason the stock fell.

The ARR walk is the number to hold onto, and the CFO gave it in full.

"We exited Q4 at roughly $0.5 billion of ARR. It's $1 billion today following acceptance of Horizon 1 by Microsoft, and that will carry through to the end of the September quarter." — Anthony Lewis, CFO

Implied ramp. Getting from $1bn of operating ARR to more than $4bn by December 31 requires Horizons 2 through 4 plus the remaining Childress, Mackenzie and Prince George capacity to be commissioned and accepted inside four months. Management says all of it is already under contract, which removes the commercial risk and leaves the construction and acceptance risk. The company delivered Horizon 1 on the schedule it named in May, which is the best available evidence on that risk.

The timing detail that matters more than the target. The step-up in reported revenue does not arrive when the ARR does.

"A significant amount of the December capacity is expected to come on late in the quarter, so we will see the reported revenue effect come through predominantly in the March quarter." — Anthony Lewis, CFO

Read literally, that pushes the reported revenue inflection to the quarter ending March 2027, one quarter later than the May framing implied. The September quarter carries $1bn of ARR for a full three months plus a $40m to $50m increase in cash SG&A, which is a combination that produces negative Adjusted EBITDA unless AI Cloud revenue rises faster than we expect.

What is locked. The audited remaining-performance-obligation schedule locks $0.9bn of ASC 606 revenue into fiscal 2027, so the balance of the consensus number depends on lease revenue commencing across Horizons 1 through 4 and the air-cooled fleet, plus contracts signed after June 30. That is achievable on the December schedule management described, and it has essentially no tolerance for slippage: a one-quarter delay on the December capacity moves roughly a quarter of the year's revenue out of the fiscal year entirely.

Guidance style. The pattern established over three updates holds: precise and reliable on things the company controls, expansive and heavily footnoted on revenue framing. Delivery dates have been met, including the Sweetwater 1 substation energisation and the Horizon 1 handoff. The GPU count target was quietly dropped this quarter without comment, which is the second operating metric to disappear in two quarters.

Analyst Q&A Highlights

Eight exchanges, and the striking feature is what was not asked. Nobody questioned the $684m loss, the collapse in Adjusted EBITDA, the SG&A guide, or the founder equity grant. Seven of the eight questions were about demand, pricing, capacity or financing, and the tone throughout was collaborative rather than adversarial.

What the FY27 capital plan covers, and how the balance gets funded

The first substantive question of the call went to the two numbers that would define the reaction: what the $25bn to $30bn actually buys, and where the money beyond the $14bn of identified resource comes from. The answer scoped the capex precisely, confirming that it excludes the GPUs for the second-half 2027 liquid-cooled facilities, and then answered the funding half by listing mechanisms and market conditions rather than a bridge.

Q: "And I just wanted to follow-up on the CapEx outlook for next year, $25 billion to $30 billion. Is that all to support the 800 megawatts that you expect to contract next year? Or is it beyond next year? And then could you maybe just talk about the financing plan beyond the $14 billion of cash, GPU prepayments and other debt financing that I think you talked about?"
— Michael Ng, Goldman Sachs

A: "So the $25 billion to $30 billion is obviously in the -- covering the financial year to June '27. So that covers all the CapEx requirements for the 2026 deployments to -- that contribute towards the $4 billion ARR target. It covers expected data center CapEx and GPU CapEx for sort of air-cooled -- the balance of the air-cooled deployments expected to come over the course of 2027 calendar year. And it also covers a significant portion of the CapEx required for the liquid-cooled deployments in the second half of 2027 calendar year. It doesn't include CapEx requirements for the GPU compute for those new liquid-cooled facilities, which will be part of the following year's capital plan."
— Anthony Lewis, CFO

Assessment: The scoping half of the answer is genuinely useful and makes the number less alarming than it first sounds, because a meaningful slice of it is 2028 capacity being pre-funded. The financing half was a list of open doors: the targeted $8bn of incremental GPU financing and prepayments, unencumbered data centres available to borrow against, and a reference to a large third-party financing partnership recently announced in the sector. No dollar bridge was offered, and the analyst did not press for one.

Contract pricing: the level, the duration and whether it holds

The pricing disclosures were the strongest part of the prepared remarks, so the obvious question was whether the headline figures describe one deal or the market. The answer was unusually direct: the numbers describe live conversations across the book, not a single outlier, and the terms are long rather than opportunistic.

Q: "And then maybe I could just follow up on some of the pricing commentary that came across quite strong, I think, in prepared remarks and the release, kind of $20 million deals what you are seeing or what you've signed. Now it's kind of somewhere around $25 million. Is that -- over what duration could you clarify? And is that more one-off? Or do you think like that is kind of like the ballpark of maybe the average you're seeing across all the conversations you're having with customers?"
— Brett Knoblauch, Cantor Fitzgerald

A: "No, we are seeing that consistently across live conversations with customers at the moment. And there are a variety of things that go into it. As Dan mentioned, we look at term length, prepayments, nature of the customer, likely growth requirements over time. But the pricing that we're seeing is relatively consistent at the moment. It continues to show an upward trend. We're seeing very strong competitive tension for near-term megawatts."
— Kent Draper, Chief Commercial Officer

Assessment: The most important answer on the call for anyone modelling 2027 and 2028. Management also volunteered that these are three-to-five-year commitments rather than spot capacity, and that it is deliberately not selling on demand today in order to build a diversified contracted base first. That is a choice to suppress near-term reported revenue in exchange for contract quality, and it is the correct one if the capacity is genuinely scarce.

Spare power inside the existing envelope

A line of questioning about power usage effectiveness produced the most quietly valuable disclosure of the day. Pressed on whether design improvements could lower PUE across the fleet, management redirected to a bigger point: the published gross-to-IT ratio is deliberately conservative, and there is headroom inside the existing grid connections that can carry more compute with no new interconnection.

Q: "We've seen a lot of deals recently with maybe lower PUEs. Are you seeing any design changes that might allow for lower PUEs at other sites that you guys have coming online or other buildings coming online in '27 and '28? Or should we kind of think of that PUE with the Microsoft deal being somewhat static?"
— Brett Knoblauch, Cantor Fitzgerald

A: "And to date, we've kept it simple, 200 megawatts of IT load for 300 megawatts of gross capacity, but we are also making it clear today that there is a reasonably sized opportunity in the portfolio to free up some of that spare power."
— Daniel Roberts, Co-Founder and Co-CEO

Assessment: Revenue that does not require a new grid connection is the highest-quality growth available in this industry, and it is entirely absent from the 5GW headline. The mechanism is unglamorous: average annual PUE runs well below the design maximum, and power-smoothing software allows a given electrical envelope to carry more racks. Unquantified, so unmodellable, but it is free option value on an asset base the company already owns.

The anchor tenant versus the AI developer

A question on preferred contract duration drew management into the clearest articulation yet of how the portfolio is being constructed, and into an unusually candid comparison of what each customer type is worth. The answer put a cost of capital on each end of the credit spectrum and then argued that the cheaper money is not automatically the better deal.

Q: "So I wanted to ask a bit on the conversations we're having for 2027 and 2028. It makes sense that we're maybe not exploring some like the really short-duration stuff now. But as you think about what you're hearing from customers in terms of the window from 3- to 5-year contracts, kind of where are you seeing most customers heading from within that time range? And what's the ideal if there's any time length in your guys' perspective?"
— Benjamin Sommers, BTIG

A: "The investment-grade anchor gets a 6% money; non-investment grade gets 9%. So it sounds like the investment grade wins on that until you start looking at the pricing, until you start looking at the prepayments funding around 50% of the GPU CapEx."
— Daniel Roberts, Co-Founder and Co-CEO

Assessment: This is the whole business model in three sentences, and it reframes the concentration debate. A 300bp funding penalty is cheap if the offsetting contract carries materially higher pricing and half the hardware cost arrives upfront. It also confirms, from management's own mouth, that the anchor contract is the low-priced one, which is the correct way to read the reported margins of the next two years.

Whether the financing market can keep pace with the asset class

The most important question of the call, asked last, was whether an industry financing at this scale can continue to find capital. It drew the longest answer of the day and the most revealing analogy: management argued that GPU financing is a market forming in real time, on the same template as property finance but with far shorter payback.

Q: "Just some high-level thoughts, maybe Anthony or Dan, on the financing environment, sustainability, of the industry to continue to finance this broader build-out at this pace, if there's anything that you're worried about there, how IREN may have some advantages given its different pieces to the business?"
— Joseph Vafi, Canaccord Genuity

A: "But let's just look at what happened 12 months ago, GPU financing barely existed as an asset class. And then in the last 3 months, we've raised $6.5 billion of it at both ends of the credit spectrum. So that's not us getting lucky with financing, that's a market forming."
— Daniel Roberts, Co-Founder and Co-CEO

Assessment: The empirical part of the answer is the strong part, and it is verifiable: two packages closed three months apart, one investment-grade and one not, at 6% and 9% headline rates. The analogy is weaker than presented. Property finance developed against assets with multi-decade lives and deep secondary markets, and GPUs have neither. The two-year payback management cites is the honest rebuttal to that objection, and it only holds if utilisation and pricing hold for those two years.

Data centre financing, the layer that has not started

A follow-up probed the one financing avenue IREN has talked about for two quarters without transacting. The answer was candid about the state of play and about why the timing has been deliberate, and it identified the real structural problem: data-centre capital is spent one to two years before the revenue arrives, which is the hardest part of the cycle to finance.

Q: "If you could just give any color on preliminary conversations around potential data center financing. I know you guys talked about potentially pursuing that down the road. So just wanted to ask around any preliminary conversations you've had there."
— Benjamin Sommers, BTIG

A: "Yes, lots of preliminary conversations, and we'll let you know when we close one."
— Daniel Roberts, Co-Founder and Co-CEO

Assessment: Every data centre in the portfolio is unencumbered, including Horizons 1 through 4, and that is a deliberate reserve of unused borrowing capacity against roughly $6.8bn of net property and equipment. It is also the single largest untapped funding source in the plan, and it has not yet produced a signed facility. Management's stated preference is to refinance each Horizon once commissioned and stabilised, which means the first of these transactions is a near-term catalyst worth watching.

The ERCOT audit and the Texas interconnection queue

The regulatory question that has been hanging over every Texas developer since early August was raised directly. Management's answer was that the directive's stated concerns match the way IREN has built from the start, and that two already-energised Texas sites put it on the right side of a tightening queue.

Q: "Just wanted to ask one about Texas. I know it's not a fun topic, but I was just curious if you could touch on what some of these dynamics have enabled from a commercial perspective, just given that you already have 2 large-scale energized sites there."
— Nick Giles, B. Riley Securities

A: "So I think everything that we've done in setting up our sites and our portfolio is in line with what Governor Abbott came out with in his directive. So I think in that sense, we actually welcome the additional transparency within the market."
— Kent Draper, Chief Commercial Officer

Assessment: Correct on the facts and self-serving in the framing, which does not make it wrong. The August 3 directive requires a verification and audit of every data centre in the ERCOT interconnection process before further approvals, and the associated Batch Zero classification process is still unresolved. A tightening queue is unambiguously good for an incumbent with energised load and bad for a developer with a paper position, and IREN is the former in Texas. It is worth noting that Sweetwater and Childress Horizons 5 and 6 are still in build, so the company is not entirely on the incumbent side of that line.

What They're NOT Saying

1. Any bridge from the capital plan to the funding plan. A $25bn to $30bn capex guide was paired with $14bn of identified resource and an $8bn target, and no reconciliation of the remainder beyond "data center financing, operating cash flows and corporate sources." The annual report's $13.6bn of commitments payable within twelve months was not mentioned at all.
2. The remaining performance obligation disclosure. $16.6bn of contracted revenue, with a recognition schedule showing $0.9bn of ASC 606 revenue landing in fiscal 2027, is the most decision-useful figure the company publishes. It appears in Note 4 of the annual report and nowhere in the release or the call, which spent their time on an operating metric the company itself says is not derived from revenue.
3. That the Microsoft and other capacity contracts are accounted for as leases. A determination that changes when and how the largest contract in the company's history hits the income statement was disclosed only in the revenue note. Neither the release nor the call discussed what that treatment does to the revenue line.
4. The 18.2m-share grant to the two Co-CEOs. Approved June 30, granted July 1, disclosed by standalone 8-K and a subsequent-events note. Equal to roughly 4.8% of shares outstanding. Absent from the annual results release and the call.
5. A GPU count, for the second consecutive quarter. The 150,000-GPU target for calendar 2026 was dropped without comment. The annual report names the fleet by model, from H100 through GB300 and VR200, and gives no unit numbers anywhere.
6. Quarterly mining operating statistics. The annual report restores hashrate, Bitcoin mined and installed capacity at the year level, which is an improvement on last quarter. The quarterly releases still carry none of it, for a segment that produced 48.6% of June-quarter revenue.
7. The size of the September-quarter revenue trough. Management guided cash SG&A up $40m to $50m and said the December capacity lands too late to show up in December revenue. It did not connect those two statements to what the September and December quarters therefore look like on Adjusted EBITDA.
8. Segment profitability after the capital cost. Cost of revenue is disclosed by segment. Depreciation, which for a GPU cloud is the dominant economic cost, is not allocated. The 87.0% AI Cloud gross margin therefore says nothing about whether the deployments earn their cost of capital.
9. The $100m tariff contingency. US Customs has asserted that mining hardware imported between April 2024 and February 2025 originated in China and assessed a 25% duty of roughly $100m. It is contested and unaccrued, and it was not mentioned.
10. Any framing of what a delay costs. The risk factors disclose that customer contracts carry "delay credits" for late delivery of contracted capacity, alongside service-level credits for uptime. Given that the entire equity story is a delivery schedule, the absence of any sizing of that exposure is a real gap.

Market Reaction

• Pre-print setup: IREN closed at $40.53 on August 27, up 7.3% year to date against 12.9% for the S&P 500, up 19.5% over the trailing 30 days and up 81.3% over the trailing twelve months. The 52-week closing range entering the print was $22.36 to $76.41, so the shares came in 47% below their high and had already given back most of a spring melt-up.
• After-hours: The release crossed at approximately 4:43 p.m. ET and shares were initially flat. The call began at 5:00 p.m. ET.
• Reaction session: Opened at $37.65 on August 28, a 7.1% gap down, traded a $34.81 to $38.05 range, and closed at $35.45, down 12.5% or $5.08.
• Volume: 89.9 million shares against a 30-day average of 46.9 million, 1.9 times normal.
• Benchmark: The S&P 500 fell 0.2% on the same session, so effectively the entire move was company-specific.
• Follow-through: Shares recovered 4.7% to $37.12 on August 31, retracing roughly a third of the decline.

The sequence is the analysis. The press release produced no reaction at all: revenue was inside the vendor consensus band, the loss was dominated by an impairment everybody knew was coming, and the ARR headline was a modest upgrade on a target already raised in July. The stock was flat at 4:43 p.m. and down 7.1% by the next open. Everything that moved it was said out loud between those two points.

Two disclosures did the damage, and neither appears in the 8-K. The first is the FY27 capital expenditure guide of $25bn to $30bn, a figure roughly twice the company's market capitalisation delivered without a supporting funding bridge. The second is the timing correction: the December-quarter capacity lands late enough that the reported revenue effect arrives "predominantly in the March quarter," which pushes the inflection a quarter beyond where the May framing put it. A market that had already been asked to wait through three quarters of declining revenue was asked to wait through two more, and handed a bill on the way out.

The August 31 recovery is the tell on what the market actually objected to. Shares rose 4.7% after the co-CEO said publicly that the capex figure "isn't an equity number," with prepayments covering roughly half of GPU capital expenditure and lenders funding most of the balance. That is the same content the call contained, restated as a funding structure rather than a spending total. The objection was never to the ambition. It was to a number of that size arriving unaccompanied.

Street Perspective

Debate: Is the capex guide a red flag or the price of a sold-out year?

Bull view: A company with a sold-out 2026, rising contract prices and customers wiring half the hardware cost upfront is supposed to have a large capital budget. The alternative to $25bn to $30bn is refusing demand. Roughly $19bn was raised in the last twelve months with only about $3bn from equity, and the two GPU financings closed three months apart at 6% and 9% prove the funding channel works without an investment-grade counterparty on every deal.

Bear view: The guide is 5.8x what the company actually spent last year and about twice its market capitalisation, and $13.6bn of it is already contractually committed and payable inside twelve months against $5.9bn of unrestricted cash. The plan leans on $8bn of financing not yet raised, a data-centre debt market the company has never transacted in, and "corporate sources." Adjusted EBITDA is $19.2m a quarter.

Our take: The bull case is stronger on the evidence and the bear case is stronger on the tail. Every element of the funding plan except data-centre debt has now been executed at least once at scale, which is more than could be said in May, and the prepayment mechanism has moved from a Microsoft-specific concession to a standard term at 45% to 55% of GPU capex. What the bears have right is that the plan has no slack. A single missed acceptance date breaks the sequencing that makes the flywheel work, because the debt is secured against contracted cash flows that only start on acceptance.

Debate: Does the contracted book convert this time?

Bull view: $16.6bn is contracted and audited, not a pipeline. Horizon 1 was delivered and accepted on the schedule named in May and earned NVIDIA Exemplar Cloud status on GB300 NVL72. The capacity for the December target is already built or in commissioning, and $1.6bn of customer cash for it is already on the balance sheet as deferred lease revenue. Conversion is now a construction schedule, not a sales problem.

Bear view: Reported revenue has fallen for three consecutive quarters, ARR exiting the June quarter was $0.5bn against a $4bn December target, and management has just moved the reported revenue inflection out by a quarter. The audited recognition schedule puts only $0.9bn of ASC 606 revenue in fiscal 2027 from contracts in force at June 30. Any fiscal 2027 revenue materially above that locked figure requires almost everything else to commence on time.

Our take: The bulls have taken the better of this argument since May, and the reason is Horizon 1. The single largest scheduling risk in the story resolved on time, which is the evidence that was missing when we initiated. The residual risk has changed character: it is no longer whether the capacity gets built and sold, it is whether the acceptance dates cluster in December or slide into March, and that difference moves roughly a quarter of a fiscal year's revenue. We would model the low end of the consensus range.

Debate: Is IREN a compute landlord or a technology business?

Bull view: The vertical stack is the differentiator. IREN owns the land, the grid connections, the substations, the buildings, the GPUs and now the orchestration software, so each layer raises the value of the one beneath it. That is why it can retrofit an air-cooled mining hall into revenue in months, why it earned a certification at the software layer, and why customers renew and expand rather than shop on price per GPU-hour.

Bear view: Strip the language away and this is an infrastructure landlord that has just told its auditors as much: the capacity contracts are accounted for as operating leases with IREN as lessor. The economics are rent on depreciating hardware, the anchor tenant is contracted at roughly half current market pricing, and the whole thing is financed against those leases. Landlords do not trade at technology multiples.

Our take: The lease classification is the honest description and the bears should be careful what they wish for with it. Lease revenue is contractual, recognised on a schedule, and largely insulated from utilisation. Applied to $16.6bn of contracted value with counterparties of this quality, that is a more defensible asset than a consumption-based cloud business, not a less defensible one. The multiple that follows should be an infrastructure multiple, and at 0.9 times the contracted book the equity is already being priced at one.

Model Update Needed

Valuation impact: At the August 31 close of $37.12 and 394.1m shares, IREN carries a market capitalisation of roughly $14.6bn and, after $7.71bn of debt principal and $244m of finance leases against $7.62bn of cash and restricted cash, an enterprise value of roughly $15.0bn. That is 0.90 times the $16.6bn of audited contracted revenue, 3.7 times the $4bn of contracted 2026 ARR, and 2.4 times fiscal 2028 consensus revenue. The comparable figure at our May initiation was 6.6 times contracted ARR. The contracted book has grown 29% since then and the multiple on it has fallen 44%, which is the arithmetic behind the rating change. We continue to hold no formal price target and will set one once the December acceptances either land or slip.

Thesis Scorecard Post-Earnings

Overall: Thesis strengthened. Four of the eight commitments we set as the May checklist were delivered outright, three were partially delivered, and one, the funding requirement, got materially harder. The two pillars that were unresolved at initiation, breadth of the customer book and the ability to deliver a liquid-cooled hyperscaler deployment on schedule, both resolved in the company's favour inside a single quarter. What is left is a financing problem of known size against a contracted book of known value, which is a better problem than the one we started with.

Action: Buy. The de-rating has run well ahead of the deterioration. The equity is capitalised at 0.9 times a contracted book that grew 29% while the multiple on it halved, and the specific triggers we published in May for moving to Outperform, a large non-strategic contract book and a materially lower entry, have both been met without either downgrade trigger firing. The position is sized for a funding risk that is real: we would cut it on a slipped Horizon 2 to 4 acceptance, on an equity raise struck below the current price, or on any sign that the targeted $8bn of GPU financing and prepayments is not clearing.