Thesis. IREN owns 5 GW of powered land bought at bitcoin-miner prices, has customers prepaying 45–55% of its GPU capex, and is pricing new three-year contracts above $20m per megawatt with $4bn of ARR contracted for 2026. The best-funded operator model in the category — still carrying crypto beta and a circular relationship with NVIDIA.
Reviewed 11 September 2026, after FY2026 results. Next review after the next earnings on 12 November 2026.
What it does
IREN was a bitcoin miner. It is now an AI GPU cloud, built on land it already owned and had already connected to the grid.
That history is the whole advantage. Getting a large site with a grid connection takes years, and IREN bought its at miner-era prices. Sweetwater, Texas is a 2 GW campus on 2,200 acres, with Sweetwater 1 energised to ERCOT in May 2026. The GPU fleet is around 150,000 chips. Founder co-CEOs Dan and Will Roberts.
"Powered land" means a site with an actual grid interconnection agreement, not just acreage. Power is the binding constraint on AI datacentre build-out, and interconnection queues run years long — which is why an ex-miner has an asset the hyperscalers can't quickly replicate.
Bull case
- FY2026 results, 27 August: $4bn of contracted ARR for 2026 capacity, with $1bn of ARR operating today. 2026 capacity is largely sold out and 2027 is in late-stage discussions
- Pricing is rising, not compressing. Recent three-year contracts priced above $20m of revenue per megawatt, with active discussions at around $25m
- Customers are funding the hardware. Prepayments now cover 45–55% of GPU capex — the smartest structure in the category, and it materially reduces the external financing need
- Two anchor contracts. Microsoft at $9.7B over five years with 20% prepayment and around $1.9B of ARR at full ramp, and NVIDIA at $3.4B over five years for Blackwell capacity at Childress
- Well funded. $6.4B of new GPU financing secured, on top of the prepayment structure
- The asset is scarce and proven. 5 GW of secured power, 480 MW of AI cloud capacity in 2026 targeting 1.2 GW in 2027, with energisation already demonstrated rather than promised
- A new frontier AI lab contract announced alongside the results
Bear case
- The revenue is partly circular. NVIDIA is both IREN's supplier and one of its largest customers, with a $3.4B contract and up to $2.1B of investment. NVIDIA benefits from seeding neoclouds because they diversify its buyer base away from hyperscalers building their own silicon. The revenue isn't fake — but the question of what share of ARR is genuinely arm's-length is unanswered
- The customer base is small and not investment-grade. Around ten AI developer customers, several of them venture-funded startups. A single delay or renegotiation moves the numbers
- Microsoft concentration on top of that
- The GPU depreciation treadmill. The fleet has to be re-bought, forever, and refinanced each time
- Execution at brutal scale with no long enterprise track record
- Competition compresses pricing eventually — CoreWeave, Nebius and Crusoe are building the same thing, and GPU-hour price compression is the consensus expectation for 2027–28
- Bitcoin still drives part of the chart independently of anything AI-related
Major customers
- Microsoft — $9.7bn announced November 2025, phased deployments at Childress through 2026, five-year average term, 20% customer prepayment, around $1.9bn of expected ARR
- NVIDIA — a five-year $3.4bn AI cloud contract from May 2026, managed cloud services on Blackwell systems across roughly 60 MW at Childress
- A frontier AI lab — new contract announced with the FY2026 results, unnamed
- Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI, Hume AI, plus an unnamed leading AI developer
- Bitcoin self-mining — not a customer relationship, but a second revenue stream that still moves the results
Weighted average contract term is around four years, with year-end 2026 ARR roughly 85% under contract.
Read-through. NVDA is the single most important name on this page, and it appears on both sides of the income statement — supplier and customer. NVIDIA's Q2 showed its AI cloud, industrial and enterprise segment at $40B, up 138% and growing faster than hyperscale, explicitly "as NeoCloud partners bring capacity online." IREN is inside that number.
That is the cross-read and the warning at once: the segment's growth partly reflects NVIDIA seeding its own customers. The honest test is new contract flow from customers with no NVIDIA relationship.
What would change the view
- New contract flow from customers with no NVIDIA relationship — the real test of whether demand is arm's-length
- Microsoft commissioning milestones through Horizon 1–4
- ARR against the $4bn contracted target, and 2027 conversion
- Revenue per megawatt — currently rising through $20m toward $25m, and the first sign of compression matters more than the level
- GPU-hour pricing across the category
- The terms of the next financing round
Update log
- 11 Sep 2026 — Reviewed. Verdict held at Constructive. NVIDIA's ACIE segment growth noted as confirmation of neocloud capacity coming online, and as the circularity risk in the same breath.
- 27 Aug 2026 — FY2026 results: $4bn contracted ARR for 2026, $1bn operating, $6.4B of new GPU financing, prepayments raised to 45–55% of GPU capex, contract pricing above $20m per MW.
- 25 Jul 2026 — Full review.
Research and education only — not investment advice.
Made with Bullet