IREN — IREN LIMITED

IREN owns 2.9 gigawatts of cheap renewable power — the exact resource that every major technology company is frantically trying to acquire — and has landed $13 billion in contracted AI infrastructure agreements with Microsoft and NVIDIA that validate the position. But the contracts are aspirational today: Q3 fiscal 2026 AI cloud revenue was $33.6 million against a $3.7 billion annual run-rate target, a gap that requires deploying 150,000 GPUs in roughly six months with no execution slippage — precisely the kind of flawless performance a 34% Q3 revenue miss suggests is not guaranteed. Interesting but requires a specific catalyst to be actionable.

---

The defining constraint of the current technology era is not compute hardware. It is power. Chip manufacturers have delivered unprecedented GPU density, but a GPU sitting in a building without reliable grid access is a very expensive piece of metal. The hyperscalers — Microsoft, Google, Amazon, Meta — have committed to spending somewhere between $660 and $690 billion on data center infrastructure in 2026 alone. A meaningful fraction of that spending cannot be deployed as quickly as they would like because new grid interconnection in the United States now takes two to five years to secure. The entities that already hold grid interconnection rights in quantity are not land developers or real estate investment trusts. They are, preposterously, Bitcoin miners.

Bitcoin mining created a generation of companies that were paid to consume electricity in remote places near cheap generation — hydroelectric dams in British Columbia, wind farms in West Texas, stranded natural gas in North Dakota. They built the substation relationships, the transmission agreements, and the physical infrastructure to take large blocks of power from grids that had nowhere else to send it. Then the economics of Bitcoin mining deteriorated: the halving events cut block rewards in half every four years, the global hashrate compounded relentlessly as competitors crowded in, and the cost to mine each coin rose as difficulty increased. What these companies were left with was not a moat in Bitcoin. They were left with a moat in power.

The AI infrastructure market is not a software business. Its economics resemble pipeline infrastructure more than cloud software: capital-intensive to build, long-contract in duration, and priced per unit of throughput. A GPU server rack in a data center does not depreciate the way a software license does; it runs continuously, drawing power and generating revenue for years. The question of who profits is therefore largely a question of who holds the cheapest, most secure power — and that question was answered, years ago, by companies whose original business plan had nothing to do with AI.

IREN Limited was founded in Sydney in 2018 by Daniel and Will Roberts, two former Macquarie infrastructure bankers who built a Bitcoin mining company around a simple insight: the lowest-cost electricity in the world tends to be stranded. Hydroelectric dams in British Columbia generate power that cannot easily be transported elsewhere. Wind farms in the Texas Panhandle produce surplus generation at off-peak hours. The brothers set about locking in offtake agreements with these generators at prices that conventional data center operators would not pay because conventional data center operators do not want to be in British Columbia. They listed the company on NASDAQ in November 2021 under the name Iris Energy at $28 per share, built out a Bitcoin mining operation running on wholly renewable power, and then watched the rest of the technology industry arrive at the same conclusion they had reached four years earlier: cheap, abundant power is the scarce resource.

In late 2024, IREN — the company rebranded from Iris Energy — announced a $9.7 billion, five-year AI cloud infrastructure agreement with Microsoft. The contract calls for IREN to deploy 200 megawatts of liquid-cooled GPU compute at its Childress, Texas campus using NVIDIA GB300 hardware. The annualized value once fully deployed is approximately $1.9 billion, at a projected 85% project-level EBITDA margin. This was followed in 2025 by a $3.4 billion, five-year strategic partnership with NVIDIA to deploy GPUs at IREN's expanding network of sites. Total contracted AI cloud revenue has reached $3.1 billion of the company's $3.7 billion target for end-2026. The company that set out to mine Bitcoin using stranded hydropower has become, on paper, one of the most significant AI infrastructure providers in the world.

The genuine advantage IREN holds is in the cost and security of its power supply. The company has locked in 2.9 gigawatts of grid-connected capacity across North America at electricity costs of approximately three cents per kilowatt-hour. This is not a marginal advantage. Conventional data center developers starting today face interconnection queues running years into the future and electricity costs that can run two to three times higher. IREN's internal cost to build one megawatt of data center capacity is approximately $4 million, versus an industry standard of $10 million for operators who must acquire land, work through interconnection queues, and build out power infrastructure from scratch. The following table shows where IREN stands relative to its closest peers on the metrics that determine who earns money in AI infrastructure:

Company	Power Cost ($/kWh)	Build Cost ($/MW)	AI Cloud GPU Price (H100/hr)	AI EBITDA Margin
IREN	$0.030–0.033	~$4M	$1.99	85% (Microsoft contract)
CoreWeave	$0.040–0.050 (colocation)	~$10M+	$6.16	57% EBITDA
Core Scientific	Not disclosed	~$10M+	Not disclosed	Not disclosed
Marathon Digital	$0.034–0.047	~$10M+	Not in AI cloud	N/A

CoreWeave is IREN's most instructive competitor in the GPU cloud market. CoreWeave has $66.8 billion in contracted backlog and 250,000 GPUs deployed — a business at a different order of magnitude. But CoreWeave is an asset-light operator: it leases space from Equinix, Digital Realty, and other colocation providers, which means it pays someone else's cost of capital for every megawatt it occupies. That colocation bill shows up as approximately 20 percentage points of margin disadvantage relative to an owner-operator like IREN. The $1.99 versus $6.16 per-hour H100 pricing differential in the spot market is partly a function of this cost structure. IREN's vertically integrated model eliminates the colocation fee entirely, which is how 85% project-level EBITDA on the Microsoft contract becomes credible math rather than promotional theater.

The limitation of this moat argument is that it depends on GPU pricing remaining elevated. The current pricing premium — IREN at $1.99, CoreWeave at $6.16 — reflects a period when GPU supply was genuinely scarce and any operator with available capacity could charge a premium. As NVIDIA scales Blackwell production, as hyperscalers build out their own infrastructure, and as the wave of announced data center projects from 2024 and 2025 comes online, the spot market rate for GPU hours will compress. IREN's infrastructure advantage protects it in a commodity price environment — it will be profitable at lower GPU rates because its cost structure is lower. But the margin it currently targets may not be achievable on uncontracted capacity once supply arrives. The multi-year contracts with Microsoft and NVIDIA provide insulation; the uncontracted 16% of the 2026 ARR target does not.

The financial profile of IREN today is the profile of a business mid-metamorphosis. Full-year FY2025 revenue was $501 million, representing 168% growth over the prior year, with $86.9 million in GAAP net income — the company's first profitable year. But these numbers are already obsolete. In Q3 fiscal 2026 (the quarter ending March 31, 2026), total revenue was $144.8 million, of which $111.2 million came from Bitcoin mining and only $33.6 million from AI cloud services. GAAP net loss was $247.8 million, reflecting $140.4 million in impairment charges on mining equipment being decommissioned to make room for GPU deployments, plus $23.7 million in unrealized losses on convertible hedge instruments. The operating business is not losing that money — the non-cash charges represent mining hardware being written down as it is retired. But GAAP is what it is, and the distance between the $87 million net income of FY25 and the current quarterly loss of $248 million reflects how much destruction precedes the construction.

Free cash flow is deeply negative, running at approximately -$1.5 billion on an annualized basis. Total debt stands at $3.96 billion against $2.3 billion in cash, a net debt position of approximately $1.66 billion. On May 11, 2026, the company announced a $2.3 billion convertible senior notes offering due 2033, sending the stock down 10% on dilution concerns. Since the IPO in November 2021 at 55 million shares outstanding, the share count has grown to approximately 332 million — a 6x multiplication that has dramatically diluted per-share economics even as the enterprise has grown substantially. Management frames this as the cost of building at speed in a capital-intensive industry. That is partly true. It is also partly the consequence of building a business that requires enormous external capital and will for some time yet.

Daniel and Will Roberts built Iris Energy from a concept into a company with real contracts and real infrastructure. That is not a trivial achievement. Former Macquarie infrastructure bankers are exactly the profile you want running a company whose competitive advantage is locking in long-duration power contracts in obscure jurisdictions — that is literally what infrastructure bankers do for a living. The Microsoft and NVIDIA contracts are sophisticated institutional transactions that reflect genuine counterparty confidence in IREN's ability to deliver. The company's stated record of never missing a construction or commissioning deadline as a public company is a meaningful data point, if it holds.

The capital allocation record is more complicated. CEO Daniel Roberts received $72.6 million in compensation in FY2025 while the company burned through cash at a rate of more than a billion dollars per year. Both founders sold approximately $66 million in stock in September 2025, near the stock's all-time high of $76.87. The Mirantis acquisition — a $625 million all-stock deal for a Kubernetes orchestration software company — was announced in May 2026, precisely as IREN was managing a 34% earnings miss, a $2.3 billion convertible offering, and a 150,000-GPU deployment race against the clock. A management team with more operational bandwidth might have deferred a software acquisition to a quieter moment. The alignment is not absent — both founders retain roughly 4% stakes each in a company with a $20 billion market cap, which is real skin in the game — but the dilution trajectory and the timing of insider sales are not the capital allocation record of a team that prioritizes per-share value above speed.

The growth trajectory of IREN's AI cloud business is the central analytical fact of this investment. The following table shows how the business has developed from its AI cloud origins through the current quarter:

Period	GPUs Deployed	AI Cloud Revenue	AI Cloud ARR (Implied)	BTC Mining Revenue	AI as % of Revenue
FY2024 (full year)	~1,000	$3.1M	~$12M	$184M	~2%
Q1 FY26 (Sep '25)	~12,000	$17.3M	~$69M	$207M	8%
Q2 FY26 (Dec '25)	23,000	~$17–20M	~$75M	$167M	10%
Q3 FY26 (Mar '26)	23,000	$33.6M	~$134M	$111M	23%
Target: Q4 FY26 (Jun '26)	150,000	$925M+ implied	$3,700M	~est. $250M	~78%

The table contains the problem with the current investment thesis in plain arithmetic. AI cloud revenue was $33.6 million in Q3 fiscal 2026, implying an annualized run rate of $134 million. Management's target for Q4 fiscal 2026 — the quarter ending June 2026, one quarter away — is $3.7 billion in AI cloud ARR. The implied quarterly revenue needed to reach that run rate is approximately $925 million. That is a 27-fold increase from Q3's reported revenue in a single quarter. The mechanism to close that gap is GPU deployment: IREN must ramp from 23,000 GPUs to 150,000 in the time it takes. The math of the deployment is internally consistent — 150,000 GPUs at roughly $20,000 per GPU per year in annualized revenue does reach $3 billion — but the operational execution required, specifically deploying 127,000 GPUs and energizing them across multiple facilities in under six months, is an extraordinary undertaking for a company that generated $33.6 million in AI cloud revenue last quarter.

The reason to take the target seriously is that much of it is contracted. The $3.1 billion in secured ARR is real paper: Microsoft's $9.7 billion five-year agreement specifies deployment of 200 megawatts of GPU compute at Childress, Texas across four phases called Horizons. Only Horizon 1 (50 megawatts) was generating revenue in Q3. When Horizons 2, 3, and 4 deploy and begin billing, the annualized revenue jump from that single customer is approximately $1.4 billion. NVIDIA's $3.4 billion five-year partnership adds roughly $680 million in annualized revenue when its air-cooled GPU capacity is activated, currently scheduled for early 2027. The gap between today's AI cloud ARR and the $3.7 billion target is therefore largely a deployment timing question rather than a customer acquisition question. Management has sold; the remaining job is construction and energization at unprecedented scale.

The penetration argument here is unusual. IREN is not a company still selling to prospective customers — 84% of its target is already under contract. The question is not whether demand exists but whether IREN can deploy the supply to meet the contracts it has already signed. Of the company's 2.9 gigawatts of secured grid capacity, the planned 480-megawatt AI cloud deployment by year end represents roughly 17% utilization. The remaining 83% of IREN's grid capacity — approximately 2.4 gigawatts — represents the multi-year growth runway beyond 2026, contractually untapped and contingent on continued execution of a buildout that has not yet demonstrated the pace required.

At $61.20 per share, IREN has a market capitalization of $20.3 billion and an enterprise value of $20.55 billion. Against trailing revenue of approximately $600 million, the stock trades at 34 times revenue — not a multiple that leaves room for execution surprises. Against the $3.7 billion ARR target, the multiple compresses to roughly 5.5 times forward revenue, and if the 85% EBITDA margin on the Microsoft contract is representative, the implied forward EV/EBITDA against $3.15 billion in EBITDA would be approximately 6.5 times — genuinely inexpensive for an infrastructure business with multi-decade asset life and contracted cash flows. The problem is that achieving the $3.7 billion ARR requires exactly the performance that Q3 fiscal 2026 called into question.

IREN's returns on net tangible assets today are poor. The company has $3.96 billion in debt and negative free cash flow, and is consuming capital at a rate that its current revenue cannot service. If the Microsoft and NVIDIA deployments proceed on schedule and generate the contracted cash flows at the stated margins, the return on the capital currently being deployed will eventually be exceptional. An infrastructure asset generating $1.9 billion in annual revenue at 85% EBITDA margins on a 200-megawatt facility that cost roughly $800 million to build is a 200%-plus return on tangible investment. But "eventually" and "if" are doing heavy lifting in that sentence. The Q3 results show a company currently generating 2% of its target AI cloud revenue while consuming over a billion dollars per year in capital. The transition from here to there is real and contracted, but it requires a pace of construction and deployment that the company has not yet demonstrated it can sustain.

The intelligent bear on IREN makes a specific argument: the Q3 revenue miss of 34% below estimates is not noise. It reflects a structural pattern of setting aspirational targets and then discovering that the operational complexity of deploying GPU infrastructure at gigawatt scale is greater than anticipated. If Horizons 2, 3, and 4 at Childress slide by even two quarters — not cancel, just slip — the $3.7 billion year-end ARR target becomes impossible, and a stock priced for that target on a 34x trailing revenue multiple will reprice violently. The answer is that the Microsoft contract contains milestone-based prepayment structures that create real financial consequences for delay, giving IREN strong incentives to deliver on schedule. But incentives and capability are different things. The Q3 revenue miss happened with those incentives already in place.

The two signals that would make this actionable: first, AI cloud quarterly revenue crossing $200 million or above, demonstrating that the Horizon 2 deployment has energized and begun billing. Second, the share count stabilizing — the $6 billion at-the-market program and the $2.3 billion convertible offering represent potential dilution that could consume any per-share value created by the revenue ramp. A company that generates $3.7 billion in AI cloud ARR at 85% EBITDA margins while simultaneously doubling its share count delivers a much smaller per-share outcome than the headline ARR figure implies. The infrastructure thesis can be right while the investment outcome is mediocre, if dilution runs faster than earnings growth.

The power is real. The contracts are real. The gap between current revenue and contracted targets is also real, and it is very large. At $61 per share the market is paying for the contracts to deploy on schedule, the GPU ramp to execute without slippage, and the dilution to stop — three conditions that have yet to be demonstrated simultaneously.