First-mover in a massive TAM (strong). AI compute infrastructure is a hundreds-of-billions-per-year market that hyperscalers alone cannot build fast enough, and the binding constraint is not GPUs but energised, permitted, cooled data-center capacity. IREN entered that market with an asset most neoclouds lack: more than 5 GW of secured grid connections accumulated during its mining years, including the 2 GW Sweetwater campus in Texas that NVIDIA has named as the flagship site of its up-to-5 GW strategic partnership. Management targets roughly 0.3 GW of IT load delivered in 2026 and 0.8 GW cumulative by 2027, and revenue per megawatt on three-year contracts has risen from roughly $10M (Microsoft, November 2025) to about $20M today with $25M under discussion. The runway is 10–20 years if AI scaling continues; it is materially shorter if it stalls.

Network effects and data flywheel (weak-to-moderate). Compute is closer to a commodity than to a platform. There is no user-side network effect: a Microsoft workload does not make IREN more valuable to Cohere. What exists instead is a supply-side flywheel — each delivered Horizon improves the construction blueprint, lowers cost and time per megawatt, and deepens the NVIDIA relationship (Exemplar Cloud status on GB300 NVL72, NVIDIA Certified Hypervisor via the Mirantis acquisition), which in turn brings earlier hardware allocation and reference customers. The 2026 customer list — Microsoft, NVIDIA, a frontier AI lab, Prometheus, Cohere, Perplexity, Figure AI, Fal AI, Higgsfield AI, Together AI, Fireworks AI — shows the flywheel is real, but it is an execution flywheel, not a data flywheel.

Disruptive technology (moderate-strong). The disruption is in the value chain, not the product: owning power, land, shells, GPUs and now the orchestration software layer lets IREN take a data-center project from grid connection to accepted GPU cluster faster and cheaper than the traditional developer-operator-cloud stack, and lets it sell a managed service rather than bare metal. Delivering 50 MW of liquid-cooled GB300 capacity to Microsoft on schedule while the industry struggles with air-cooled deployments is the proof point. The software layer (managed services, on-demand and premium inference pricing) is the option that could convert commodity compute into higher-margin services, but it does not yet show in revenue.

AI-disruption-resistance (strong — 2 anchors). IREN satisfies anchor (a) Physical — its value chain ends in substations, cooling loops and racks that no AI agent can replicate — and anchor (c) AI infrastructure — the more general AI agents proliferate, the more compute they consume. Both anchors are reinforced by the scarcity of grid connections, which is a multi-year permitting moat. No Section 4 AI-disruption flag is required; the relevant risk is the mirror image — dependence on AI scaling continuing.

Net 7.0/10: a strong, physically anchored executor in a huge market, but a competitor set (CoreWeave, Nebius, Crusoe, Applied Digital, Cipher, TeraWulf, Hut 8, Core Scientific) running the same playbook and hyperscaler customers with pricing power cap this well short of monopoly. Pricing today reflects GB300 scarcity, and scarcity pricing mean-reverts.

Trait 1 — Missionary vision (20%) — 7/10
Co-founders and co-CEOs Daniel and William Roberts articulate a clear thesis — the digital world scales instantly, the physical world does not, so owning energised land at scale is the durable asset of the AI era — and capital allocation traces to it: the mining fleet is being decommissioned in favour of AI infrastructure and the 5 GW pipeline was assembled years before the current demand wave. The vision is specific and has guided a genuine pivot, but it is framed in capacity and contracts rather than in a mission that reaches beyond being the fastest builder.

Trait 2 — Radical long-termism & skin in the game (25%) — 7/10
The brothers each hold about 6% of the company (roughly 12% combined) and have not sold shares in the past twelve months, and they have accepted a year of falling reported revenue and a $703M GAAP loss to reposition the business for a decade. Against that, IREN is an Australian-domiciled company with a single share class and no founder voting control, and management has diluted shareholders by roughly 45% in ten months to fund the build. Long-termism is evident; structural protection of it is not.

Trait 3 — Product & customer obsession (20%) — 7/10
Management talks in operational specifics — megawatts of IT load, revenue per megawatt, acceptance dates, NVL72 liquid-cooling density — and the Mirantis acquisition plus NVIDIA Certified Hypervisor and Exemplar Cloud status show real investment in the customer-facing software layer rather than bare-metal leasing. The customer roster has broadened well beyond Microsoft within two quarters. What is missing is any disclosed customer-level metric: utilisation, renewal rates or net revenue retention are not yet reported.

Trait 4 — Execution velocity (20%) — 9/10
This is the standout trait and the reason the company is on the watchlist at all. Horizon 1 (50 MW, liquid-cooled GB300) was delivered and accepted by Microsoft in August 2026 as promised; contracted ARR reached $4.0B against the $3.7B target set one quarter earlier; total power secured rose from 4.5 GW to more than 5 GW while the mining decommissioning ran in parallel; and Horizons 2–4 are in commissioning and late-stage construction for December 2026. Delivering a harder design on schedule during a business-model switch is rare.

Trait 5 — Capital efficiency & financial discipline (10%) — 5/10
The weakest trait. IREN raised about $19B in twelve months — $2.1B of ordinary shares, $3.0B of convertibles, $3.6B of investment-grade GPU financing at ~6% and $2.8B of sub-investment-grade GPU financing at 9% — and guides FY2027 capex at $25–30B, more than one and a half times the market capitalisation. Customer prepayments funding 45–55% of GPU capex and a fully unencumbered data-center portfolio are genuine mitigants, and the unit economics per megawatt are improving, but a business whose annual capex exceeds its market value cannot survive a two-to-three-year capital-markets freeze without existential dilution. The Costco-algorithm exception applies to the operating model, not to the balance sheet.

Trait 6 — Talent magnetism & organisational scaling (5%) — 8/10
Headcount nearly tripled in FY2026 and five C-suite executives were recruited from NVIDIA, AWS, Oracle and Google, alongside the Mirantis engineering team. That is strong evidence the company can attract the talent an AI-cloud operator needs; it is also an integration risk at this pace, and the culture is still visibly the two founders.

Valuation — FLAG on trailing, WITHIN RANGE on contracted
On the framework's primary base — the most recent full fiscal year, FY2026 ended 30 June 2026, revenue $707.0M — IREN trades at about 24.8x sales, far above the sub-5 entry target. That number measures a revenue base that is being deliberately dismantled: $578M of it is Bitcoin mining that will be switched off by December 2026. The economically relevant base is contracted annualised recurring revenue of $4.0B for 2026 capacity ($4.7B including the NVIDIA contribution scheduled for 2027), on which the market capitalisation of $17.5B is 4.4x (3.7x). On enterprise value — adding $7.4B of debt and subtracting $5.9B of unrestricted cash — the multiple is about 4.8x contracted ARR. Entry discipline is therefore satisfied only on the forward, contracted base, and only if the Horizons are delivered; the trailing figure is flagged.

Revenue and margin trajectory
Reported revenue fell from $144.8M in Q3 to $137.2M in Q4 FY2026, missing consensus by about 13%, because mining capacity was decommissioned faster than AI-cloud capacity came online. Beneath that, AI-cloud revenue rose from $33.6M to $70.5M in one quarter (51% of the total, from 9% a year earlier) and FY2026 AI-cloud revenue of $128.8M was roughly eight times FY2025. Adjusted EBITDA fell to $19.2M in Q4 and the margin compressed from 54% to 35% for the year — the expected cost of running two business models during a switch. Contract pricing has moved sharply in IREN's favour: three-year contracts are at roughly $20M per megawatt of IT load, up about 125% since November 2025, with $25M under discussion. Revenue per megawatt rising while cost per megawatt falls is the operating-leverage setup the transition is meant to produce; it becomes visible in reported numbers only from the March 2027 quarter, after Horizons 2–4 are accepted.

Balance sheet and path to profitability
At 30 June 2026 IREN held $7.6B of cash ($1.7B restricted for the Microsoft build) against $7.4B of non-current debt, with $14B of cash plus committed GPU financing available. GAAP net loss of $702.6M for the year is dominated by $638.8M of non-cash impairments on decommissioned mining hardware; the cash operating business remains EBITDA-positive. The concern is forward, not backward: FY2027 capex guidance of $25–30B must be funded by customer prepayments (45–55% of GPU capex), GPU debt at 6–9% covering 90–96% of the remainder, and equity — including the $6B at-the-market programme and NVIDIA's option to buy 30M shares at $70. The path to profitability is credible on contract economics (five-year hyperscaler contracts against GPUs financed over similar terms), but it is also entirely dependent on capital-markets access remaining open at current spreads, which makes this a Pattern B-sensitive balance sheet rather than the fortress the framework prefers. Score 5.5/10: attractive forward entry multiple, deliberate and improving unit economics, but leverage and dilution that no asset-light platform company carries.

Capital-markets dependence and dilution
Shares outstanding have grown from roughly 272M at the November 2025 peak to 394M, a ~45% increase, and FY2027 capex of $25–30B exceeds the market capitalisation. Every future Horizon is funded at whatever price debt and equity markets offer at the time. A risk-off period in AI credit — a neocloud default, a hyperscaler capex cut, a scaling-law scare — would raise the cost of capital exactly when the company most needs it, and the dilution-adjusted return to today's shareholder could lag the growth of the enterprise substantially.

Scaling-law and GPU-pricing sensitivity
Current $20–25M per megawatt pricing reflects GB300 scarcity. If frontier-model returns to scale flatten, or Blackwell supply catches up, contract prices would fall, margins compress and GPU collateral would depreciate faster than the debt secured against it. A software company can keep creating value on today's models; a leveraged compute builder cannot. This is the core thesis risk and the company itself concedes it.

Customer concentration in Microsoft
The $9.7B Microsoft contract (200 MW at Childress, five years) is the majority of live and near-term contracted ARR, and Horizons 2–4 are built to its specification. Management has broadened the roster to NVIDIA, a frontier lab, Cohere, Perplexity and others, but declined to disclose Microsoft's share of ARR. A delivery dispute, a scope change or non-renewal after year five would be material.

Execution across a global multi-site build
Delivering 0.3 GW in 2026 and 0.8 GW cumulative by 2027 across Texas, British Columbia, Oklahoma, Australia and Spain, while tripling headcount for the second year running and integrating a software acquisition, is a construction and organisational load few companies have carried. Horizon 1 proves the blueprint at 50 MW; Horizons 2–4 in December 2026 will prove it at 200 MW. A slip would push contracted ARR into 2027 and strain the financing plan.

Commodity economics and competition
CoreWeave, Nebius, Crusoe, Applied Digital and several former miners are executing the same power-to-AI playbook, often with hyperscaler anchor tenants. IREN's advantages — owned power, unencumbered data centres, liquid-cooled delivery record, NVIDIA endorsement — are advantages of degree. Over a decade, GPU-hour pricing is likely to behave like any capacity market, and the terminal multiple should be underwritten as infrastructure (5–8x sales), not as a platform.

Reported-revenue optics through December 2026
Consolidated revenue will keep falling as mining is switched off, and each quarter of decline invites earnings-day drawdowns of the kind seen in August 2026 (−7% after hours on a 13% miss). This is a Pattern C feature rather than a thesis break, but it is a source of repeated volatility for a stock with a beta above 4.