Crypto & Digital Assets
A Bitcoin miner that won the AI-infrastructure lottery — NVIDIA both rents IREN's GPUs and bought its equity at $70, validating the liquid-cooling pivot Culper shorted; the bet is now execution and dilution, not survival.
Research
The verdict
A Bitcoin miner that won the AI-infrastructure lottery — NVIDIA both rents IREN's GPUs and bought its equity at $70, validating the liquid-cooling pivot Culper shorted; the bet is now execution and dilution, not survival.
IREN Limited (Nasdaq: IREN; legal name still Australian-incorporated, HQ Level 5, 55 Market Street, Sydney) is, in its own current words, "a vertically integrated provider of AI Cloud Services, delivering large-scale data centers and GPU clusters for AI training and inference" . That self-description is brand-new. The company IPO'd in November 2021 as **Iris Energy**, a renewable-powered **Bitcoin miner**, and rebranded to IREN in 2024 . As of the March-2026 quarter it runs two businesses: Bitcoin mining (still the majority of revenue) and AI Cloud Services (the strategic future), and it has publicly committed to transitioning out of Bitcoin mining entirely at its flagship Childress campus ``.
How it makes money, plainly:
. (Unusual for a miner — IREN is *not* a Bitcoin-treasury play like MARA; it is a pure power-to-hashrate converter.) Installed hashrate ~**38 EH/s** as of 2026-03-31 .. ~**150,000 GPUs installed or on order** as of 2026-03-31 .The asset base is power. IREN controls seven sites with executed grid-connection agreements totaling 4,510 MW: Childress TX (750 MW), Sweetwater 1 TX (1,400 MW), Sweetwater 2 TX (600 MW), an Oklahoma site (1,600 MW), and three British Columbia sites — Canal Flats (30 MW), Mackenzie (80 MW), Prince George (50 MW) . 100% renewable (clean generation or RECs). Management now markets **~5 GW of secured power globally** including the pending Spanish capacity . Land-and-power is the moat (Lens 3) — the scarce input in the AI build-out is not GPUs, it is energized megawatts, and IREN has been accumulating them since 2018.
Key customers (the pivot's whole story is here):
. Crucially, NVIDIA also took an equity instrument (see Lens 7/9).Key suppliers: NVIDIA (GPUs), Dell (the integration/supply channel — ~$2.3bn Canada + ~$1.2bn USA purchase agreements signed 2026-03-04, and ~$5.8bn of Dell GPUs tied specifically to the Microsoft deal) . Bitmain historically for ASIC miners.
Contract structure is the bull case in one line: the AI contracts are multi-year, take-or-pay-like, with prepayments and a strategic-investor anchor — a profile far more durable than spot Bitcoin mining. The risk is customer concentration: Microsoft + NVIDIA are essentially the entire AI book today.
Upstream inputs → IREN → end customer, named at every link:
. Spain capacity arrives via the **Nostrum Group (Ingenostrum, S.L.)** acquisition (~490 MW, ~EUR 165m) .Chokepoints: (a) NVIDIA GPU allocation (mitigated by the equity tie-up); (b) ERCOT energization timing; (c) capital markets access — uniquely load-bearing here, because IREN must fund >$9bn of spend before the contracts pay out in full (Lens 11/13). Tariff risk is live: IREN disclosed disputes over Bitcoin-miner country-of-origin (Indonesia/Thailand/Malaysia vs. China, +25% tariff claim) and a proposed 100% US semiconductor tariff ``.
The moat is energized, contracted megawatts in the right places — accumulated early and cheaply. IREN began securing renewable-rich, grid-connected land in 2018, years before "AI data center" was a category. That gives it 4,510 MW of grid-connection-agreement capacity (5 GW marketed) at a moment when power, not GPUs, is the binding constraint on AI compute. Replicating that land/power bank takes 3–5 years of permitting and interconnection queues — a durable, time-based moat.
Vertical integration is the second moat: IREN owns the land, builds the data centers, runs the power procurement, and operates the GPU fleet, so it captures margin a pure GPU-reseller (a "neocloud" with leased shells) cannot, and it controls energization timelines. The Nostrum acquisition internalizes construction/engineering; Mirantis internalizes the software layer.
Bargaining power is asymmetric and improving. The NVIDIA equity placement is the single most important moat signal: NVIDIA does not take equity in vendors it sees as fungible. By committing investment rights tied to delivery of up to 600,000 GPUs, NVIDIA has effectively (a) guaranteed IREN priority allocation and (b) publicly certified IREN's liquid-cooling execution — the exact competence Culper's 2024 short ("a Prius at the Grand Prix") said IREN lacked ``. That is a remarkable reversal of the bear narrative.
Where the moat is weak (be honest): GPU-cloud compute itself is close to a commodity — rental rates are set by an oversupplying market (CoreWeave, Nebius, hyperscaler self-build). IREN's switching costs over customers are modest once a contract lapses; the durable edge is the power underneath the GPUs, not the GPUs. Brand/marketing moat: negligible — this is an infrastructure business, won on cost-per-MW and delivery reliability, not brand.
segments.csv is header-only (not populated), so all segment figures below are `` from the income statement, not the CSV.
By product line (USD thousands) ``:
| Period | Bitcoin mining rev | AI Cloud rev | Total | AI % of total |
|---|---|---|---|---|
| FY2023 (yr-end Jun) | 75,509 | 0 | 75,509 | 0% |
| FY2024 | 184,087 | 3,105 | 187,192 | 1.7% |
| FY2025 | 484,629 | 16,394 | 501,023 | 3.3% |
| 9M FY2026 (to Mar-26) | 511,502 | 58,280 | 569,782 | 10.2% |
| Q3 FY2026 (Mar-26) | 111,160 | 33,635 | 144,795 | 23.2% |
The trend is the thesis. Total revenue compounded ~6.6× in two years (FY23 $75.5m → FY25 $501m). But the mix is now inflecting hard: AI Cloud went from 1.7% of revenue (FY24) to 23% in the latest quarter, growing +94% sequentially (Q2→Q3 FY26, $17.3m → $33.6m) and ~9× YoY . Meanwhile Bitcoin mining revenue fell YoY (Q3 FY26 $111.2m vs $141.2m) and sequentially (from $167.4m in Q2 FY26) — by design, as IREN decommissions miners to repurpose Childress for AI ``. So the headline "revenue flat YoY" masks a deliberate substitution of a declining commodity stream for a contracted growth stream.
By geography: as of 2026-03-31, 70% of non-current assets in the USA, 30% in Canada ``. Europe (Spain, via Nostrum) is pending. The center of gravity is decisively Texas (Childress + the two Sweetwater mega-sites + the contracts anchored there).
All `` unless noted.
, but the more important May catalysts (NVIDIA deal + equity placement, Microsoft hedge, Mirantis/Nostrum) dominated the tape — IREN was ~$54 in Jan and ~$59 by mid-June .Unusual vs. its own history: the $140.4m impairment is the first large one since FY2023's $105m, and management flagged a further ~$520m of impairment coming as Childress mining is fully retired ``. Also new: a $111.8m debt-conversion-inducement expense (9M) from inducing convertible holders to convert early — a real cash/shareholder cost of the capital-raising blitz.
transcripts/ is empty, so this is ``-grounded.
Across the last ~4 quarters the management narrative has completely re-centered from "Bitcoin miner with an AI optionality" to "AI infrastructure company with a declining mining tail." The phrases that have risen: "vertically integrated," "5 GW," "annualized run-rate revenue (ARR)," "liquid cooling," "Horizon," "Microsoft," "NVIDIA," "contracted." The phrases that have faded: hashrate guidance, "low-cost Bitcoin mining," EH/s targets — mining is now discussed as something being exited, not grown. On the Q3 FY26 call (2026-05) management leaned into the ARR ladder ($3.1bn → $3.7bn → $4.4bn) and the 5 GW pipeline to redirect attention from the GAAP loss and the revenue "miss" ``. Sentiment: high conviction, high promotion — the founders are selling a vision and backing it with real contracts, but the gap between today's ~$570m 9-month revenue and the $4.4bn ARR target is enormous, and the calls increasingly ask investors to underwrite future delivery.
Peer set: AI-pivoting Bitcoin miners + pure-play neoclouds. Multiples are `` with source/date or n/a. Nothing fabricated.
| Company | Ticker | Mkt cap (USD) | Note / contracted AI backlog | EV/Sales | P/E |
|---|---|---|---|---|---|
| IREN | IREN | ~$21B (2026-06-17) `` | $9.7bn MSFT + $3.4bn NVDA; ~$3.1bn ARR contracted, $4.4bn target | n/a — not cleanly sourced (rev base mid-transition) | ~72x Q3'26 `` |
| CoreWeave | CRWV | ~$43B (Mar-26) `` | $99.4bn revenue backlog (Mar-26); FY25 rev $5.13bn | P/S ~7x `` | n/a |
| Nebius | NBIS | n/a | ~$50bn backlog; Q1'26 rev $399m (+684% YoY) | P/S ~62x `` | n/a |
| TeraWulf | WULF | n/a | ~$12.8–13bn contracted AI rev (Fluidstack/Core42, Google backstop + ~14% stake) `` | n/a | n/a |
| Cipher Mining | CIFR | ~$9.18B (Jun-26) `` | ~$9.3bn HPC backlog (AWS 300MW + Fluidstack) `` | n/a | n/a |
| Applied Digital | APLD | n/a | ~$11bn contracted (CoreWeave 400MW, 15-yr) `` | n/a | n/a |
| Core Scientific | CORZ | n/a | "furthest along" in AI pivot (CoreWeave partner) `` | n/a | n/a |
Read: the AI-miner cohort is being re-rated as infrastructure, not crypto — JPMorgan explicitly reframed the group as "integrated cloud businesses" ``. IREN's ~$21bn cap is below CoreWeave's ~$43bn but its contracted book ($9.7bn + $3.4bn) is in the same league as peers carrying smaller caps (CIFR ~$9bn cap on ~$9.3bn backlog), suggesting IREN is neither obviously cheap nor obviously expensive on backlog — it is priced on execution probability. Clean cross-comparable EV/Sales is genuinely unavailable because every name is mid-transition with a revenue base that doesn't yet reflect the contracted ARR — so I will not invent one.
The tape reveals what the market actually trades IREN on: AI-contract headlines and short reports, far more than Bitcoin.
. Bernstein defended .Pattern: IREN is an AI-infrastructure momentum stock with a Bitcoin beta and a short-report tail-risk. It rallies on contracts/allocation news (Microsoft, NVIDIA, GPU counts) and sells off on (a) short reports questioning execution and (b) dilution/capex prints. Bitcoin price matters at the margin (mining revenue) but is no longer the primary driver.
Co-founders / Co-CEOs: Daniel Roberts and William Roberts — brothers .
insider-transactions.csv not present, so precise current insider % is n/a.Forensic lens. Every figure labeled.
Regulatory findings (required):
):** A putative securities class action (D.N.J., filed 2022-12-14; claims under Exchange Act §§10(b)/20(a) and Securities Act §§11/12(a)(2)/15 tied to the IPO and Nov-2021→Nov-2022 statements) was **dismissed *with prejudice* on 2026-02-18**; lead plaintiffs **filed a Third Circuit appeal on 2026-03-13** . Company calls the claims meritless. Net: the core securities case is largely won at the trial level, on appeal.:** (a) **Canada CRA / GST dispute** — CRA asserts a Canadian permanent establishment; ~**$27.7m GST** under dispute; IREN appealed to the Tax Court of Canada (2025-06-23); deemed "reasonably possible but not probable," **no loss contingency booked** . (b) NYDIG / Non-Recourse SPV equipment-financing default litigation (IE CA 3/4 Holdings, in PwC receivership) — claims against the SPVs and IREN largely unsuccessful; a ~$20m loss contingency was recognized/settled in FY25 ``.Bottom line, Lens 10: no regulator enforcement and the securities case essentially won — but the accounting deserves a skeptic's eye: serial impairments, a giant non-GAAP-to-GAAP wedge, a $112m inducement expense, ~100%-LTV GPU leases, and ~39% dilution in 10 months. The books are aggressive and capital-hungry, not fraudulent.
Bottom-up from actuals + the company's contracted ARR ladder. Every input labeled; outputs ``. This is a transition company where EPS is the wrong primary metric — the loss is impairment/D&A-driven — so I anchor on revenue + Adjusted EBITDA and treat GAAP EPS as a derived, noisy line.
Inputs:
FY2026 (ending Jun-2026) ``: revenue ≈ $0.80–0.85bn [9M actual $569.8m + a Q4 with AI ramping ~$45–60m and mining ~$90–100m → ~$0.21–0.23bn Q4]. Adj. EBITDA ≈ $0.30–0.34bn [40% on the year, less Q4 transition drag]. GAAP EPS: deep negative, ≈ $(1.50)–$(2.50). The print to watch is Adj. EBITDA and ARR-under-contract, not EPS.
FY2027 (ending Jun-2027) ``: revenue ≈ $2.2–3.0bn [Horizons 1-4 (Microsoft, $1.9bn ARR) fully live for much of the year + NVIDIA $3.4bn/5yr ($0.68bn/yr) + residual mining; ramping toward the $3.7–4.4bn ARR exit rate but not fully in-period]. Adj. EBITDA ≈ $1.1–1.6bn [50–55% margin as contracted AI dominates]. GAAP EPS ≈ break-even to modestly positive, ~$0.00–$1.00.
FY2028 (ending Jun-2028) ``: revenue ≈ $3.5–4.5bn [full $4.4bn ARR run-rate in-period + initial Sweetwater 1 ramp]. Adj. EBITDA ≈ $1.9–2.6bn. GAAP EPS ≈ $1.50–$3.50.
Base call (the scoreable one): IREN FY2027 (Jun-2027) revenue ≥ $2.4bn — directly tied to whether Horizon 1-4 energizes on schedule and the Microsoft/NVIDIA ARR converts to recognized revenue. I'd put p ≈ 0.55: contracts are signed and prepaid, but ERCOT energization timing and GPU-delivery phasing are the historical slip points. (Per --watchlist rules, NOT logging this to forecast.ts in the unattended loop — recorded here as the seed for a future tracked forecast.)
GAAP EPS for any of these years is genuinely low-information (impairment/D&A/derivative-mark dominated); the honest forward metric is revenue + Adjusted EBITDA against the ARR ladder, with share count as the per-share haircut.
Adversarial, institutional.
Bull case. IREN spent 2018–2024 quietly hoarding the one input the AI build-out can't manufacture — energized, grid-connected, renewable megawatts (4,510 MW agreed, ~5 GW marketed) — and is now converting it into ~$13bn of contracted, multi-year, prepaid AI revenue anchored by the two best counterparties on earth: Microsoft ($9.7bn) and NVIDIA ($3.4bn + an equity stake). The operating business already prints 40% Adjusted EBITDA margins and positive operating cash flow; the GAAP losses are transition impairments, not a broken model. The single most powerful bull fact: NVIDIA took equity at $70/share tied to delivering up to 600,000 GPUs — Jensen's company does not invest in vendors it thinks will fail to execute, which detonates the entire "can't do liquid cooling" short thesis. If Horizons 1-4 energize on schedule and ARR climbs $3.1bn → $4.4bn, a company throwing off ~$2bn of Adj. EBITDA in FY28 at a ~$21bn EV is not expensive. Optionality on top: Sweetwater (2 GW), Spain (Nostrum), software margin (Mirantis), and a Bitcoin-mining call option they can re-energize if BTC rips.
Bear case (permanent-impairment risks). (1) Capital intensity is existential, not incidental — IREN must fund >$9bn of spend (GPUs, data centers) before the contracts fully pay out, on a base of negative GAAP earnings; if capital markets tighten, the whole ARR ladder stalls and the equity is the residual claimant behind $3.69bn of converts + $3.6bn of new secured debt + GPU leases. (2) Dilution is relentless and compounding — +39% shares in 10 months; bear models assume ~65% more shares by 2035 ``, which can leave shareholders with too little of the upside even if the business succeeds. (3) GPU-cloud commoditization / oversupply — CoreWeave ($99bn backlog), Nebius, and hyperscaler self-build are all adding capacity; GPU rental rates can fall hard, compressing the very margins underwriting the thesis, and GPUs depreciate on a brutal 3-4 year clock. (4) Customer concentration — Microsoft + NVIDIA are the AI book; a renegotiation, slippage, or non-renewal by either is a step-change risk.
Pre-mortem (it's Dec-2027 and the thesis broke): Horizon energization slipped 2-3 quarters on ERCOT interconnection batching; the flagged ~$520m impairment ballooned as more mining sites stranded; a soft BTC tape gutted the mining cash that bridged the build; and IREN had to fund the gap with another large ATM tranche near a depressed price — diluting holders ~20% in a single year. The contracts were fine; the financing and timing weren't.
Is the multiple too high? On today's ~$0.8bn revenue, ~$21bn EV is rich (and the ~72x P/E is meaningless). On the contracted $4.4bn ARR / ~$2bn future Adj. EBITDA, it is reasonable-to-cheap. So the multiple is entirely a bet on execution probability and dilution — there is no "value" floor here, only a forward curve.
Contrarian view (what the market is refusing to see): The crowd still files IREN under "Bitcoin miner" and prices in crypto-beta and short-report risk. The repriced reality is that NVIDIA's equity check converted IREN from a speculative miner into a quasi-anointed NVIDIA capacity partner — closer to a CoreWeave than to a MARA — and that re-rating is only partly done. The inverse contrarian point: the market may also be under-pricing how much dilution will siphon the equity upside even in the bull outcome.
I am short IREN. Here is how the money actually breaks.
A levered, sub-NAV Ethereum proxy run by the best operators in the trade — own the discount-closing buyback, not the "treasury premium" that already died; the bet is ETH plus a 0.83x→1.0x mNAV mean-reversion, fighting a 95% drawdown and a structurally bigger rival.
A 25 EH/s green Bitcoin miner that liquidated its entire treasury to fund a credible-but-unfunded pivot to AI compute — the stock now trades on a CAD $3.5B Gigafactory dream that $23M of cash cannot pay for.
A levered Solana proxy whose only value-creation engine — issuing stock above NAV to buy SOL — has stalled exactly as mNAV collapsed to ~1.0x; below par the flywheel runs in reverse and the $83M Q1 equity wipeout, three material weaknesses, and a SOL-collateralized margin book make this a forced-deleveraging short into the next SOL drawdown.