At a glance
IREN signed US$2.8bn of new multi-year AI cloud contracts on 20 July 2026, spread across about ten customers rather than the two names, Microsoft and NVIDIA, that anchored its earlier book.
Customers are prepaying about 45% of the GPU cost up front, a strong sign the contracts are firm.
IREN lifted its 2026 revenue run-rate target from US$3.7bn to more than US$4bn, about 85% now under contract.
The new names are mostly venture-funded AI labs, which spreads IREN’s risk but on lighter credit than a hyperscaler.
The contracts and capacity are offshore. IREN’s Australian project, the 800MW Bundey campus in South Australia, targets first power from 2028.
Customer base widens from two anchors to about ten names
IREN said on 20 July 2026 it had signed US$2.8bn of new multi-year cloud contracts with about ten AI companies: Microsoft, NVIDIA, Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI, Hume AI and one it did not name. The figure is US$2.8bn of total contract value, not US$28bn.
IREN is the largest Australian-founded neocloud by market value, founded in Sydney in 2018 as Iris Energy and still headquartered at 55 Market Street, though it runs its computing halls in the United States, Canada and Spain. As we set out in our state of play on IREN’s Australian position, its megawatts are offshore and its earlier book rested on two names: a US$9.7bn Microsoft deal and a US$3.4bn NVIDIA contract. The real change is the customer base widening from two anchors to about ten.
Customers prepay about 45% of the GPU cost
Under those contracts, customers are prepaying about 45% of the cost of the GPUs, up front, before the chips are installed and earning anything.
Normally the operator buys the GPUs first with its own money and earns it back slowly as customers rent them. Here the customers cover almost half the hardware bill in advance, the way you pay a builder a large deposit before construction rather than after. A customer who puts 45% down before the racks are switched on has shown, in cash, that the contract is real.
IDC’s Dave McCarthy, writing about the deals, said the prepaying customers are “essentially acting as project co-investors to lock down guaranteed capacity.” Paying that much up front lowers what IREN has to raise itself. Those prepayments, plus the US$3.65bn investment-grade facility it closed in June, are also why IREN holds about US$7.6bn in cash, US$1.7bn of it restricted for the Microsoft financing.
Revenue run-rate target lifts past US$4bn, 85% contracted
IREN also raised its 2026 run-rate revenue target, or ARR, from US$3.7bn to more than US$4bn. ARR is not booked revenue. It takes what IREN expects to be earning late in the year and annualises it, so it is a projection.
Two things temper it. About 85% is under contract, leaving roughly 15%, on the order of US$600m, still to be signed. And IREN’s own release flags the risk that contracted GPU capacity may not convert to revenue on schedule. The prepayments are the hard evidence here; the run-rate target is the ambition attached to them.
New customers skew to venture-funded AI labs
Spreading the book across ten names reduces the risk of leaning on one or two. The trade-off is who the new customers are: mostly venture-funded AI labs such as Perplexity, Figure AI, Together AI and Fireworks AI, not hyperscalers with decades of cash flow.
As we noted in our piece on Meta’s compute resale plan, a hyperscaler contract is bankable for a decade, while a contract with a younger lab is only as solid as its funding. IREN is broadening beyond its two anchor names, and taking on lighter credit to do it; one new name, Fluidstack, is itself a compute broker. Bernstein had already ranked IREN below CoreWeave and Nebius on revenue per megawatt of contracted capacity, while keeping a Buy rating. The book is broader, and its credit is thinner.
Self-built capacity scales from 3MW to 480MW in a year
A year ago IREN ran about 3MW of its own AI computing. It expects 480MW this year and 1.2GW in 2027, roughly a 160-fold jump in twelve months that costs billions to build. Customers prepay about 45% of the chips; IREN funds the rest and wears the depreciation on hardware that dates fast, with NVIDIA now shipping a new platform roughly every year. IREN is scaling this fast on fast-ageing equipment while carrying the half its customers do not prepay.
IREN’s Australian plans centre on Bundey, from 2028
The deal sits offshore, where IREN runs its capacity today, across the United States, Canada and Spain. The Australian angle is indirect for now: the contracts strengthen IREN’s revenue base and cash position, the kind of balance sheet a capital-heavy local build would eventually draw on. Its Australian project, the 800MW Bundey campus in South Australia, is still at the grid-connection stage and targets first power from 2028.
The read fits the tier we mapped across Australia’s neoclouds: the market is paying for contracted demand backed by secured power, with sovereignty as the premium on top, the same test we set out on the two markets forming in AI compute. IREN’s new book, with money down and multi-year terms, sits on the safe side of that line.
What IREN reported | The figure | What it signals |
New contract value | US$2.8bn total, ~4-year average term | Fresh multi-year demand across about ten customers |
Customer prepayments | ~45% of the GPU cost, paid up front | Firm commitment and the source of much of the cash pile |
2026 revenue run-rate target | Raised to more than US$4bn, ~85% contracted | Ambition, with ~15% still to be signed |
Self-built capacity | ~3MW a year ago to 480MW now, 1.2GW in 2027 | Very fast growth, and the funding and depreciation risk under it |
Source: IREN disclosures, July 2026.
What to watch
Three things will show whether this converts. IREN’s next quarterly update, where the test is how much of the more-than-US$4bn run-rate has become billed revenue. Any move on Bundey from grid agreement to construction, and whether IREN names the customers it will serve. And whether IREN contracts any Australian capacity, which would turn its local presence from pipeline into demand.