IREN swapped $30 million of bitcoin revenue for $30 million of AI revenue — and its total came out $28,000 lower than a year ago

IREN mines bitcoin and is turning itself into an AI cloud. The transition is real, it is enormous, and the most striking thing in the filing is what it has produced so far: nothing, on the top line.

Revenue in the quarter to 31 March was $144,795,000. In the same quarter a year earlier it was $144,823,000. Down twenty-eight thousand dollars — 0.02% — because AI Cloud revenue grew by about $30.1m and bitcoin mining revenue fell by about $30.1m. The company's own release never makes that comparison; it anchors to the sequential quarter, which fell.

The headline loss looks worse than the business. Of the $(247.8)m net loss, $318.9m is non-cash — impairments, depreciation, share comp and a mark on capped calls — which is more than the entire loss. Operating cash flow was positive $75.3m.

And the contracts everyone quotes had not started. At the balance-sheet date the $9.7bn Microsoft contract contributed exactly zero to remaining performance obligations, because IREN books them only as tranches are delivered and accepted, and there had been nil. The first one was accepted on 13 August.

The number the release does not print

IREN's Q3 FY26 release leads with a sequential comparison: revenue "decreased to $144.8m (vs. Q2 FY26 $184.7m)". That is true and it is the less interesting number.

The year-over-year figure is not in the release at all. Q3 FY26 revenue was $144,795,000. Q3 FY25 revenue was $144,823,000. The difference is twenty-eight thousand dollars, or 0.02%.

Underneath that stillness, everything moved. AI Cloud services went from $3,581k to $33,635k — up 839%, and up 94% on the prior quarter. Bitcoin mining fell by almost exactly the same amount. A company that has raised billions, signed a $9.7bn contract and rebuilt itself around GPUs has, after a full year, precisely the revenue it started with.

That is not an accusation. It is what a substitution looks like when the two curves cross. But it is the fact a reader most needs and the one the release does not surface.

What the loss actually measures

The $(247.8)m net loss reads like a company in trouble. It is mostly demolition.

That is $71m more than the entire net loss, and net cash from operating activities was positive $75.3m — with working-capital swings roughly cancelling, so the figure is close to clean. Segment cash margins ex-depreciation are 68.2% for mining and 86.3% for AI Cloud. Neither is a true gross margin, because the $121.2m of depreciation sits as a reconciling item in the segment note rather than being pushed down — a caveat worth carrying, particularly for the AI Cloud figure, which is the margin the bull case leans on and which has never been charged for the GPUs producing it.

The impairment is the point: it is the deliberate write-down of mining hardware being decommissioned, and Note 22 warns of approximately $520 million more to come. Mining hardware still carries at $1,041.8m gross against GPU hardware at $811.4m — the asset base being destroyed is still larger than the one replacing it.

The contract, read rather than characterised

This is where the report changed its own mind.

The natural bear reading of a $9.7bn headline is that it is a ceiling: Microsoft pays only for GPUs it accepts, so the number describes an opportunity rather than an obligation. That reading is wrong, and the document that settles it is filed — the Statement of Work went in as Exhibit 10.3 to the December quarter's 10-Q.

Section 3.2.a.v: "The Service Fees are non-refundable, non-cancellable and constitute a firm minimum commitment by Microsoft to Partner regardless of whether Microsoft utilises any or all of the Services, except as otherwise provided for in this Agreement."

Exhibit F disapplies Sections 7–17 of Microsoft's standard purchase-order terms, which removes the Section 14 termination-for-convenience right that would otherwise let Microsoft walk at will. What remains are two narrow exits: a per-tranche termination if IREN misses the Delivery Delay Window, and a 60-day-cure material-breach right that expressly carves out late delivery as not a material breach. On any other termination, Section 6.c requires Microsoft to "pay an amount equal to the TCV less any Service Fees already paid."

The ceiling half is real — fees "will not exceed $9,666,845,337.60", total contract value shrinks by the tranche value of anything terminated, and billing multiplies GPU quantity "less any GPUs that have not been Accepted by Microsoft." But a cap on price is not an absence of commitment. The risk in this contract is IREN's execution, not Microsoft's demand. There is also a $1,932,839,884.80 upfront payment schedule — exactly 20% of tranche value — credited back pro rata after the 24th month.

And yet none of it was in the accounts

Both things are true at once, which is the whole report.

Note 5: "The Group includes amounts in unsatisfied RPO only for tranches that have been delivered and accepted, of which there have been nil as of the reporting date."

So at 31 March, a firmly committed $9.7bn contract contributed exactly zero to remaining performance obligations. Total unsatisfied RPO was $710,272k, of which $307,950k falls in the twelve months to March 2027. Against the "$3.1bn ARR under contract" headline that is a 10.1× gap — and annualised current revenue of $579m is nearly double the entire GAAP twelve-month backlog.

The first tranche was accepted on 13 August 2026: Horizon 1, 50MW of IT load, the first of four buildings totalling approximately 200MW. Twelve months earlier IREN had, on its own co-CEO's account, "approximately 3MW of self-built AI Cloud capacity." The FY2025 10-K had targeted Horizon 1 for energization by the end of calendar 2025; acceptance came about eighteen months after the project was announced.

NVIDIA on three sides of the table

NVIDIA is IREN's GPU supplier, a $3.4bn customer under a five-year cloud services agreement, and a prospective shareholder: investment rights over 30,000,000 ordinary shares at $70.00, up to roughly $2.1bn gross, running to 7 May 2031.

The mechanism is the part worth knowing. Those rights "vest based on certain volumes of deliveries of up to 600,000 NVIDIA GPUs" — so NVIDIA's right to buy IREN equity is earned as IREN buys NVIDIA hardware. Roughly one share per twenty GPUs.

Two honest qualifications. The $2.1bn is a ceiling, payable only on full exercise and subject to regulatory limits — and at $41.88 the stock is 40% below the $70.00 strike, so none of it is money today. And the Q3 10-Q contains no related-party disclosure of NVIDIA anywhere — the phrase does not appear in the filing, its exhibits or its XBRL. That is defensible on timing, since both agreements were signed on 7 May, after the balance-sheet date. It is worth watching whether the FY2026 10-K takes a different view.

The balance sheet

Commitments went from $368,805k at 30 June 2025 to $11,902,471k at 31 March 2026 — 32× — with $11,899,054k payable within twelve months against $2,213,274k of cash at that date.

The funding response was fast and large. $3.0bn of 1.00% convertible notes due 2033 on 14 May, at a ~$73.07 conversion price, with $201.3m spent on capped calls. Then on 29 May a ~$3.6bn secured project financing at the Hardware 3 SPV: a $1.5bn delayed-draw term loan at SOFR+2.25% and $2.1bn of 5.96% senior notes due 2031, secured against the Hardware 3 assets including the GPUs and the Microsoft cash flows, with a 1.05:1.00 debt-service-coverage covenant. Cash reached ~$7.6bn at 30 June, of which $1.7bn is restricted against that facility.

The dilution is structural. Shares outstanding rose 32.3% in nine months; the antidilutive overhang is 192.4m shares, 86.2m of it from converts; a $6.0bn shelf was filed in March. Worth correcting a tempting simplification: it is not true that every strike is out of the money. The 2029 and 2030 converts have conversion prices of $13.64 and $16.81 on $445.7m of principal — deeply in the money at $41.88. It is the newest layers ($51.40, $70.00, $73.07, $85.63) that are not.

What the price assumes

At the 21 August close of $41.88, enterprise value is about $16.72bn. Set against headline total contract value of roughly $15.9bn — the $9.7bn Microsoft and $3.4bn NVIDIA contracts plus $2.8bn announced on 20 July — that is about 1.05× EV to contract value.

That is more expensive than CoreWeave, at $94.48bn of EV against $103.7bn of RPO (0.91×, or ~0.73× including the >$25bn of early-Q3 commitments it discloses but excludes), and cheaper than Nebius at $62.22bn against $37.5bn of RPO (1.66×).

But the comparison is not like-for-like, and that matters more than the ranking: IREN is measured on total contract value, CoreWeave on RPO plus committed backlog, Nebius on strict ASC 606 RPO. On its own GAAP RPO of $710.3m, IREN trades at 23.5× — and at roughly 124× annualised AI Cloud revenue. The screen EV is also stale in both directions: it predates the May convertible and the Hardware 3 draw, and equally predates the cash build to $7.6bn and the 11,981,668 shares issued for Mirantis on 3 August.

Risk — each isolated, do not blur

1. Acceptance (dominant). Revenue exists only as tranches are delivered and accepted. The Limited Parent Guarantee explicitly covers "any shortfall in Hardware 3's payment obligations attributable to a tranche of GPU services that Microsoft does not accept or terminates" — the company has guaranteed the consequence of its own delivery risk.
2. The transition gap. Bitcoin revenue is being switched off on a schedule; GPU revenue arrives on an acceptance schedule. The two are not synchronised, and roughly $520m of further impairment sits between them.
3. Near-term commitments. $11.9bn falling due inside twelve months is a very large number against a business with $579m of annualised revenue, even with $7.6bn of cash.
4. Dilution. A 32.3% share-count increase in nine months, 192.4m of overhang, a $6.0bn shelf, and 30m NVIDIA rights.
5. Concentration. Two customers — Microsoft and NVIDIA — account for essentially the entire contracted story.
6. Schedule. Horizon 1 slipped from a calendar-2025 energization target to an August 2026 acceptance. Three more Horizons are targeted for 2026.

Horizon and sizing (kept separate)

Horizon. The near term is mechanical and unusually legible: how many of the remaining three Horizon tranches are accepted before the FY2026 10-K, and what happens to GAAP RPO when they are. Watch the RPO line, not the ARR headline — it is the one that moves only when revenue becomes real.

Sizing considerations (not a recommendation). IREN is the only company in this section that owns its power and its land rather than leasing capacity, which is a genuinely different risk profile from the other neoclouds — slower to build, harder to displace. It is also the only one deliberately destroying a profitable business to fund the new one. The multiple embeds delivery on a schedule that has already slipped once.