IREN Signed $2.8B in AI Contracts. The CRE Read on a New Data Center Capital Stack

On July 20, 2026, IREN Limited (NASDAQ: IREN) said it signed new multi-year cloud services contracts with leading AI developers representing $2.8 billion in total contract value, and raised its year-end AI Cloud annualized run-rate revenue target from $3.7 billion to more than $4 billion. About 85 percent of that revised target is now under contract, the company says. Its named customer base spans Microsoft, NVIDIA, Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI, Hume AI, and one further unnamed leading AI developer, and the portfolio carries a weighted average contract term of approximately four years. Those figures are IREN’s, filed with the Securities and Exchange Commission as an exhibit to a Form 8-K, not PSV findings.

Two operating details sit under the headline. IREN says recent contracts include customer prepayments representing about 45 percent of the associated GPU capital expenditure, which reduces the company’s net funding requirement for those deployments, and that it held roughly $7.6 billion in cash as of June 30, 2026, including $1.7 billion of restricted cash tied to GPU financing for the Microsoft contract at its Horizon data centers. On capacity, Co-Founder and Co-CEO Daniel Roberts said the platform expanded from about 3 megawatts of self-built AI Cloud capacity a year ago to 480 megawatts being delivered this year, with 1.2 gigawatts targeted for 2027. The Horizon 1 through 4 facilities named in the filing sit at IREN’s 750 megawatt Childress, Texas campus, per the company’s site.

This is a data center real estate story wearing a cloud-revenue headline. IREN owns and builds the physical plant: the land, the on-site substations, the power interconnect, and the liquid-cooled halls that house the GPUs. What changed on July 20 is not that AI demand exists, it is how the buildout gets paid for. When a developer can point to four-year contracts covering 85 percent of a raised revenue target, and to customer prepayments funding close to half of the associated equipment cost before a hall fills, the asset starts to underwrite like a credit-tenant, take-or-pay development rather than a speculative one. Contracted cash flow, not a demand thesis, is doing the work.

For an operator, that reframes the counterparty and the capital question at once. The scarce inputs in this asset class are powered land and speed to energization, not construction know-how, and the players winning contracts are the ones who already control an interconnect and can deliver megawatts on a customer’s schedule. It also concentrates risk in a way worth naming: contracted revenue is only as good as the counterparties behind it, prepayments shift some financing risk onto the customer but tie the developer to that customer’s roadmap, and equipment-backed debt against fast-depreciating GPUs is a different animal than a mortgage on a building. Contracted does not mean riskless.

Treat a disclosure like this as a trigger to build a capital-stack and contract-coverage file from primary records: the Form 8-K and its exhibit, the company’s prior filings on the Microsoft agreement and its financings, the campus’s local permits and utility and interconnection records as they appear, and the grid operator’s queue for the relevant load. An AI assistant reads each record and returns a cited memo: how much revenue is actually contracted versus targeted, the weighted term and counterparty mix, what share of capex is covered by prepayments and equipment financing versus corporate capital, which megawatts are energized versus announced, and what changed since the last check, every figure traced to the document it came from.

The reviewer is the capital markets or acquisitions lead working with an analyst, and the approval gate is explicit: no land bid near a campus, no supply position, and no underwriting assumption moves on a model summary alone. The same review produces the questions a human sends: to the utility on interconnection and load for the specific site, to the county on permit and entitlement status, and to counsel on how prepayment and take-or-pay terms actually allocate risk if a customer’s roadmap slips. The tool assembles the record. It does not decide whether the contracted cash flow is worth the exposure.

The judgment calls stay human. Whether contracted revenue at these terms de-risks a specific site, how to weigh a developer’s concentration in a handful of AI-developer counterparties, and whether equipment-backed leverage against depreciating GPUs belongs in a real estate underwrite are principal decisions, not summary outputs. So is the concentration question that runs through every data center brief on this desk: a platform that rises on a few customers’ capital plans also carries those customers’ cycles.

The honest cautions. The $2.8 billion, the above-$4 billion target, the 45 percent prepayment figure, and the capacity numbers are IREN’s own, filed with the SEC but forward-looking, and a run-rate target is not realized revenue. One unnamed customer means part of the counterparty credit is not public. PSV has not audited IREN’s contracts or financing and states no view on the stock. The open questions worth tracking will land in filings and dockets, not the press cycle: how much of the 480 megawatts is energized and contracted on schedule, what the Childress interconnection and the GPU financing terms actually require, and whether the prepayment structure holds if AI-developer demand cools. Verify at the substation, not the run rate.