IREN owns its power, its land and its data centres, and rents the NVIDIA GPUs inside them to Microsoft, NVIDIA and AI developers on three- to five-year contracts. At $41.30 the market prices it 15% below our base case and just inside the bottom of our fair-value band: the price is our central case discounted at 10.4% instead of 10.0%, or compute prices settling about $0.3m per megawatt a year below ours. The real argument is not how many gigawatts get built. It is how long a GPU keeps earning. Our band is $38.90–60.95, and we rate the shares fairly valued, with outcomes spread very widely either side.
The price sits 15% below our base case, and the gap is one discount-rate choice wide. At a 10.0% WACC our modal base case is $48.60. At $41.30 the market is paying for the same business at about 10.4% — the rate our own inputs give with a 5% equity risk premium rather than 4.6%. Held at 10.0%, the price needs long-run fleet revenue of $15.2m per megawatt of IT load a year instead of our $15.5m, or a 74.8% steady-state margin instead of 76%, or about half the capacity we assume gets built after 2027. The market is not capitalising the $20–25m per megawatt that new contracts earn today; it prices a normalisation slightly deeper than ours. The shares fell 15% from the 22 September close with no company announcement, while CoreWeave and Nebius were little changed.
The value sits in how long a GPU earns, not in how many gigawatts get built. Doubling every megawatt built after fiscal 2027 adds only about $13.85 a share, because new capacity at normalised pricing earns only a modest spread over the cost of capital. Holding the fleet price constant, lengthening GPU life from IREN’s own five-year depreciation policy to the six years CoreWeave uses lifts the base case from $48.60 to about $91; four years takes it below zero. That single assumption moves value more than any growth input, and it is the variable to watch.
The distribution is lopsided, and both tails are real. In our bear case a pricing bust after the committed 2026–27 build leaves discounted cash flows below the debt (−$4.40 a share on the model alone), so the equity is worth little more than a $4.00 floor for undeveloped grid-secured power. In the bull case IREN delivers management’s plan, prices hold above equilibrium for longer, and the shares are worth $164.80. Weighted 25 / 45 / 25 / 5, the scenarios give $82.30, twice the price. The band anchors the rating, so the rating is fairly valued; the weighted value says the payoff is asymmetric upwards, bought with genuine ruin risk.
Run the model in Section 02 backwards at our 10.0% discount rate and $41.30 needs one of four things, each on its own with everything else held at the base case. Long-run fleet revenue of $15.2m per megawatt of IT load a year from fiscal 2032, instead of $15.5m. Or a steady-state adjusted EBITDA margin of 74.8% instead of 76%. Or about half the capacity we assume is built after fiscal 2027 — 1.6GW of IT load in service by June 2036 instead of 2.9GW. Or new capital after fiscal 2040 earning about 9%, below the cost of capital, so that growth destroys value; even with growth earning exactly its 10% cost, the base case is $44.35. None of these is a collapse. Each is our normalisation, drawn a little deeper.
The simplest reading is the discount rate. The price equals our base case at a WACC of about 10.4%, which is what the same inputs give with a 5% equity risk premium instead of 4.6%. Our 10.0% sits at the low end of what current Treasury yields support, and we say so among the known weaknesses in Section 03. At 10.5% the base case is $39.15 and the price sits 5% above it. So the market and our model agree on the business; they differ by less than half a point on what it costs to own it. That is why the verdict is fairly valued: the price is inside the band, 6% above its lower edge, and the disagreement that remains is a matter of judgement about the rate, not about the megawatts.
The shares closed at $48.55 on 22 September and traded at $41.30 at 13:37 New York time on 29 September, 15% lower. There was no company announcement in between. The week brought SemiAnalysis’s ClusterMAX 3.0 ratings on 23 September, which put IREN’s managed cloud service in its lowest category and CoreWeave and Nebius in its highest, and two broker initiations, Rothschild & Co Redburn at Neutral and Jones Trading at Hold. Over the same sessions CoreWeave fell 1% and Nebius rose 2%, while the powered-land landlords Applied Digital, Cipher Digital and TeraWulf each fell about 12%. We discuss the rating in Sections 06 and 07; it does not change any input in our model, because the anchor contracts are bare metal and the exposure it describes sits in years four and five of the fleet’s life, which is already the swing factor.
IREN reports $4.0bn of annualised run-rate revenue contracted for its 2026 capacity, of which $1.0bn was operating on 26 August after Microsoft accepted the first 50MW at Childress, Texas. A five-year $3.4bn NVIDIA contract, about $0.7bn a year, ramps from early 2027. Drawn to scale, $bn of annualised run-rate:
Against that, reported revenue was $0.71bn in fiscal 2026, and our base case for fiscal 2027 is $2.75bn: most of the December-quarter capacity arrives late, and the chief financial officer said on 27 August that its revenue lands “predominantly” in the March 2027 quarter. Run-rate revenue is a company metric, contracted GPU-hour pricing times 8,760 hours, and the company itself warns that recognised revenue may be materially lower. The contracted book covers 2026 capacity. Everything the valuation depends on after that — the roughly 0.5GW due in 2027 and all capacity beyond — is uncontracted, and the price it signs at is the whole question.
Two things decide this valuation, and only one is observable soon. The first is the price per megawatt at which 2027 capacity is contracted: recent three-year deals are above $20m and live talks near $25m, against our $15.5m long-run fleet average. The first contracts are expected in the first half of 2027 and the first results after this report around 4–5 November 2026. The second is how a GPU fleet earns in years four and five, which nobody has yet observed through a full cycle. At $41.30 the price asks for slightly less than our base case on both.
At $41.30 the shares sit inside the sensitivity grid — 15% below the base case, 6% above the band’s lower edge and half the probability-weighted value — so the rating rests on the band as house practice requires, and the weighted value is disclosed rather than averaged in. On multiples the shares look cheap: about 4.3 times year-end contracted run-rate revenue against about 8 for Nebius, and under 3 times our fiscal 2028 base-case EBITDA. Those multiples ignore the capex that follows, and Section 03 explains why the cash flows are the better lens. The shares are 46% below their 5 November 2025 closing high of $76.41.
FY2036 revenue $41.7B · FY2040 return on capital 12.3% · 423.8m shares · terminal value is 134% of enterprise value.
This is an unlevered discounted cash flow that builds IREN megawatt by megawatt over fourteen fiscal years, 2027 to 2040, with GPU refresh and share-based compensation charged as costs. Each year’s revenue is the average IT load in service times fleet revenue per megawatt times utilisation; fiscal 2027 is set directly at $2.75bn because the December-quarter capacity arrives late in the year. Adjusted EBITDA is a margin on that revenue. Share-based pay is charged at $370m, $420m and $470m in fiscal 2027–29 and 2% of revenue thereafter, so unvested awards are kept out of the share count. Cash tax is nil through fiscal 2030 on loss carry-forwards and bonus depreciation, rising to 23% from fiscal 2035. GPUs are paid half a year ahead of delivery; data centres 20 / 60 / 20 across the two years before; development capex for land, grid and long-lead items adds 6% of build capex; payables run at 8% of capex. New customer prepayments are credited back over the three years after receipt, and Microsoft’s 20% prepayment, $1.84bn on the balance sheet, is credited against its fees in fiscal 2029–31. The balance sheet at 30 June 2026 opens fiscal 2027, so every fiscal 2027 flow is counted; flows are valued as at 23 September 2026 with a mid-year convention, the first at 31 December 2026.
The terminal value uses the value-driver formula: fiscal 2041 operating profit after 23% tax, times one minus growth over the return on new capital, over the discount rate less growth. At a 12% return on new capital and 3% growth, growth after fiscal 2040 adds only a modest spread, which is why terminal growth barely moves the answer and the return on new capital does. On base settings the present value of fiscal 2027–40 cash flow is −$7.2bn and of the terminal value $28.5bn: the business consumes cash for four years and the value arrives late, which is also why the discount rate matters so much.
Four operating settings are the thesis. Build path (assumed). Management targets about 0.3GW of IT load in 2026 and 0.8GW cumulatively by end-2027. We assume the 2026 programme lands and the tail of the 2027 programme slips about six months, then 300–400MW a year through fiscal 2033, fading to 3% growth by fiscal 2040: about 3.3GW of IT load, roughly 4.9GW gross, inside the more-than-5GW announced pipeline. The slider scales every megawatt built after fiscal 2027 in proportion. Pricing (derived and assumed). Fiscal 2028’s $16.1m fleet average is a weighted build-up: Microsoft’s 200MW at $9.7m, NVIDIA’s ~50MW at about $13.6m, and the rest at $20m, the disclosed floor for recent three-year contracts. The average peaks at $17.5m as 2027 capacity fills, then falls to $15.5m from fiscal 2034 as three-year contracts roll off and new supply arrives. At $15.5m and $32m of GPUs per megawatt, IREN earns about 12% on invested capital, two points over the cost of capital — a modest premium for owning energised power, and our view of equilibrium. Margins (disclosed anchor). AI cloud cash gross margin was about 87% in the June quarter and the Microsoft contract’s project EBITDA margin was given as about 85%; after corporate costs of about $0.5–0.6bn a year we take 62% in the ramp year, 73% in fiscal 2028 and 76% at scale. GPU life (company policy). IREN depreciates GPUs over five years; the model buys every vintage again at then-current cost at the end of its life. The slider moves depreciation and the refresh cycle together.
Capex anchors are disclosed. Microsoft’s GPUs cost $5.8bn for 200MW, about $29m per megawatt; the company gave all-in data-centre costs of $14–16m per megawatt for Horizons 1–4. The base case’s fiscal 2027 capex is about $15bn, below the $25–30bn guide, because the slower 2027 path moves spend into fiscal 2028. The 10.0% discount rate is built from a 4.96% risk-free rate (the ten-year Treasury on 21 September 2026), a fundamental beta near 1.8 (an asset beta of about 1.1 relevered at the target debt weight; the statistical beta above 4 reflects volatility, not systematic risk) and a 4.6% equity risk premium, for a cost of equity of about 13.3%; an after-tax cost of debt of about 6% blends 6.0% investment-grade and 9.0% sub-investment-grade GPU debt, future data-centre debt and the economic cost of the convertibles, at a 45% debt weight. The same inputs with a 5% premium give about 10.4%, where the base case is $40.95. No separate premium is added for dependence on external funding. Each half point moves the base case by $9–11 a share.
Share count and bridge. 394.06m basic shares (FY2026 Form 10-K cover) plus 29.74m from the 2029 3.50% and 2030 3.25% notes, which are deep in the money and counted as converted: 423.8m in the base case. The 2031 0% ($1.0bn, conversion $85.63), 2032 0.25% ($1.15bn, $51.40), June 2033 1.00% ($1.15bn, $51.40) and December 2033 1.00% ($3.0bn, $73.07) notes are below their conversion prices at the base value and are deducted at face, $6.3bn; above a conversion price a series is counted as converted, with its capped call credited up to the cap. Also deducted: $938m of drawn GPU financing, $244m of finance leases, the $40m Mirantis cash consideration paid in August 2026, and NVIDIA’s right to buy 30m shares at $70, valued by Black-Scholes at the model-implied price with 75% volatility ($788m at the base value). Cash of $7.62bn includes $1.72bn restricted, mostly for Microsoft GPU capex. Unvested RSUs, including the co-CEOs’ 18.2m granted on 1 July 2026, are not added, because share-based pay is charged as a cost. The bridge is iterated until the price at which the notes and the right are counted equals the value it produces.
The number above is not a fact about IREN. It is what these assumptions produce, and the two that decide it are not visible in any filing yet: the price at which capacity after 2026 is contracted, and how many years a GPU earns before it is replaced. The third is a judgement about the cost of capital for a levered builder of depreciating hardware. The rating holds for a long-run price between $14.8m and $15.6m per megawatt, or a steady margin between 73% and 76.4%: a little above our assumptions on either, the band’s lower edge rises above the price and the reading turns undervalued. It does not survive a change in GPU life: four years makes the shares overvalued, six makes them undervalued — which is why that is the variable to watch.
Each card loads its published path into the model above and prints that path’s value: −$4.40, $48.60, $164.80 and $364.40. All four use the same 10.0% rate and 3% terminal growth, so the differences between them are operating paths only; construction delay, pricing and funding risk sit in the paths and the weights, not in the discount rate. The bear is valued at a $4.00 floor in the weighting, not at its negative model value: undeveloped grid-secured power has a value at distressed prices, and the GPU debt sits in financing subsidiaries with limited parent guarantees. The floor is a judgement and is not modelled line by line. Weighting the four at 25 / 45 / 25 / 5 gives 0.25 × $4.00 + 0.45 × $48.60 + 0.25 × $164.80 + 0.05 × $364.40 = $82.30.
| Input | Bear (25%) | Base (45%) | Bull (25%) | Super-bull (5%) |
|---|---|---|---|---|
| IT MW, June 2027 / 2028 | 400 / 620 | 450 / 780 | 500 / 900 | 520 / 950 |
| IT MW, FY2036 (FY2040) | 720 (775) | 2,920 (3,350) | 4,000 (4,560) | 5,500 (6,440) |
| Fleet revenue per MW, $m | 14.5 FY28 → 9.5 from FY33 | 16.1 FY28, 17.5 peak → 15.5 from FY34 | 19.0 peak → 17.5 | 21.0 peak → 20.0 |
| Utilisation and realisation | 84–86% | 87% FY28 → 94% | 90% → 95% | 91% → 96% |
| FY2027 revenue | $2.45bn | $2.75bn | $2.95bn | $3.00bn |
| Adjusted EBITDA margin, steady | 70% | 76% | 78% | 80% |
| GPU capex per MW, $m | 30 → 33 | 30 → 32 | 30 → 32 | 30 → 32 |
| Data-centre capex per MW, $m | 15.0 | 14.0 | 13.0 | 13.0 |
| GPU refresh | None: halls re-let as colocation at $1.9m/MW | Every 5 years at then-current cost | Every 5 years | Every 5 years |
| Customer prepayment, % of GPU capex | 30% → 5% | 35–40% → 15% | 45% → 20% | 50% → 25% |
| Return on new capital after FY2040 | = WACC | 12% | 15% | 18% |
| Common to all | WACC 10.0%, terminal growth 3.0%; cash tax nil in FY2027–30, 10% in FY2031–32, 18% in FY2033–34 and 23% from FY2035; share-based pay $370m / $420m / $470m in FY2027–29, then 2% of revenue; payables 8% of capex; development capex 6% of build capex; Microsoft’s prepayment credited in FY2029–31 | |||
| Value per share | $4.00 floor (model −$4.40) | $48.60 | $164.80 | $364.40 |
| $bn unless stated, FY to June | FY27 | FY28 | FY29 | FY30 | FY31 | FY33 | FY36 | FY40 |
|---|---|---|---|---|---|---|---|---|
| IT MW, year end | 450 | 780 | 1,160 | 1,520 | 1,860 | 2,430 | 2,920 | 3,348 |
| Revenue | 2.8 | 8.6 | 15.8 | 22.0 | 26.7 | 33.7 | 41.7 | 48.1 |
| Adjusted EBITDA | 1.7 | 6.3 | 11.8 | 16.8 | 20.3 | 25.6 | 31.7 | 36.5 |
| Share-based pay (charged) | (0.37) | (0.42) | (0.47) | (0.44) | (0.53) | (0.67) | (0.83) | (0.96) |
| Depreciation and amortisation | (3.4) | (5.9) | (8.5) | (10.6) | (12.9) | (16.7) | (19.8) | (23.1) |
| EBIT | (2.1) | (0.0) | 2.8 | 5.7 | 6.9 | 8.3 | 11.0 | 12.4 |
| Capex including refresh | (14.9) | (17.3) | (17.9) | (18.2) | (15.6) | (21.6) | (16.6) | (22.8) |
| Working capital | 1.6 | 2.9 | 1.2 | (1.2) | (2.4) | (0.1) | (0.8) | (0.5) |
| Unlevered free cash flow | (12.0) | (8.5) | (5.3) | (3.0) | 1.1 | 1.7 | 10.9 | 9.5 |
| Return on invested capital | −9% | 0% | 5% | 9% | 11% | 10% | 13% | 12% |
Free cash flow is unlevered and charges share-based pay as a cost. Return on invested capital is EBIT after 23% tax over net GPU and data-centre capital. The capex steps in fiscal 2032 and 2037 are the five-year refresh of the fiscal 2027 and fiscal 2032 fleets. Present value of fiscal 2027–40 cash flow −$7,170m; terminal value $28,457m; enterprise value $21,287m.
| Base case, $m | Value |
|---|---|
| Present value of free cash flow, FY2027–40 | −7,170 |
| Present value of terminal value (value driver, 12% return on new capital) | 28,457 |
| Enterprise value | 21,287 |
| + Cash including restricted, 30 June 2026 | 7,620 |
| − Drawn GPU financing | −938 |
| − Finance leases | −244 |
| − Mirantis cash consideration, August 2026 | −40 |
| − Convertibles below conversion price, at face (2031, 2032, June 2033, December 2033) | −6,300 |
| − NVIDIA right, 30m shares at $70 (Black-Scholes at the implied price, 75% volatility) | −788 |
| Equity value | 20,597 |
| ÷ Shares: 394.06m basic + 29.74m from the 2029 and 2030 notes as converted | 423.8m |
| Value per share | $48.60 |
The grid holds the base operating path fixed and varies the discount rate against terminal growth. The bold rows are the published fair-value band, $38.90–60.95: WACC 9.5–10.5% by terminal growth 2.5–3.5%.
| WACC ↓ / terminal growth → | 2.5% | 3.0% | 3.5% |
|---|---|---|---|
| 9.0% | $70.81 | $72.58 | $74.73 |
| 9.5% | $58.52 | $59.61 | $60.93 |
| 10.0% base | $47.99 | $48.60 | $49.35 |
| 10.5% | $38.89 | $39.16 | $39.51 |
| 11.0% | $30.96 | $30.98 | $31.05 |
Terminal growth barely matters, because the 12% return on new capital sits close to the discount rate; the discount rate matters a great deal, because most of the value arrives after fiscal 2030. The $41.30 price sits between the 10.0% and 10.5% rows, 6% above the band’s lowest cell. The grid is reproducible from the model above by moving the last two sliders.
Modal base case $48.60, 18% above the price. Fair-value band $38.90–60.95, the WACC 9.5–10.5% by growth 2.5–3.5% grid, which anchors the rating; the price is inside it, near its lower edge: fairly valued. Probability-weighted value $82.30 across the four scenarios at our own weights, about twice the price. The skew is strongly to the upside because the bear case is capped by limited liability while the bull and super-bull are not. The two anchors disagree, and the report rates on the band, as house practice requires.
| Single input changed, all else at base | Base | Price implies, 10.0% WACC | Price implies, 10.5% WACC |
|---|---|---|---|
| Discount rate | 10.0% | 10.4%: the price equals the base case | |
| Long-run fleet revenue per MW, FY2032 on | $15.50m | $15.20m | $15.59m |
| Steady-state adjusted EBITDA margin | 76.0% | 74.8% | 76.4% |
| Return on new capital after FY2040 | 12% | 8.9% — below the cost of capital | 13.6% |
| IT load by June 2036 (build after FY2027 as a multiple of the base path) | 2.92GW (1.0×) | 1.62GW (0.47×) | 3.76GW (1.34×) |
At our 10.0% rate the price asks for a little less than the base case on any one operating input: $0.3m per megawatt of long-run pricing, 1.2 points of margin, or about half the post-2027 build. Only the growth reading is severe — the price would need growth after fiscal 2040 to destroy value — and that is a symptom of how little value the base case gives growth in the first place: even at a return on new capital equal to the 10% cost of capital the base case is $44.35. At 10.5% the direction flips and the price needs slightly more than the base case: $15.59m per megawatt, a 76.4% margin, or a third more capacity. The price is bracketed by our base case at 10.0% and 10.5%, and the market is not paying for megawatts for their own sake, nor capitalising super-normal returns forever.
| Risk | Priced in | Not also in |
|---|---|---|
| Construction and grid delays | Build path (six-month slip) and scenario weights | WACC, margins |
| Pricing-cycle normalisation | Fleet price path and the 12% return on new capital | WACC, utilisation |
| GPU obsolescence | Five-year refresh capex at full cost | Price path (held per megawatt) |
| Customer credit outside Microsoft and NVIDIA | Utilisation and realisation (94% steady) and the bear case | WACC |
| Funding and future dilution | Leverage-adjusted cost of equity in the 10.0% WACC, with no separate premium; the markets-closed tail in the bear case; new equity assumed issued at fair value | Cash flows, terminal |
| Microsoft prepayment credit-back | Working capital, −$1.84bn across FY2029–31 | Equity bridge |
| Convertible and NVIDIA dilution | Equity bridge (as converted above the conversion price; Black-Scholes for NVIDIA) | Cash flows |
| Governance, litigation, Texas policy | Not priced; disclosed in Sections 07 and 08 | — |
| Managed-service quality | Not priced; disclosed in Sections 06 and 07 | — |
On forward run-rate revenue IREN trades at a discount to Nebius and near CoreWeave; on trailing numbers every metric is meaningless. All prices are 29 September 2026 intraday quotes taken at 13:37 New York time.
| Company | Mkt cap | EV | Growth y/y, latest qtr | Adj. EBITDA margin | EV / revenue | Forward anchor |
|---|---|---|---|---|---|---|
| IREN $41.30 | $16.3bn | $17.4bn | AI cloud +1,860%; total −27% | 14% (Q4) | 6.2× FY27E | EV / December 2026 contracted run-rate $4.0bn: 4.3× (3.7× on $4.7bn including NVIDIA) |
| CoreWeave $85.82 | $47.3bn | ≈$75bn | +112% | 59% (Q2) | ≈5.9× FY26E | EV / $104bn backlog ≈ 0.73×; EV / FY26E EBITDA ≈ 10× |
| Nebius $240.06 | $65.3bn | ≈$66bn | +454% | 41% (Q2) | ≈20× FY26E | EV / December 2026 run-rate guide ($8bn midpoint) ≈ 8.3× |
| Applied Digital $25.12 | $7.3bn | n.c. | +407% (FQ4) | 16% (FQ4) | n.c. | Landlord model; ~1.4GW leased to hyperscale tenants |
| Cipher Digital $16.75 | $6.9bn | n.c. | n.c. | n.c. | n.c. | Landlord model; HPC leases |
| TeraWulf $15.20 | $7.6bn | n.c. | n.c. | n.c. | n.c. | Landlord model; HPC leases |
Market capitalisation is the 29 September 2026 intraday quote times the latest filing share count: IREN 394.06m; CoreWeave 551.5m, all classes; Nebius 271.9m; Applied Digital 291.5m; Cipher 415.0m; TeraWulf 499.0m. IREN’s enterprise value counts the 2029 and 2030 notes, which are above their conversion prices, as equity and all other debt at face, less $7.62bn of cash including restricted cash at 30 June 2026; it excludes the $1.84bn Microsoft prepayment. CoreWeave and Nebius enterprise values use net debt from results summaries and filing extracts and are approximate (≈). Landlord enterprise values are not computed (n.c.): they carry project-level debt and tenant-funded fit-outs that make the comparison unlike for like. P/E, forward P/E and PEG are not meaningful for any of the three neoclouds, all loss-making; IREN fiscal 2027 consensus EPS is −$3.92. Price to free cash flow is not meaningful: IREN’s fiscal 2026 free cash flow was about −$2.3bn and fiscal 2027 consensus is about −$20.7bn. Consensus as at 21 September 2026.
On the one forward metric the three neoclouds share — enterprise value against year-end 2026 run-rate revenue — IREN is the cheapest: about 4.3× against about 8.3× for Nebius. On enterprise value per megawatt the picture is similar: about $22m per megawatt of IT load targeted for end-2027, against about $18m per gross megawatt of CoreWeave’s contracted power, roughly $25m per IT megawatt once converted. On our base-case fiscal 2028 EBITDA of $6.3bn, IREN trades at about 2.8×, which looks absurdly cheap until one remembers that every neocloud must spend roughly its entire EBITDA again, and more, on capex for years. EV multiples that ignore the capex are the wrong lens; the cash flows above are the right one.
Is the discount already paying for the risk? Partly. IREN’s discount to Nebius reflects a smaller, more concentrated book — Microsoft is about half the 2026 run-rate — a heavier dependence on external funding, founder control, and now an independent rating of its managed service at the bottom of the industry. Our 10.0% WACC carries no separate premium for funding risk: that risk sits in the leverage-adjusted cost of equity and, in its severe form (capital markets closing), in the bear case. We add no multiple discount and do not haircut the cash flows for it. What could expand the multiple: 2027 capacity contracted at $25m per megawatt with high prepayments; a first data-centre financing on investment-grade-like terms; evidence that three-year-old GPUs re-contract at a high share of their original price; NVIDIA exercising its $70 right. What could contract it: a fall in spot or short-term GPU pricing, a customer default, a missed Horizon delivery date, or a large equity raise at a discount.
Intrinsic value is not a price target and this report does not publish one. But in twelve months the market will price IREN on next-twelve-months EBITDA, and the multiple regime it applies is worth being explicit about. At 6.5 times next-twelve-months EBITDA in September 2027, the range CoreWeave now trades in, our base-case numbers support about $87 (5× gives $62; 8×, $111). It sits well above the price and above our cash-flow value, because neocloud multiples today capitalise EBITDA without charging the capex that follows it. This is a market-regime estimate, not a fair-value estimate, and it is not the anchor of the rating; we disclose the tension rather than resolve it.
GPU life. We use IREN’s own five-year depreciation life. Holding the fleet price constant, six years gives about $91; that overstates, because older GPUs earn less, but it shows where the value is most fragile. Four years gives a negative equity value. Above consensus near term. Base fiscal 2028 revenue of $8.6bn is about 14% above the $7.55bn consensus because we credit disclosed pricing on 2027 capacity that is not yet contracted. The bear floor is a judgement. The bear model value is negative; the $4.00 floor values undeveloped grid-secured power at distressed prices and reflects the ring-fencing of GPU debt in financing subsidiaries; neither is modelled line by line. Items not credited. Capped calls on the 2029 and 2030 notes and the 2029 prepaid forward are not credited, together worth perhaps $0.50–1.25 a share; convertible time value is ignored on both sides, per house convention. Derived openings. Operating IT load at 30 June 2026 (about 40MW), the opening split of plant and equipment, and the cash-tax path are derived, not disclosed; peer enterprise values for CoreWeave and Nebius are approximate. Discount rate. Our 10.0% WACC is at the low end of what its inputs support: with the ten-year Treasury at 4.96% (21 September 2026), a 1.8 beta and a 5% equity premium give about 10.4%; 10.0% requires a 4.6% premium. At $41.30 the price equals the base case at almost exactly that 10.4%, so the market’s reading and the fuller-premium reading of our own inputs coincide. At 10.5% the base case is $39.15 and the band $30.95–49.35, with the price near its middle; the rating would still read fairly valued. A thin margin on the upside. The price is 6% above the band’s lower edge; a long-run price $0.1m per megawatt above ours, or a margin half a point higher, would lift the lower edge above the price. Two anchors disagree. The band says fairly valued; the probability-weighted value, about twice the price, says undervalued. We rate on the band, as house practice requires. And the market data are intraday: the $41.30 price was captured during the 29 September session at 13:37 New York time, and the close will differ.
IREN reported its fourth quarter and fiscal 2026 on 27 August. Revenue of $137.2m missed a consensus near $157m, adjusted EBITDA fell to $19.2m, and the shares closed at $40.53 that day and $35.45 the next, down 12.5%. The quarter was the trough of a transition: mining was switched off before the GPUs came on.
| Metric | Q4 FY26 | Q3 FY26 | Change | Why it matters |
|---|---|---|---|---|
| Total revenue, $m | 137.2 | 144.8 | −5% | Trough quarter of the transition |
| AI cloud revenue, $m | 70.5 | 33.6 | +110% | The only line that matters from here; the first quarter above mining |
| Bitcoin mining revenue, $m | 66.7 | 111.2 | −40% | Wound down by the end of 2026 |
| AI cloud cash gross margin derived | 87.0% | 86.3% | +0.7 pts | Supports a high steady-state EBITDA margin |
| SG&A including share-based pay, $m | 128.3 | 81.8 | +57% | The platform is being built ahead of revenue |
| Adjusted EBITDA, $m | 19.2 | 59.5 | −68% | Margin 14%, from 41% |
| Net loss, $m | (684.0) | (247.8) | n.m. | Mostly $450.4m of non-cash impairments on mining hardware |
| Operating cash flow, $m | 1,811.1 | 75.3 | n.m. | Microsoft’s prepayment: $1.72bn of deferred revenue |
| Capex, $m | 1,978.2 | 1,355.3 | +46% | Heading to $25–30bn in fiscal 2027 |
| Operating run-rate revenue | ≈$0.5bn | n.d. | $1.0bn at 26 August after Horizon 1 |
Source: fourth-quarter and fiscal 2026 results release, Form 10-K and conference call, 27 August 2026. Consensus as reported after the print and not independently verified.
One: the revenue ramp is real but late. Revenue fell 5% on the quarter as mining revenue dropped by $44.5m while AI cloud added $36.9m. The chief financial officer said most December-quarter capacity arrives late, so its revenue lands “predominantly” in the March 2027 quarter.
Two: pricing has moved decisively. Three-year contracts are pricing above $20m of annual revenue per megawatt with about a two-year payback on compute capex, and live discussions sit near $25m. This is the single most important disclosure on the call, because pricing, not volume, sets the value.
Three: customers are funding the build. Recent prepayments cover 45–55% of GPU capex; combined with 90% GPU financing on non-investment-grade deployments, total funding exceeds the GPU cost. The data-centre portfolio, including Horizons 1–4, is entirely unencumbered and available for future asset financing.
Four: the capex bill is enormous. Fiscal 2027 capex guidance is $25–30bn against about $14bn of cash, committed GPU financing and prepayments. The company targets a further ~$8bn of GPU financing and prepayments; the rest must come from data-centre financing, operating cash flow and “corporate sources”.
Five: costs step up ahead of revenue. Headcount nearly tripled in fiscal 2026, and cash SG&A for the September quarter is guided up $40–50m on the quarter. AI cloud’s direct cost of revenue was $9.2m on $70.5m, an 87% cash gross margin before depreciation, consistent with the ~85% project EBITDA margin the company gave for the Microsoft contract; group margins collapsed because SG&A, including $42.9m of share-based pay, rose to $128.3m.
Guidance: run-rate revenue above $4bn by the end of the December 2026 quarter, all under contract; $1bn held through the September quarter; fiscal 2027 capex of $25–30bn; mining effectively decommissioned by December 2026; September-quarter cash SG&A up $40–50m. The tone was confident and thesis-led: the co-chief executive opened with industry token-growth charts and framed the scarcity of physical capacity as structural. The chief financial officer was more measured, repeatedly noting that actual capex and funding depend on schedules, costs and “the overall fundraising environment”.
Goldman Sachs asked what the $25–30bn covers and how the gap beyond $14bn gets funded. Macquarie probed the returns on converting British Columbia sites to liquid cooling. Cantor asked whether $20–25m per megawatt is representative and over what term; the answer was three to five years, consistent across live talks. B. Riley asked about the Texas governor’s data-centre directive and expansion options. BTIG asked about data-centre financing: many preliminary conversations, none closed. Canaccord asked whether industry financing can keep pace.
| Positive surprises | Negative surprises |
|---|---|
| A new multi-year contract with an unnamed frontier AI lab | Revenue below consensus |
| The pricing disclosure: $20m-plus per megawatt, talks near $25m | A sharper EBITDA fall than expected |
| $2.8bn of GPU financing for non-investment-grade customers, funding 90% of GPU capex | The size of the capex guide against secured funding |
| Mirantis named an NVIDIA-certified hypervisor | No data-centre financing closed yet |
What to watch next. Microsoft’s acceptance of Horizons 2–4 and the operating run-rate printed with the first-quarter fiscal 2027 results; the first contracts for 2027 capacity, and at what price per megawatt; the first data-centre-level financing; first-quarter cash SG&A against the +$40–50m guide; and whether prepayment terms hold at 45–55% as supply grows.
IREN, formerly Iris Energy, was founded in Sydney in 2018 by the brothers Daniel and William Roberts, who remain co-chief executives. It built its business mining Bitcoin at sites with cheap, largely renewable grid power in British Columbia and Texas. Since 2024 it has converted those sites, and a much larger development pipeline, into an AI cloud platform. The company now describes itself as vertically integrated across three layers: the data centres themselves — land, grid connections, substations, cooling; the compute inside them — NVIDIA GPUs, storage and networking, owned by IREN; and a software and services layer added with the August 2026 acquisition of Mirantis.
The commercial model is GPU cloud services: customers contract dedicated capacity, mostly bare metal, for multi-year terms, and pay for GPU hours. That is different from the “powered shell” landlords such as Applied Digital, Cipher Digital or TeraWulf, which lease buildings to a tenant who brings its own GPUs. IREN carries the GPU capital and the GPU risk, and it keeps the GPU margin. Because it owns the buildings it pays no colocation rent, which is why its economics differ from CoreWeave’s, much of whose capacity is leased.
| Stream | FY2026, $m | FY2025, $m | Direction |
|---|---|---|---|
| AI cloud services | 128.8 | 16.4 | $70.5m in the June quarter, the first quarter above mining |
| Bitcoin mining | 578.2 | 484.6 | Being decommissioned by the end of December 2026 |
| Total revenue | 707.0 | 501.0 | +41% year on year |
The forward mix is almost entirely AI cloud: $4bn of run-rate revenue contracted for 2026 capacity, $1bn of it operating after Microsoft accepted Horizon 1 at Childress, plus the five-year $3.4bn NVIDIA contract from early 2027. Microsoft’s $9.7bn five-year contract, signed in November 2025, works out at about $9.7m of revenue per megawatt a year; recent three-year deals price above $20m.
Three trends matter. First, compute pricing turned sharply upwards through 2026. IREN says three-year contract pricing rose about 125% between November 2025 and August 2026 and five-year pricing about 70%, with recent three-year deals above $20m of annual revenue per megawatt of IT load and live discussions near $25m; Nebius independently reported new deals averaging above $20m per megawatt in its second-quarter 2026 letter. Second, customers are paying in advance: IREN cites prepayments of 45–55% of GPU capex on recent contracts, and Nebius reported 50–60%. Third, the financing market for GPUs has formed quickly: IREN raised $3.6bn of investment-grade GPU debt at about 6.0% for the Microsoft contract and $2.8bn for non-investment-grade customers, including $2.4bn at a 9.0% fixed rate led by Blue Owl and PIMCO. Against those trends sit the classic risks of a capital-spending boom: supply catching up, GPU generations turning over every year or two, and customers who are mostly still loss-making AI developers.
The closest competitors are the other GPU-owning neoclouds: CoreWeave, with about $104bn of revenue backlog and 1.5GW of active power at 30 June 2026, and Nebius, with $3bn of run-rate revenue in June and guidance of $7–9bn by December 2026. Hyperscalers are both customers and competitors: Microsoft is IREN’s largest customer and also builds its own capacity. Powered-land and data-centre landlords — Applied Digital, Cipher Digital, TeraWulf, Galaxy, Hut 8 — compete for the same grid connections and construction labour but sell a different product. IREN’s differentiators are owned, energised sites (750MW at Childress, a 2GW Sweetwater hub whose first substation was energised on schedule in May 2026, and British Columbia), no third-party rent, and an NVIDIA partnership that includes a five-year right for NVIDIA to buy up to 30 million IREN shares at $70.
| Disclosed, $m | Q4 FY26 | FY2026 | FY2025 |
|---|---|---|---|
| Revenue | 137.2 | 707.0 | 501.0 |
| Adjusted EBITDA | 19.2 | 245.7 | 269.7 |
| Net income (loss) | (684.0) | (702.6) | 86.9 |
| Impairments, mostly mining hardware | 450.4 | 638.8 | 7.2 |
| Share-based compensation | 42.9 | 205.0 | 42.6 |
| Operating cash flow | 1,811.1 | 2,100.4 | 245.9 |
| Capex plant, hardware, intangibles | 1,978.2 | 4,440.7 | 1,372.7 |
| Cash including restricted, period end | 7,619.5 | 7,619.5 | 564.5 |
Fiscal 2026 is a transition year and its numbers flatter little. Operating cash flow was lifted by a $1.72bn increase in deferred revenue in the June quarter, which is Microsoft’s 20% prepayment; it is credited back against Microsoft’s fees in years three to five of the contract, so it is a timing benefit, not earnings. The net loss reflects $638.8m of mostly non-cash impairments as mining hardware is retired. The balance sheet at 30 June 2026 held $7.62bn of cash, including $1.72bn restricted and mostly earmarked for Microsoft GPU capex, against about $6.75bn face value of convertible notes across six series, $938m of drawn GPU financing and $244m of finance leases. The share count rose about 42% in a year, to 394.06m on the latest filing cover, as the company sold $4.7bn of equity in fiscal 2026.
Disclosed drivers: delivery of Horizons 2–4 to Microsoft in the December 2026 quarter, taking operating run-rate revenue towards $4bn; the NVIDIA contract from early 2027; a target of about 0.8GW of IT load delivered cumulatively by end-2027 (1.2–1.25GW gross); the Sweetwater, Kiowa (Oklahoma), Bundey (South Australia) and Badajoz (Spain) developments for 2028 onwards; and more than 5GW of announced pipeline. Assumed in our base case: 2027 capacity contracted at about $20m per megawatt, a roughly six-month slip on the tail of the 2027 programme, and 300–400MW a year of new capacity thereafter.
Risks, in order of weight: a turn in the compute-pricing cycle before the 2027–28 capacity is contracted; GPU economic lives shorter than five years; funding a $25–30bn fiscal 2027 capex programme without destructive dilution; customer credit outside Microsoft and NVIDIA; construction and grid-connection delays, including Texas large-load reforms; and governance, where the founders’ B-class shares carry 15 votes per ordinary share held. Section 07 scores each.
| Source | Score / 5 | Assessment against CoreWeave, Nebius and the powered-land landlords |
|---|---|---|
| Cost advantage | 4 | Owned data centres mean no colocation rent — roughly $1.5–2m per megawatt a year for a lessee, our estimate — and sites were secured at Bitcoin-era prices with low-cost, largely renewable power. This is the core of the moat |
| Regulation and access | 4 | Grid interconnection queues, especially in Texas, are lengthening and tightening. Energised sites with approved connections are hard to replicate within three years |
| Scale | 3 | About 0.8GW of IT load targeted by end-2027, against CoreWeave’s 4.2GW of contracted power. Big enough to win hyperscaler and frontier-lab contracts; not the largest |
| Distribution | 3 | The NVIDIA partnership (up to 5GW of DSX-aligned deployment) and the Microsoft relationship open doors; most contracts are still sold directly |
| Customer loyalty | 3 | Together AI and Fireworks AI renewed and expanded; the unnamed July 2026 developer has been named as Prometheus. Early evidence only |
| Brand | 2 | Credible with hyperscalers after Microsoft and NVIDIA; little pull with end users. CoreWeave has the stronger brand, rated Platinum by SemiAnalysis’s ClusterMAX 3.0 against Underperforming for IREN |
| Switching costs | 2 | Three- to five-year bare-metal contracts lock in revenue, not loyalty. Mirantis orchestration could raise stickiness for smaller customers over time |
| Intellectual property | 2 | In-house data-centre design and the Mirantis software stack. Useful, but replicable |
| Network effects | 1 | None of substance. More customers do not make the product better for other customers |
| Data | 1 | No proprietary data advantage |
| Total | 25 / 50 | Narrow moat, expanding |
Strongest advantage: energised power and owned buildings. In a market where the binding constraint is getting electricity to a building, IREN’s energised sites are an asset that competitors need years and billions to replicate. Owning the buildings removes rent and lets IREN capture the whole stack’s margin; the ~85% project EBITDA margin disclosed on the Microsoft contract is the proof point.
Biggest threats. GPU capital is not a moat: anyone with financing can buy the same chips. If power constraints ease — new generation, relaxed interconnection, behind-the-meter gas — the scarcity rent on IREN’s sites shrinks. Hyperscalers building their own capacity could also turn from customers into competitors. In a downturn, the customers most likely to churn are the non-investment-grade AI developers that make up most of IREN’s non-Microsoft book. Service quality is the other weak point: the customers most likely to take three-year-old GPUs are smaller developers who buy a managed service, and that is where SemiAnalysis currently rates IREN near the bottom of the industry.
Direction: expanding, for now. Each contract adds customers and references, and each energised site extends a lead measured in years. The moat is narrow because it rests on a temporary scarcity, grid connections, rather than a permanent one.
| Concern | Severity | Evidence and why it matters |
|---|---|---|
| Funding and dilution | HIGH | Shares up about 42% in a year to 394.06m; $4.7bn of equity sold in fiscal 2026. Fiscal 2027 capex of $25–30bn against about $14bn secured means billions more must be raised. The 0–1% converts are cheap in cash, but the six series convert into about 127m shares in total — about 30m from the deep-in-the-money 2029 and 2030 notes — partly offset by capped calls |
| GPU economic life | HIGH | GPUs are depreciated over five years (FY2026 10-K) while most new contracts run three. What happens in years four and five is the swing factor: holding the price constant, a six-year life raises our base case from $48.60 to about $91; a four-year life takes it below zero |
| Customer concentration and credit | MED-HIGH | Microsoft is about $1.94bn of the $4bn contracted 2026 run-rate (derived: $9.7bn over five years). Most of the rest is non-investment-grade AI developers. Prepayments and 90% financing in ring-fenced vehicles mitigate, but a single large default would dent the ramp |
| Governance and pay | MED-HIGH | Two B-class shares give the founders 15 votes per ordinary share held until about 2033; on our arithmetic from July 2026 Form 4 holdings that is a little over half the votes on a poll. The co-CEOs received 18.2m time-vested RSUs on 1 July 2026, worth about $790m at grant, with no performance hurdle |
| Metric quality | MEDIUM | Run-rate revenue is non-GAAP and the company states recognised revenue may be materially lower. Revenue has missed consensus repeatedly through the transition; first-quarter fiscal 2027 revenue consensus was cut about 44% in the three months to mid-August |
| Capital-structure complexity | MEDIUM | Six convertible series, capped calls, a 2029 prepaid forward, ring-fenced GPU financing entities with limited parent guarantees, and NVIDIA’s 30m-share right at $70. Each is disclosed; together they make per-share value hard to read |
| Execution and grid | MEDIUM | Roughly 500MW of IT load targeted in 2027 alone, with more than 4,000 workers on site. Texas large-load interconnection reforms and a state verification of data centres add schedule risk, though the 2GW Sweetwater hub was conditionally included in an ERCOT large-load process (8 September 2026 release) |
| Managed-service quality | MEDIUM | SemiAnalysis’s ClusterMAX 3.0 (23 September 2026) places IREN’s managed service in its Not Recommended – Underperforming category, citing reported power outages, storage and network failures and GPU errors at Prince George and Mackenzie in British Columbia; CoreWeave and Nebius are rated Platinum. The evidence came through resellers and customer feedback, not a full hands-on test, and Childress and Sweetwater appear better built. The anchor contracts are bare metal; the exposure is re-letting GPUs to smaller customers in years four and five |
| Legal and tax | LOW-MED | Two putative securities class actions — one from the 2021 IPO period, one filed in October 2024 after a July 2024 short report on Childress — remain unresolved; a Canadian GST and permanent-establishment dispute is before the Tax Court. None is quantified by the company |
| Accounting noise | LOW | Fiscal 2026 included $558.5m of unrealised gains on financial instruments, largely capped calls, and $638.8m of impairments. Both are disclosed and non-cash; neither drives the valuation |
| Insider selling | LOW | The only open-market sales we found were 1m shares by each co-CEO at $33.13 in September 2025. The new RSUs carry a two-year post-vesting holding period |
Overall red-flag score: 6 out of 10. The flags cluster around one theme: this is a business that must raise and spend very large sums before it earns them, and whose long-run returns depend on an asset life nobody has yet observed through a full cycle. There is no evidence of accounting manipulation; the concerns are structural.
| Strength | Weight | Evidence and why it matters |
|---|---|---|
| Pricing power, corroborated | HIGH | Three-year pricing up about 125% since November 2025 to above $20m per megawatt; Nebius reported the same range independently. At that price the compute pays back in about two years, so contracts signed in 2026–27 are profitable even if pricing later falls |
| Customer-funded build | HIGH | Prepayments of 45–55% of GPU capex plus 90% GPU financing mean the equity funds little of the compute. The entire data-centre estate is unencumbered, a financing source not yet tapped |
| Energised power | HIGH | Grid connections are the scarcest input in the industry. IREN has 750MW at Childress, an energised Sweetwater substation, British Columbia sites with spare headroom, and more than 5GW announced. Our base case monetises only about 3.3GW of IT load by fiscal 2040, and our bear case prices undeveloped power only as a floor |
| Counterparty validation | MED-HIGH | Microsoft accepted Horizon 1 in August 2026 and NVIDIA awarded Exemplar Cloud status on GB300 NVL72. NVIDIA is both a customer and a prospective shareholder. Independent testing is less favourable; see managed-service quality above |
| NVIDIA’s $70 right | MEDIUM | If exercised, it brings $2.1bn of equity at a 69% premium to the $41.30 price. We charge it as dilution in the bridge ($788m at the base value), not as a positive |
| Cheap capital already raised | MEDIUM | $6.75bn of converts at 0–3.5% coupons, $1.84bn of interest-free customer prepayment, and 6% investment-grade GPU debt. The cost of capital embedded in the existing book is below our 10.0% WACC |
| Short interest | LOW-MED | About 94m shares sold short, roughly 24% of shares outstanding (secondary data source). Not a fundamental, but a crowded short raises the chance of sharp moves on good news |
Overall green-flag score: 7 out of 10. The strongest positive is not a forecast: it is disclosed pricing, disclosed prepayments and delivered capacity. What remains unproven is duration: how long pricing stays above the level that merely earns the cost of capital.
| Dimension | Score / 5 | Assessment |
|---|---|---|
| CEO track record | 4 | Daniel and William Roberts founded the company in 2018, took it public in November 2021, survived the 2022 mining downturn and executed the move into AI cloud: Microsoft ($9.7bn), NVIDIA ($3.4bn plus a strategic partnership) and a frontier lab in under a year |
| CFO credibility | 3 | Anthony Lewis joined as Chief Capital Officer in July 2025 and succeeded Belinda Nucifora as chief financial officer. His financing record is strong — 6.0% investment-grade and 9.0% non-investment-grade GPU debt, 0–1% converts; his reporting record as CFO is short |
| Guidance accuracy | 3 | Run-rate targets have been met and raised ($3.7bn to more than $4bn in July 2026) and Horizon 1 was delivered and accepted in August 2026. Reported revenue has lagged the street through the transition, and the conversion from run-rate to revenue has been slower than headline targets suggest |
| Transparency | 3 | Good disclosure of contract terms, financing and pricing per megawatt. Heavy reliance on a non-GAAP run-rate metric; the frontier-lab customer is unnamed; no disclosure of 2027 pricing until contracts sign |
| Capital allocation | 3 | Raising low-coupon converts with capped calls and extracting customer prepayments is skilful. The December 2025 induced conversion of in-the-money 2029/2030 notes cost a $111.8m inducement expense and was funded by a 39.7m-share placement |
| Acquisitions | 3 | Mirantis (about 12m shares and ~$40m cash, closed 3 August 2026) and Nostrum (Spanish sites, ~490MW) are small and strategically sensible; too early to judge returns |
| Buybacks | n.a. | No buyback. Capped calls and a 2029 prepaid forward act as dilution hedges |
| Dilution | 2 | Share count up about 42% in a year; $4.7bn of equity sold in fiscal 2026; further issuance likely to fund fiscal 2027 |
| Insider ownership | 4 | Each co-CEO held about 14.5m ordinary shares plus 9.1m unvested RSUs after 1 July 2026 (Form 4). Economic stake roughly 7% combined; voting control a little over half through the B-class shares (derived) |
| Compensation | 2 | 18.2m time-vested RSUs, about $790m at grant, on 1 July 2026, with a six-year vest-and-hold and no further grants until fiscal 2031. The long holding period is a plus; the absence of performance conditions and the size relative to earnings are not |
| Board quality | 3 | Independent directors approved the grants unanimously with an outside consultant. The 2025 AGM added a Nasdaq-aligned quorum and a director-election structure, but the B-class nomination rights remain until about 2033 |
| Communication style | 3 | Promotional in framing, disciplined in the numbers; answers on funding are candid about dependence on market conditions |
| Average | 3.0 | Capable owner-operators with weak shareholder protections |
Do they act like long-term owners? Mostly, yes. The founders have sold almost nothing, the new awards cannot be monetised until fiscal 2029 at the earliest, and they have repeatedly chosen longer contracts and diversified customers over headline deal size. The counterweight is control: with a majority of votes on our arithmetic, outside shareholders cannot discipline pay or dilution decisions, and the 2026 award suggests the board will not do so on their behalf.
| Window | Event | Upside | Downside | Confidence · source |
|---|---|---|---|---|
| Oct–Nov 2026 | Hyperscaler September-quarter results and 2027 capex outlooks | Capex raised again, supporting pricing | Any “digestion” language | High · company calendars |
| Early Nov 2026 | Q1 FY27 results (consensus date about 4–5 November) | Run-rate of $1bn or more held; 2027 contracts named at $20m or more per megawatt | Cash SG&A above the +$40–50m guide; Horizon 2 slips | High · Q4 call; consensus services |
| Oct–Dec 2026 | Horizons 2–4 delivered to Microsoft | On time; run-rate towards $4bn | Delay credits; termination rights if dates are missed | Medium-high · FY26 release; Microsoft 8-K |
| 24 Nov 2026 | Annual general meeting, held virtually | Governance concessions | Votes on dilution authorities | High · FY26 Form 10-K |
| By Dec 2026 | Mining decommissioned | A cleaner profit and loss account | Further impairments | High · Q4 call |
| H2 2026–2027 | First data-centre financing | Low-cost asset debt against an unencumbered estate | Terms that signal tightening credit | Medium · Q4 call |
| Feb 2027 | Q2 FY27 results | Operating run-rate of $4bn or more at December | Run-rate below target; a longer revenue lag | High · pattern of prior years |
| Early 2027 | NVIDIA $3.4bn contract ramps at Childress | +$0.7bn of run-rate revenue | Slippage | Medium-high · 8 May 2026 release |
| H1 2027 | Contracts for 2027 capacity (~0.5GW) | $25m or more per megawatt with 50% prepayment | Pricing below $20m; lower prepayments | Medium · Q4 call (“late-stage”) |
| May 2027 | Q3 FY27 results | First full quarter at ~$4bn run-rate | Margin below about 70% | High · pattern |
| 1 Jul 2027 | First tranche of co-CEO RSUs vests, then a two-year hold | — | Dilution optics | High · 1 Jul 2026 8-K |
| Aug 2027 | FY27 results and the FY28 capex guide | Capex funded without large equity issuance | Guide raised with a funding gap | High · pattern |
| H2 2027 | Sweetwater 1 and Horizons 5–6 liquid-cooled capacity | On schedule | Texas grid reforms; delays | Medium · FY26 release |
| Open-dated | NVIDIA right to buy 30m shares at $70, to May 2031 | $2.1bn of equity at a premium | — | Low · 7 May 2026 release |
| Open-dated | Securities class actions; Canadian GST dispute | Dismissal | Settlement cost | Low · FY26 Form 10-K |
No dividend and no buyback programme. Macro items to watch: interest rates, since the sensitivity grid shows half a point moves value by $9–11 a share; the timing of NVIDIA’s next GPU generation, which reprices older fleets; and any change in US export controls affecting AI hardware. Confidence reflects the timing of the event, not the direction of its outcome.
BULLGrowth. $4bn of contracted run-rate revenue for 2026 is nearly six times fiscal 2026 revenue, and it is already signed. Demand exceeds supply at every neocloud, IREN has sold out 2026, and it is in late-stage talks for most of 2027. With 5GW of power, the runway is a decade.
BEARContracted run-rate is not revenue, and every quarter of the transition has shown the lag. Beyond 2026 the capacity is uncontracted, and the price it signs at is the whole question. Growth that earns only the cost of capital creates no value, and the reverse DCF shows the market is not paying for it either.
BULLValuation. At about 4.3 times year-end contracted run-rate revenue IREN is the cheapest neocloud, against about 8 for Nebius, and under 3 times our fiscal 2028 EBITDA. The probability-weighted value is $82.30, twice the price; sell-side targets averaged about $79 on 21 September. The shares fell 15% in a week on a third-party service rating and two cautious initiations, with no change in any contract.
BEARThose multiples ignore that IREN must spend $15–30bn a year for years. On a proper cash-flow basis the modal value is $48.60 only at a 10.0% discount rate, which is low for a levered builder of depreciating hardware when ten-year Treasuries yield almost 5%. At 10.5% it is $39.15, below the price. The market is simply using the fuller rate.
BULLBusiness quality. Owning the power and the buildings removes rent and captures an 85% project margin. Grid connections are the scarcest input in the industry and IREN has them. The managed-software layer that third-party ratings test matters little to Microsoft and NVIDIA, which buy bare metal and run their own.
BEARGPUs are a commodity anyone can finance. The scarcity of power is temporary, and the product, a three-year GPU rental, reprices every contract cycle. And the service is weak: SemiAnalysis rates it Underperforming, citing outages and GPU errors in British Columbia, while CoreWeave and Nebius are Platinum. The customers who rent three-year-old GPUs are exactly the ones who read that rating.
BULLRisks. Customers pay half the GPU cost up front and lenders fund 90%; losses on a bad cohort sit largely with ring-fenced lenders, not the parent’s equity.
BEARIn a pricing bust the committed fiscal 2027–28 build earns below the cost of capital and the model equity goes negative. Customers outside Microsoft and NVIDIA are mostly loss-making start-ups.
BULLFinancials. $7.6bn of cash, 0–1% convertible coupons, $14bn secured for fiscal 2027, and an unencumbered data-centre estate still to be financed.
BEARFiscal 2027 free cash flow is deeply negative on any assumption — consensus about −$21bn. The funding gap will be closed partly with equity, and the share count already rose 42% in a year.
BULLManagement. Founders who called the pivot early, delivered on schedule and financed it brilliantly.
BEARFounders who control the vote and gave themselves $790m of unconditional stock.
BULLCatalysts. Horizons 2–4, the NVIDIA ramp and 2027 contracts at $25m per megawatt all land in the next nine months.
BEAREach is a binary delivery test, and the first data-centre financing and the next equity raise will show what capital really costs.
Near term, the bull has the better evidence: pricing, prepayments and delivery are disclosed facts, not forecasts, and they make the 2026–27 cohort profitable almost regardless of what follows. Long term, the bear has the better argument: nothing yet shows how a GPU fleet earns in years four and five, and IREN’s value is more sensitive to that than to anything else. Net, the price sits 15% below our modal value but inside the band, and the gap is almost exactly the difference between our 10.0% discount rate and the 10.4% our inputs support with a fuller equity premium, so neither side has an edge large enough to act on. Most uncertain: GPU economic life, the price of 2027 capacity, and the terms of the first data-centre financing. Verify next: operating run-rate at the first and second quarters of fiscal 2027, disclosed price per megawatt on 2027 contracts, re-contracting rates on expiring older GPUs across the industry, and the share count after fiscal 2027 funding. The committee’s ruling is that the shares are fairly valued, near the lower edge of the band, with a payoff skewed strongly upwards and a real tail below it.
IREN owns big buildings full of specialised computer chips, NVIDIA GPUs, in places with cheap electricity, mostly Texas and British Columbia. It rents the computing power of those chips to companies that build artificial-intelligence products, such as Microsoft and NVIDIA itself. It used to use the same buildings to mine Bitcoin and is switching that off by the end of 2026.
Customers sign contracts, usually for three to five years, and pay by the hour for the chips they reserve. Many now pay a large part in advance. IREN pays for the chips and the buildings up front and earns the money back over the contract.
Demand for AI computing is growing faster than the world can build power-connected buildings. IREN already has the power connections, which are the hardest part, and has signed contracts worth about $4bn a year for capacity coming online in 2026 — against $0.7bn of total revenue last year.
A GPU is paid for up front and earns rent until it is too old to be wanted. IREN’s accounts assume five years. Most contracts last three. If the chips keep earning for six years, the business is worth far more than today’s price; if only four, it is worth far less. Nobody has watched a fleet like this through a whole life yet, and that is most of the uncertainty in this stock.
Our estimate of what the business is worth is $38.90–60.95 a share, most likely about $48.60, once you allow for the huge spending still to come. At $41.30 the shares cost about 15% less than our central estimate and sit near the bottom of that range, so we call them fairly valued. If things go well they are cheap; if they go badly, shareholders could lose most of their investment.
Right: the new buildings open on time, 2027’s capacity is rented at today’s high prices, customers keep paying in advance, and the chips keep earning for longer than five years. In that world the shares could be worth several times today’s price.
Wrong: rental prices for AI chips fall as more supply arrives, as they did in 2024–25; the company has to raise tens of billions on bad terms; a big customer fails; or smaller customers avoid IREN because of its service rating. The founders also control the vote.
The results in early November, and in particular whether new contracts for 2027 are signed at $20m or more per megawatt a year. Whether Microsoft’s remaining buildings go live by December. And how the next round of money is raised, and at what price.
| Status | What it covers |
|---|---|
| Verified from IREN filings and releases | Revenue by stream, adjusted EBITDA, net income, impairments, share-based compensation, cash flow, capex, cash and restricted cash, deferred revenue, convertible note terms, GPU financing and finance leases, share counts, contract values, pricing per megawatt, prepayment ranges, financing terms, capacity targets, guidance, the co-CEO equity grants and the B-class share terms — from the FY2026 results release and Form 10-K (27 August 2026), the fourth-quarter conference call (27 August 2026), the third-quarter results and NVIDIA contract and partnership releases (7 May 2026), the $2.8bn contracts and run-rate target release (20 July 2026), the Microsoft contract release and Form 8-K (3 November 2025), the convertible note releases (11 and 13 June 2025; 14 October 2025; 3 and 8 December 2025; 12 and 14 May 2026), the Form 8-K on the co-CEO grants and Forms 4 (1 July 2026), the Form 8-K on the Mirantis resale registration (August 2026) and the ERCOT release (8 September 2026). |
| Verified from market data | The $41.30 intraday price of 29 September 2026, taken at 13:37 New York time, and peer prices taken in the same minute; the 22 September 2026 close of $48.55 and the daily close series behind the 52-week high; share counts from each company’s latest filing cover; the ten-year Treasury yield (21 September 2026). Peer operating figures from CoreWeave’s second-quarter 2026 results (11 August 2026), Nebius’s second-quarter shareholder letter and Form 6-K (12 August 2026) and Applied Digital’s fiscal 2026 results (27 July 2026). |
| Third-party, attributed | SemiAnalysis, ClusterMAX 3.0 (23 September 2026), and Blockspace’s coverage of it: the site-level findings on IREN rest on resellers and customer feedback rather than a full hands-on test and are not corroborated. Consensus estimates and price targets from S&P Global Market Intelligence via StockAnalysis (21 September 2026). Short interest from a secondary data service. The Rothschild & Co Redburn (23 September 2026) and Jones Trading (28 September 2026) initiations as reported in the financial press. |
| Our estimate or judgement | Every scenario input in Section 03; Microsoft and NVIDIA revenue per megawatt; AI cloud cash gross margin; operating IT load at 30 June 2026 (about 40MW); the opening split of plant and equipment and the cash-tax path; the founders’ voting share; enterprise value per megawatt and every multiple, with CoreWeave and Nebius enterprise values approximate; the fundamental beta and the 10.0% rate; the 25 / 45 / 25 / 5 weights and the $4.00 bear floor; the sensitivity grid, the reverse DCF, the $38.90–60.95 band and the $48.60 central estimate; and all scores in Sections 06, 07 and 08. The model on this page is the model that produced those figures — the scenario cards load the published paths and the sensitivity grid is reproducible from the base settings. |
| Still open | The identity and terms of the frontier-lab contract; the pricing of 2027 capacity; how GPU fleets re-contract after their first term. The $4.7bn “projected adjusted net income” figure attributed to IREN in some press coverage of the July 2026 awards could not be traced to a filing and is not used. |
Method. Unlevered discounted cash flow, methodology version 1.3, built as a capacity model for a contracted-infrastructure owner: fourteen explicit fiscal years, 2027–2040, valued as at 23 September 2026 with the 30 June 2026 balance sheet opening fiscal 2027 and a mid-year convention, and a value-driver terminal value at 3.0% growth and a 12% return on new capital. Free cash flow to the firm is adjusted EBITDA less share-based compensation, cash tax and capex including GPU refresh, plus working capital, which carries customer prepayments and their credit-back. Convertibles are counted as converted above their conversion prices, with capped calls credited to the cap, and at face below; NVIDIA’s right is valued by Black-Scholes at the model-implied price. One discount rate and one terminal growth rate are used across all scenarios. Scenario weights are owned by the analyst, the fair-value band is read from the sensitivity grid, and the reverse DCF is run against the market price. The rating changes to undervalued if the price falls below $38.90 with no change in the evidence, and to overvalued above $60.95; the conclusion is most sensitive to GPU economic life, the discount rate and the long-run price per megawatt, in that order. Ratings are undervalued, fairly valued or overvalued according to where the price sits relative to the band; Sonde does not issue buy, sell or hold recommendations or price targets. The reference price used throughout is the $41.30 intraday quote of 29 September 2026, and every multiple in this report scales directly with it.