Summary
- IREN announced $2.8bn of new multi-year AI cloud contracts and raised its year-end 2026 AI Cloud annualised run-rate revenue target from $3.7bn to more than $4bn.
- The company says roughly 85% of that target is under contract; its portfolio has a weighted average stated term of about four years.
- Recent contracts include prepayments equal to approximately 45% of the estimated GPU capital expenditure associated with those deployments, reducing but not eliminating IREN’s funding and execution burden.
- ARR is a non-GAAP operating metric built from commissioned GPU hourly prices multiplied by 8,760 hours, plus annualised storage and ancillary revenue. Delivery, testing and customer acceptance still stand between a contract and reported revenue.
Eight thousand seven hundred and sixty hours sit inside IREN’s new revenue headline. They are the number of hours in a year, and the multiplier the company uses when it converts the hourly price of commissioned GPUs into annualised run-rate revenue.
That denominator is more revealing than the familiar “$4bn-plus” headline. IREN has found buyers: $2.8bn of new multi-year agreements lift the contracted share of its raised year-end target to about 85%. The open question is now less whether demand exists than whether machines arrive, pass testing, win customer acceptance and stay utilised at the prices assumed by the run-rate calculation.
This is commercial validation with an engineering dependency attached.
A contract can reserve an hour before the GPU serves it
IREN defines AI Cloud ARR as the hourly price of GPUs commissioned by 31 December 2026, multiplied by 8,760, with annualised storage and ancillary revenue added. It explicitly says the figure is an operating metric, not a GAAP measure and not a substitute for revenue recognised under accounting rules.
The distinction is not technical small print. A year-end snapshot annualises capacity as if the relevant configuration and price applied for a full year. The income statement, by contrast, records service as it is delivered under the applicable accounting treatment. A rack accepted late in December may contribute to the run-rate measure while contributing only a short period of service revenue during 2026.
IREN’s raised target also rests on physical assumptions. It expects 480MW gross of AI Cloud capacity to be delivered during 2026, compared with approximately 3MW of self-built capacity a year earlier, and targets 1.2GW for 2027. Revenue is expected to ramp only after the data-centre delivery sequence is followed by GPU commissioning, testing and customer acceptance.
“Approximately 85% under contract” should therefore not be rewritten as 85% operational, 85% billed or 85% recognised. It measures commercial coverage of a target whose equipment and service base is still being assembled.
Prepayments change the funding equation, not the acceptance test
The strongest term in the announcement may be the customer prepayment rather than the contract total. IREN says agreements executed since 1 June include amounts payable before service delivery equal to roughly 45% of the estimated GPU capital expenditure associated with those deployments.
That moves part of the hardware-funding exposure toward customers. A buyer willing to provide capital before receiving service is supplying more than an expression of interest. It reduces the net amount IREN must finance and provides a harder signal that the customer wants the reserved compute.
But the percentage has a deliberately narrow denominator. It refers to estimated GPU capex associated with recent deployments whose contracts include prepayments. It is not 45% of all IREN capital spending, all 480MW planned for 2026 or the company’s 1.2GW target. Terms vary by contract, and IREN warns that future contracts may not repeat them.
Nor does money received in advance make an unfinished GPU cluster operational. IREN still owns procurement, integration, cooling, networking, software, commissioning and performance risk. A customer still needs the promised system to meet acceptance conditions. Prepayment improves the financing bridge; it does not remove the bridge.
Four years of contracts meet two years of construction
The portfolio’s weighted average stated contract term is about four years, calculated by weighting each agreement by its contribution to ARR. That provides more duration than a spot cloud booking, but it also exposes the operator to a different set of timing risks.
The first clock is contractual. IREN must perform over multi-year terms and retain the pricing and utilisation assumed in its run-rate.
The second clock is the 2026 delivery programme. The planned 480MW has to move through construction, hardware arrival and acceptance before year end if it is to support the target definition.
The third is the 2027 expansion to a targeted 1.2GW. That is a management target, not commissioned capacity or a disclosed customer commitment. Scaling from hundreds of megawatts to more than a gigawatt magnifies supply-chain, power, cooling and operational demands even when customer interest exceeds planned supply.
The new deals provide evidence on the first clock. They do not make the second or third run faster.
Liquidity is large, but not all of it is free
IREN reported preliminary cash and cash equivalents of approximately $7.6bn at 30 June. The balance is meaningful support for an infrastructure programme, but $1.7bn of it is restricted in connection with GPU financing for the Microsoft contract at Horizon 1–4.
That restriction matters for the same reason the prepayment denominator matters. Gross liquidity, customer-supported GPU purchases and freely deployable corporate cash answer different questions. Adding them together rhetorically would overstate the capital available for every new deployment.
The company’s customer list now spans Microsoft, NVIDIA, Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI, Hume AI and a newly referenced leading developer, across bare-metal and managed services. That list demonstrates a broader commercial surface, but IREN does not say that every named company signed one of the $2.8bn of new agreements. The counterparties to much of the new value remain unnamed.
The independent market response was immediate, but share-price relief is not an operating milestone. The more durable checkpoints will be disclosed commissioning progress, customer acceptances, actual utilisation, recognised revenue and the relationship between run-rate expectations and cash conversion.
IREN has moved the debate. A company that once had to prove there was demand for its AI pivot can now point to multi-year contracts and customers willing to prepay part of associated GPU costs. Its next proof is more physical: turn a contracted hour into a tested GPU hour, then keep doing so 8,760 times.