IREN: Quietly building the next AI cloud backbone

The mood around Iren changed drastically within a week. Two weeks ago, most still filed it under Bitcoin mining, cheap power, efficient rigs, a likely beneficiary of rising grid demand. Then a $9.7 billion Microsoft contract arrived alongside a clear plan to stand up liquid‑cooled GPU capacity in Texas and Canada. Same company, new footing in a different light. The headline is big; the meaning of this deal is even bigger. Iren is moving from selling mined coins to supplying reserved AI capacity and being paid for reliability.

What actually changed

Three key developments arrived at once. First, the Microsoft contract, with about 20% paid upfront, confirmed that demand is real and backed by cash, not just intent. Second, the latest quarter showed that the balance sheet can support the initial build through cash reserves, supplier financing, and a zero-coupon convertible bond due 2031, meaning there is no cash interest burden during construction. Third, NVIDIA named Iren a Preferred Partner, and in a market where access to GPUs matters more than price, that partnership carries weight. Supply, power, and capital are finally working in sync.

What Iren does, if you have no idea what Iren does.

Iren designs and runs data halls, liquid‑cooled and power‑dense, on very cheap, flexible electricity. It sells two things. One, self‑mined Bitcoin at the low end of the cost curve. Two, reserved AI clusters under multi‑year contracts where the customer pays for capacity. Childress, Iren’s strategic 750 Mega-watt data centre located on 576 sweet American acres in West Texas is the lead site with its own grid connection, cooling, and on‑site power design, so timelines rest with Iren, not a landlord. If you read “capacity provider,” think utility‑like service, you pay for uptime and delivery.

The engine that drives returns

Mining built the operating discipline. Iren’s cost to produce one Bitcoin sits near $41,000, while realised revenue per coin has been around $93,000. That spread earned trust in their ability to run efficiently. The same playbook now applies to the AI segment. Once a data hall is operational, the first half of its capacity mainly covers fixed costs. After usage climbs past roughly 70%, most additional revenue turns into profit because major expenses such as power, cooling, and infrastructure are already covered while clients pay for reserved capacity.

Where my view differs from the consensus

Most analysts see 2026 revenue around $1.3 billion, assuming AI makes up only a small portion of sales and mining stays near 50 EH, or exa hashes per second. I think that view underestimates how these contracts scale once capacity is installed and utilization rises. If Iren deploys around 100,000 GPUs by the end of 2026 and Microsoft, along with other large clients, expands as planned, I believe annual recurring revenue from AI could exceed $2 billion by late 2026. The test is clear. If by December 2026 GPU count is below 70,000 or AI revenue falls short of $1.2 billion, the thesis fails and I would downgrade.

Scorecard so far

Fiscal 2025 built the foundation, with revenue around $501 million and solid adjusted EBITDA. Management kept mining output near 50 EH while redirecting power and capital toward AI expansion. In the first quarter of fiscal 2026, revenue came in around $240 million with adjusted EBITDA of about $91.7 million. Liquidity stands near $1.8 billion, supported by roughly $400 million in GPU financing that helps de-risk near-term commitments. Citigroup has become the fourth-largest shareholder with about 6.4 million shares worth roughly $400 million, a clear sign that major institutions are backing the strategy. The company also issued a $1.125 billion zero-coupon convertible due 2031, which carries no cash interest and can convert into shares later, with capped calls that limit dilution up to a higher share price. In short, the balance sheet is built for growth.

The visual below tells the story better than any sentence could. Revenue has climbed from roughly $55 million a few quarters ago to over $240 million today, while EBITDA moved from loss-making territory to around $90 million at its peak. The small pullback in Q3 reflects the cost of growth rather than a slowdown, with new halls being commissioned and early AI capacity not yet fully monetised. The market seems to agree, the share price has surged more than 140% year on year, mirroring the step up in scale and credibility. What matters now is converting that build into consistent profitability as utilisation rises.

Moat direction, widening if uptime holds

Texas power is cheap and flexible. ERCOT, the Texas grid operator, rewards smart curtailment, temporarily dialing down power when the grid needs capacity. Iren designed for this and controls the on‑site elements that make delivery predictable. Add NVIDIA Preferred Partner status and a blue‑chip anchor, and the practical switching cost rises. For Microsoft to move, it would need to recreate a liquid‑cooled, supply‑constrained GPU grid with comparable reliability and price. That is difficult. Hence the prepayment. Now I would like to mention, major players like Microsoft are outlaying this cost to Iren to limit their overheads - if they suddenly need to pull back on their Capex spending - Microsoft come out with less bruises and cuts.

What can possibly go wrong?

There is execution risk if parts arrive late or cooling systems take longer to install, which would push revenue recognition further out and test investor confidence. Policy risk is another factor, since any change in ERCOT rules could shift the power cost equation. There is also a concentration risk, with one major cloud provider driving most of the initial AI revenue. None of this is a surprise, but it means investors should watch the energisation dates at Childress, monitor any ERCOT policy updates, and pay close attention to contract terms covering penalties and delivery delays.

Catalysts over the next 12 months

The first milestone is full commissioning at Childress, meaning the site is cleared to operate at full capacity. Next comes proving it can run at performance targets, followed by filling at least 70% of available racks within two quarters. That 70% mark matters because most costs are fixed, so once usage reaches that level, nearly every additional dollar of revenue turns into profit. If Microsoft expands capacity ahead of schedule, it confirms the contract structure is working as intended, where payment is secured for reserved capacity whether or not it is fully used. Securing a second major cloud provider or a large enterprise client would diversify demand and reduce dependency risk. The monthly updates on mining output and power costs remain important too, since they track the legacy cash flow that continues to support the build phase.

Valuation, quick reality check

Enterprise value sits near $16.4 billion. On 2026 adjusted EBITDA around $500 million, the multiple is about 32x, enterprise value divided by EBITDA. That is expensive if you think miner, acceptable if you think contracted capacity provider mid‑ramp, and fair for a company with milestone risk. You do not need multiple expansion to win. You need the denominator to grow as halls fill, and the multiple to ease toward the mid‑teens by 2028. If GPUs slip 20% and ARR trails plan, equity IRR falls below my hurdle.

My four conditions before I call it inevitable

One, commission on time. Two, usage over 70% within two quarters of go live. Three, publish SLA results, uptime and penalties, plus realised PUE at Childress. Four, secure hardware allocation beyond the current tranche so 2026 deliveries are protected. And five, free cash flow. I need to see it turn positive and build quarter after quarter before conviction is set. Free cash flow is the ultimate truth in this story, it tells you if operations can fund expansion without leaning on fresh equity or debt. It measures discipline, pricing power, and the durability of margins once the hype fades. The chart below shows how those swings have narrowed sharply, but consistency is what will confirm this business has matured into a self-funding engine.

How the money actually moves

Most of Iren’s sales go directly to enterprises, with almost no spend on marketing. Deals usually follow a set path: first an NDA and due diligence, then a multi-year contract with about 20% paid up front, followed by staged delivery, steady billing, and renewal options around year five. Once AI sites are fully operational, they should deliver project margins above 80% if energy costs stay within plan. The Bitcoin mining side typically earns a mid-50% cash margin when hashprice, or revenue per unit of mining compute, holds steady. Upfront payments and daily Bitcoin sales support working capital. Free cash flow will remain negative through the build phase, then turn positive as data halls reach capacity and depreciation overtakes new spending.

Competitive view

Among public miners, Iren ranks near the top for power cost and efficiency. Its advantage shows up in cost per coin and the flexibility to reduce output when power prices rise. Marathon, Riot, CleanSpark, Cipher, and Hut 8 all chase similar inputs, but few deliver with the same consistency. On the AI side, the real competition comes from operators able to deploy liquid-cooled GPU infrastructure at scale, maintain close to 99.999% uptime, and secure dependable chip supply. That list is very short.

Key metrics to track

The next wave of AI customers needs to move from onboarding to paid service quickly—ideally within 30 days and covering about 90% of planned GPU deliveries. If that slips for two straight quarters, capital spending should slow until conversion improves. Average revenue per unit should stay steady or rise, with minimal churn through each contract term. If pricing trends negative twice in a row, it’s time to reassess indexation. Mature AI customers should contribute more than 80% of revenue, and if that mix weakens, I’d check power and cooling performance. Net revenue retention, meaning growth from existing clients after churn, should hold at or above 100%.

Verdict

I’m still building conviction here, but watching the story closely. The mining side continues to fund operations, yet the real pivot depends on how well they deliver on AI. If execution matches the plan, this could reshape the company’s entire income base. The market already prices in a fair amount of optimism, so I’m waiting for proof before leaning in. Each new data hall that goes live and sustains usage above 70% reduces risk and supports valuation. But if by December 2026 GPUs remain below 70,000 or AI revenue falls short of $1.2 billion, I’ll step back. For now, I’m holding a measured position and letting results earn the next move.