$9.7B Microsoft Contract $2.8B New AI Developer Contracts Mirantis Closed — Aug 3 $3.4B NVIDIA AI Cloud Contract 5GW Secured Power — 3 Continents $4B+ ARR Target — ~85% Contracted $3.65B Investment-Grade Financing — A / A(low) 800MW South Australia 490MW Spain — Nostrum Closed Added to Russell 1000 BIP Playbook — 18% Annualised Returns $9.7B Microsoft Contract $2.8B New AI Developer Contracts Mirantis Closed — Aug 3 $3.4B NVIDIA AI Cloud Contract 5GW Secured Power — 3 Continents $4B+ ARR Target — ~85% Contracted $3.65B Investment-Grade Financing — A / A(low) 800MW South Australia 490MW Spain — Nostrum Closed Added to Russell 1000 BIP Playbook — 18% Annualised Returns
Updated Thesis — August 2026

THE AI POWER PLAY OF THE DECADE

IREN built the scarcest moat in AI infrastructure: 5GW of secured renewable power across three continents, purpose-built GPU data centres, and $15.9B of signed contracts — Microsoft, NVIDIA and ten AI developers. Credit markets have already re-rated it — rating agencies stamped the GPU facility investment-grade. Equity markets still price it like a bitcoin miner. That gap is the thesis.

Read the Thesis Explore the Sites
$15.9B
Total Contracts Signed
5 GW
Secured Power
$4B+
YE2026 ARR Target — ~85% Contracted
$9.7B
Microsoft Contract
$3.4B
NVIDIA Contract
5 GW
Secured Power
$2.8B
New Contracts — Jul 20
$4B+
YE2026 ARR Target
A / A(low)
GPU Facility Rating
480MW
Operating by YE2026
The Bull Case

WHY IREN WINS

Six structural advantages in AI infrastructure — none fully priced by the equity market, though the credit market has already begun to price them.

Secured Power Is the Scarcest Asset
36 GW US power shortfall by 2028. 18-year average transmission build times. IREN's 5GW of secured, grid-connected power — now spanning North America, Spain and South Australia — cannot be replicated on any timeline that matters. This is the moat. Not the NVIDIA logo.
🤝
$15.9B Contracted — And No Longer Two Names
Microsoft: $9.7B, 5 years, 76,000+ GB300s across Horizons 1–4, 20% prepaid. NVIDIA: $3.4B, 5 years, air-cooled Blackwell inside existing Childress capacity, ramping early 2027. Plus $2.8B signed July 20 across eight AI developers — Perplexity, Figure AI, Together AI, Fluidstack, Fireworks AI, Fal AI, Hume AI and one unnamed. Year-end ARR target lifted to $4B+, ~85% under contract.
🔬
Air-Cooled Childress Discovery
Dan Roberts shelved liquid-cooled conversion. Air-cooled retrofits across 18 buildings. ~$4M/MW vs. $15M/MW liquid-cooled. 2028 net income ~30% above prior model.
💰
Credit Re-Rated Before Equity
$3.65B investment-grade GPU facility — rated A / A(low), arranged by Goldman and J.P. Morgan against the Microsoft offtake. Plus $3.0B converts at 1.00%. The rating agencies underwrote these cash flows a year before the equity market will. That sequence is the BIP analogue.
🏗️
Execution Track Record — Now Being Tested
Sweetwater 1 substation energized on schedule. 5 EH/s to 50 EH/s in two years. But Horizon 1 is the live credibility variable — the liquid-cooled build is the hardest thing IREN has attempted, and September is where the claim gets marked to market.
🌊
Full Stack — Power to Orchestration
IREN owns land + power + datacenter, enabling FMV GPU leasing — cycling into best silicon each generation. Mirantis closed August 3 — adding workload orchestration, monitoring and support on top, plus 1,500+ enterprise customers. Not a new strategic vector: a margin enabler that moves IREN from landlord to full-stack neocloud.
Asset Portfolio

THE POWER MAP

5GW of secured, grid-connected power across three continents — North America, Europe and APAC. Two new regions added since April: Spain via Nostrum (closed June 15) and an 800MW campus in South Australia (transmission agreement signed June 3).

Site Legend
BC Canada (×3)
160MW · Operational
Childress, TX
750MW · Live + MSFT
Sweetwater 1+2, TX
2,000MW · Energizing
Oklahoma — "Kiowa"
1,600MW · 2028 Ramp
Spain — Nostrum CLOSED
490MW · GW+ Pipeline
Bundey, South Australia NEW
800MW · 2028 Energization
5 GW
3 CONTINENTS
🏔️
● Live
Prince George
📍 British Columbia, Canada
50MW
Capacity
~23k
B300 GPUs
1.1
PUE
Hydro-powered flagship GPU site. 100% renewable. Air-cooled at industry-leading 80kW rack density. At maximum capacity with $500M+ ARR. The proof-of-concept for everything that follows.
🌲
● Live — Expanding
Mackenzie
📍 British Columbia, Canada
80→100MW
Capacity
~33k
B300 GPUs
~$785M
ARR Potential
Expansion from 80MW to 100MW confirmed via podcast — not yet in official guidance. 85%+ EBITDA margins on stranded hydro at ~$700K/MW capex.
⚡ Expansion not priced in by any sell-side model
💧
● Live
Canal Flats
📍 British Columbia, Canada
30MW
Capacity
~12k
B300 GPUs
<$0.03
$/kWh
IREN's original site. Canadian Rockies. Additional expansion headroom flagged by management. Ultra-low hydro cost creates structurally highest-margin segment.
🏭
● Live + MSFT
Childress, Texas
📍 Childress County — West ERCOT
750MW
Total
480MW
Live by YE26
60MW
NVIDIA Air-Cool
The crown jewel — and now the home of both anchor contracts. 576-acre freehold, 345kV ERCOT. Horizons 1–4 (liquid-cooled) fully contracted to Microsoft at $9.7B, on track for 480MW by year-end. NVIDIA's separate $3.4B contract deploys air-cooled Blackwell inside existing capacity. Horizons 5–6 follow in the 2027 build.
⚠️ Horizon 1 is the live credibility test. No confirmed handover as of Aug 6. Watch DDTL draw cadence, not social-media delivery claims.
● Substation Energized
Sweetwater Hub
📍 Fisher County — West ERCOT
2,000MW
Total Hub
1,800 ac
Land
<10ms
Latency
Sweetwater 1 substation energized on schedule — management delivered the date. Initial Sweetwater 1 phase sits in the 2027 build (1,210MW total with Childress H5–6). Sweetwater 2 ramps 2028+. 700,000+ liquid-cooled GPU runway. Designed for Vera Rubin. ERCOT Batch Zero.
🚀 Still uncontracted. The next hyperscaler deal most likely lands here. 40 miles from OpenAI's 1GW Stargate campus.
🌾
◆ Pipeline
Oklahoma — "Kiowa"
📍 Oklahoma — SPP Grid
1,600MW
Capacity
2,000 ac
Land
SPP
Grid
Diversifies away from ERCOT. Google and Meta already in Oklahoma. IREN's Q3 FY26 filing names Sweetwater and Kiowa as the 2028+ capacity ramp. Dangubic's institutional co-investment playbook is purpose-built for funding this without diluting shareholders.
💡 Watch for a co-investment SPV structure here — sovereign capital at asset level, IREN retains control + promote.
🇪🇸
● Closed June 15, 2026
Spain — Nostrum
📍 Spain — Iberian Grid · Europe Entry
490MW
Secured
GW+
Pipeline
50+
Local Team
Ingenostrum, S.L. — founded 2009 as a renewables engineering firm, rebranded 2025 with data centres as the core business. IREN bought an operating team that already holds secured power and understands Spanish permitting — not a greenfield entry. Abundant renewables, strong fibre, constructive AI policy.
🌍 First European footprint. Operating under the IREN brand. Terms not disclosed.
🇦🇺
◆ New — June 3, 2026
Bundey, South Australia
📍 ~78 miles NE of Adelaide — APAC Entry
800MW
Planned
4 × 330kV
Feeder Exits
2028
Energization
Transmission connection agreement signed — four 330kV feeder exits at the utility substation supporting up to 800MW without requiring network upgrades. That clause is the whole point: it removes the single longest pole in the tent. SA grid targets 100% net renewable by 2027. Submarine fibre into Singapore, Indonesia, South Korea and Japan.
🏠 The Sydney-headquartered company finally plants a flag at home. ~200 ongoing jobs, 500+ in construction.
📶 THE CAPACITY LADDER — PER Q3 FY26 FILING
Management now discloses the build as a dated ladder rather than an aspiration. Each rung is a checkable claim — which is exactly what makes the next four earnings prints binding.
480MW
2026 — Operational
Horizons 1–4 at Childress, delivery by year-end. Operational capacity fully contracted. Year-end CY26 ARR target raised July 20 from $3.7B to more than $4B, of which ~85% is under contract. NVIDIA’s Blackwell deployment ramps on top from early 2027.
1,210MW
2027 — In Build
Childress Horizons 5–6, Childress air-cooled capacity, and the initial Sweetwater 1 phase. NVIDIA's $3.4B contract ramps from early 2027 inside this window.
5 GW
2028+ — Secured Power
Sweetwater and Kiowa ramp from 2028. Nostrum adds 490MW in Spain plus a GW+ pipeline. Bundey (800MW) energizes from 2028 under a signed transmission agreement.
Note — Project Iris Monarch retired from this deck. The April version of this page carried a suspected ~1GW Childress asset sourced from county land records and satellite shape-matching. That inference has been overtaken by disclosure: IREN now names its pipeline formally at 5GW across Childress, Sweetwater, Kiowa, Spain and Bundey. The land-records method also proved to be the kind of analysis that produces confirmation loops without touching the variables that actually move the price — contracted revenue, hyperscaler announcements, and management disclosure. Cut, not because it was wrong, but because it was never load-bearing.
The Anchor Contracts
$15.9B
Two Anchors + Ten AI Developers

Microsoft — $9.7B. Signed November 3, 2025. 5-year term. 76,000+ NVIDIA GB300 GPUs across Horizons 1–4 at Childress, liquid-cooled, 480MW by year-end 2026. 20% prepayment ($1.93B) received.

NVIDIA — $3.4B. Signed May 7, 2026. 5-year AI Cloud contract for air-cooled Blackwell inside 60MW of existing Childress capacity — no new build required. Ramping from early 2027.

AI developers — $2.8B. Signed July 20, 2026. Eight new multi-year cloud contracts across bare metal and managed services. ~45% of the associated GPU capex prepaid by customers; ~4-year weighted average term across the portfolio.

Capital Structure
A / A(low)
GPU Facility Rating
$3.65B
IG Facility — Closed Jun 1
1.00%
$3.0B Convert Coupon
🏛️
Credit Re-Rated First — A / A(low)$3.65B investment-grade GPU facility closed June 1, arranged by Goldman and J.P. Morgan against the Microsoft offtake. Rating agencies underwrote these cash flows before the equity market would. This is the leading indicator, not a footnote.
🤝
NVIDIA — 5GW Strategic PartnershipBeyond the $3.4B contract: collaboration across IREN's full 5GW pipeline. NVIDIA received a 5-year right to buy up to 30M ordinary shares at $70 strike — up to $2.1B invested, subject to conditions including regulatory.
⚖️
But: NVIDIA Ties Aren't DifferentiatingCoreWeave, Lambda, Nebius, Crusoe and others hold materially similar equity-linked NVIDIA partnerships. The warrant is validation, not a moat. Secured power and the operational stack are the moat.
💵
$3.0B Converts at 1.00%32.5% conversion premium; $201.3M capped call pushes the effective cap to ~$110.30 (~80% effective premium). Materially cheaper all-in than Nebius's raise.
🧩
Mirantis Closed — August 3, 2026Aggregate consideration of 13,673,894 ordinary shares priced off the $45.66 May 1 close (~$625M), plus ~$40M of cash, RSUs and other consideration. Operates as a standalone subsidiary. Inaugural partner in the NVIDIA AI Cloud Ready Initiative; k0rdent AI integrates with NVIDIA DSX OS. Dilution is real: ~12.6M new shares, with 11,981,668 registered for resale.
💧
$7.6B Cash — June 30, 2026Includes $1.7B restricted against the Horizon 1–4 GPU financing. Customer prepayments now cover ~45% of GPU capex on recent contracts, cutting the net funding requirement on those deployments.
📡
DDTL Draw Cadence — The Real-Time SignalBetween prints, the delayed-draw term loan cadence is the best available proxy for Horizon handoff. More actionable than management commentary.
The Concentration Critique — Partially Answered
Through June the strongest bear argument was not power, financing or execution. It was that IREN had one real customer. Bernstein said it in print: behind peers on building enterprise demand outside Microsoft. That critique was correct, and it is on this page in the rating box below.

Two events since have moved it. On July 20 IREN signed $2.8B across eight AI developers and lifted the year-end ARR target past $4B with ~85% under contract. On August 3 it closed Mirantis — the managed-services and orchestration layer, with 1,500+ enterprise customers attached. Management's own framing is worth noting: the acquisition already facilitated several of the announced and prospective contracts. The software layer was not bought after the demand; it was part of winning it.

What this is not. Eight AI developers are not a hyperscaler. Perplexity and Figure AI do not underwrite Sweetwater. The add trigger on this page is a third anchor — Meta, Oracle, Amazon, Google scale — and it has not fired. Counting the July cohort as that trigger would be a rule change dressed as good news. It is best read as a gate cleared, not a trigger pulled: the concentration objection is now weaker, which is what has to be true before the anchor matters.
The Dream Team

ASSEMBLED FOR A $50–100B COMPANY

Twelve months. Eight senior hires. A CFO from 22 years at Macquarie. A Chief Innovation Officer who writes the global liquid-cooling standards. A Head of BD from CVC and KKR. And since April: product from Oracle Cloud, development from Google, security from Nutanix and NVIDIA, hyperscaler BD from AWS. The pattern widened from infrastructure finance to cloud operators — which is exactly what you'd expect from a company that now has to run the thing it financed.

The Macquarie pattern
built the balance sheet.
The cloud pattern
has to deliver the product.
👤
Daniel Roberts
Co-Founder & Co-CEO
Founded 2018 · Since inception
Co-owner, Palisade Investment Partners ($6B infra manager)
Macquarie Group — infrastructure finance
University of Western Australia
Why It Matters~14M ordinary shares via Awassi Capital trusts plus 21.8% Class B voting power. Granted a further 9,099,328 RSUs on July 1, 2026 — time-vesting over four years, each tranche locked a further two years, final tranche unsellable until FY2033. No further equity grants until FY2031. See Governance →
👤
Will Roberts
Co-Founder & Co-CEO
Founded 2018 · Since inception
Vice President, Macquarie Group — infrastructure & finance
University of Western Australia
Why It MattersSame structure — ~14M shares, 21.8% Class B, and an identical 9,099,328 RSU grant. VP at Macquarie gave him the toolkit to structure $9.7B offtakes and investment-grade GPU facilities. The playbook was already in his head before the sector existed.
💼
Anthony Lewis
Chief Financial Officer
Joined July 2025 → CFO September 2025
22 years at Macquarie Group
Co-Treasurer — global funding, liquidity & capital management
Regulatory reporting across all jurisdictions
Why It MattersDelivered the $3.65B facility at investment grade — A / A(low), plus $3.0B of converts at a 1.00% coupon. Getting a GPU-backed facility rated single-A is the single most underrated event of the past year: it means the offtake is being treated as utility-grade paper. Architect of retrospective DC financing ($1–2B non-dilutive).
🔧
John Gross
Chief Innovation Officer
Appointed February 17, 2026
Vice Chair, ASHRAE — writes global liquid-cooling standards
CTO, Prometheus Hyperscale
20+ years data center engineering · PE licensed
Why It MattersThe man who writes the industry's cooling standards now designs IREN's next-gen data centers. Entire engineering firm joined with him. Directly relevant to the Horizon 1 liquid-cooled build — and to why IREN's air-cooled fallback at Childress exists at all when OpenAI's Abilene expansion stalled on liquid-cooling reliability.
🤝
Boris Dangubic
Head of BD, Strategy & M&A
Joined ~March 2026 · Head of BD, Strategy & M&A
Senior Managing Director, CVC Capital Partners (€205B AUM)
Director Private Equity, KKR (Latitude Financial deal)
Deutsche Bank IBD, Sydney · UWA — same as Roberts brothers
Why It MattersKKR + CVC background is the exact Brookfield co-investment playbook: sovereign wealth at SPV level, IREN retains control and promoted interest. Since April he has had a busy year — Mirantis (~$625M, closed Aug 3), Nostrum (~€165M), Awaken. Owns the structures that fund Sweetwater, Kiowa and Bundey without diluting shareholders.
📡
Kent Draper
Chief Commercial Officer
Active through all recent earnings calls
Deep data center commercial background
GPU cloud (CSPaaS) deal structuring
Why It MattersGPU cloud has "particularly compelling returns" with 2–3 year ROI. "Time-to-data-center is the key decision point." Owns the origination funnel for the third anchor customer — the one catalyst that would justify adding to the position.
🧭
Kambiz Aghili
Chief Product Officer
Announced July 2, 2026 · San Francisco
Oracle Cloud Infrastructure
Leads product strategy for IREN's AI Cloud platform
Why It MattersIREN has never had a CPO. Hiring one — from OCI, based in SF, not Sydney — says the AI Cloud is being built as a product with a roadmap, not a colocation contract with GPUs in it. This is the direct answer to the Bernstein critique that IREN is behind CoreWeave and Nebius on enterprise cloud.
🏗️
Michael Nudelman
Chief Development Officer
Announced July 2, 2026 · San Francisco
Google — data center development
CyrusOne · Beale Infrastructure
Why It MattersDirects global data center development and expansion of the 5GW portfolio. A Google + CyrusOne background is the counterparty-side view of exactly what a hyperscaler needs to see before signing. Hired the same month the pipeline went from two countries to four.
🔐
Eric Hammersley
Chief Information Security Officer
Announced July 15, 2026 — newest hire
VP Engineering & Chief Product Security Officer, Nutanix
Software product security architecture, NVIDIA HPC
Chief Engineer supporting the Joint Chiefs of Staff (J6) · US Navy veteran
Why It MattersSecurity across all three platform layers: data centers, compute, software. Read it as procurement plumbing — hyperscalers and frontier labs run security diligence before they sign, and a named CISO with NVIDIA and federal credentials is a box that has to be ticked. Hiring a CISO is what a company does when it expects more counterparties, not fewer.
🏦
Ryan Boers
From Quadrant Private Equity — Role Not Yet Disclosed
Reported August 2026 · AFR · Not yet announced by IREN
Quadrant Private Equity — Partner, Buyout
KKR Sydney — Principal, 2 years
Archer Capital — 7 years
Why It MattersFifteen years of Australian control-deal experience — originate, execute, manage, exit. Two readings, and they are not the same trade. Either this is asset-level co-investment structuring for Sweetwater, Kiowa, Bundey and Spain, which is precisely the Brookfield playbook this page is built on. Or it is more M&A capacity at a company that has closed three acquisitions in twelve months, partly in stock. The first is the thesis executing. The second is where capital-allocation discipline becomes the question. Reported by the AFR from an LP communication; Quadrant’s site still lists him. No IREN release.
☁️
Christopher Sailer
Hyperscaler Business Development
Surfaced May 2026 · Social-media sourced
AWS — corporate deal strategy
Worked on AWS × Anthropic partnership
Multi-year AWS × OpenAI partnership
Why It MattersThe most thesis-relevant hire on this page. Sailer sat on the counterparty side of AI-lab compute deals — he knows how these get structured because he structured them. IREN is building in-house origination rather than renting it from banks. Caveat: sourced from social media, not a filing or press release. Treat accordingly.
🔗 The Network — Widened
Phase 1 — the balance sheet (2025→early 2026). Dan Roberts (Macquarie infra) → Will Roberts (Macquarie VP) → Anthony Lewis (22 years Macquarie, Co-Treasurer) → Boris Dangubic (Deutsche Bank / KKR / CVC, same UWA as the Roberts brothers). That cluster produced a $9.7B offtake, an investment-grade GPU facility and $3.0B of cheap converts.

Phase 2 — the product (mid-2026). Christopher Sailer (AWS deal strategy) → Kambiz Aghili (Oracle Cloud Infrastructure) → Michael Nudelman (Google, CyrusOne) → Eric Hammersley (Nutanix, NVIDIA, Joint Chiefs J6). Not one Macquarie name among them. Cloud operators, hired to answer how do we run it and sell it.

Correction — this is not a rotation. An earlier version of this page argued the hiring signal had moved on from balance sheet to product. Ryan Boers breaks that reading: a Quadrant buyout partner, ex-KKR Sydney, reported in August. The capital track never closed — it widened alongside the product track. Which is the more demanding interpretation, not the more comfortable one. A company that has finished financing hires operators. A company still hiring principal-investment partners is telling you the capital structure is not finished — Sweetwater, Kiowa, Bundey and Spain are unfunded, and the Brookfield playbook this page rests on requires exactly this skill set to bring institutional money in at asset level while retaining control and promote. Read Boers as confirmation that management knows what still has to be built, not as evidence it is built.
Unit Economics

WHY THE NUMBERS WORK

Levered Cloud IRR
~400%
Microsoft GPU Deployment
$260M equity → $5.8B GPU capex → $1.94B ARR annually
DC capex sunk. Every future cycle compounds free.
$4M
Air-cooled CapEx/MW
vs $15M liquid-cooled
38mo
EBITDA Payback
vs 42mo Horizons
Peer Comparison — Capital Cost
CompanyGPU FinancingGPU ModelDC Ownership
IRENIG — A / A(low) FMV Lease ✓Freehold ✓
CoreWeave~10%Owned (risk)Leased ✗
NebiusNVIDIA-backedOwnedLeased ✗
Applied Digital>10%MixedMixed
The FMV Lease Moat IREN owns datacenter + power + land. That enables FMV GPU leasing — cycling into best hardware each generation. CoreWeave and Nebius own GPUs but lease datacenters. They cannot replicate this structure.
The Mix Shift

BITCOIN OFF, AI ON

Three quarters of reported data and a model of the next six. Bitcoin mining revenue has already fallen 52% from its peak — deliberately, by decommissioning ASICs to free capacity for GPUs. The question this chart answers is when the mix flips. The question it can't answer is whether the flip is what re-rates the stock.

REPORTED MODELLED 0 200 400 600 800 REVENUE $M / QTR 0% 25% 50% 75% 100% BITCOIN % OF REVENUE 240 Q1 FY26 Sep-25 NOV 2025 185 Q2 FY26 Dec-25 FEB 2026 145 Q3 FY26 Mar-26 MAY 2026 137 Q4 FY26 Jun-26 SEP 2026 165 Q1 FY27 Sep-26 NOV 2026 327 Q2 FY27 Dec-26 FEB 2027 573 Q3 FY27 Mar-27 MAY 2027 708 Q4 FY27 Jun-27 AUG 2027 785 Q1 FY28 Sep-27 NOV 2027 97% 91% 77% 60% 36% 13% 5% 3% 1% MAXIMUM-PAIN PRINT Revenue troughs · Sep 16 CROSSOVER · BTC <50% Nov 2026 — the thesis print AI Cloud revenue Bitcoin mining revenue Bitcoin % of total (right axis) Modelled — not company guidance
Sept 16, 2026 — Q4 FY26
~$137M
Revenue Trough · BTC ~60%
Third consecutive revenue decline. GAAP loss still carrying ASIC decommissioning impairments. Horizon 1 barely billing. The model says this print should look terrible.
Nov 2026 — Q1 FY27
36%
Crossover · AI Overtakes BTC
AI Cloud passes Bitcoin for the first time in company history. Revenue inflects back up. The same print this deck already called the thesis validation event.
May 2027 — Q3 FY27
<5%
Bitcoin Becomes a Rounding Error
All four Horizons plus NVIDIA live. ~$573M in the quarter. At this point no honest analyst can call IREN a miner — and the comp set has to change.
How The Model Is Built Reported: Q1 FY26 $240.3M ($232.9M mining / $7.3M AI) → Q2 FY26 $184.7M ($167.4M / $17.3M) → Q3 FY26 $144.8M ($111.2M / $33.6M). Mining down 28% then 34% QoQ; AI roughly doubling each quarter.

Mining ramp-off holds a 26–36% quarterly decay — the observed rate. It is not a Bitcoin-price forecast; the decline is mostly deliberate, and the $140.4M of non-cash impairment IREN booked in Q3 FY26 on retired mining hardware is the receipt. BTC at ~$60K makes it faster, not different.

AI ramp-on is built bottom-up from the contracts, not from a growth rate: Prince George ($0.5bn ARR under contract ≈ $125M/qtr at capacity, ramping now) + Microsoft ($1.9bn/yr average ≈ $475M/qtr, Horizons 1–4 to 480MW by end CY26) + NVIDIA ($0.7bn/yr ≈ $175M/qtr, ramping from early CY27).

Internal check: the model's Q1 FY28 AI quarter of $775M annualises to $3.10bn — the $3.1bn ARR currently under contract, to the decimal. That is a consistency check, not a validation. Any bottom-up build from the same contracts had to land there. The real uncertainty isn't the destination, it's the slope — and the slope is entirely Horizon delivery timing.
Does A Low Bitcoin % Actually Cause The Re-Rate?
Probably not — and the chart above quietly admits it. Look at what drives the orange line down. Mining falls from $233M to $10M, which is real. But total revenue goes from $240M to $785M. The percentage collapses mostly because the denominator explodes. A chart of "Bitcoin %" is a chart of "AI Cloud revenue" wearing a hat. It re-encodes information the deck already has; it doesn't add any. Something that carries no independent information can't be an independent cause.

The record cuts against the mix trigger. IREN's mining share has already fallen from 97% to 77% in two quarters — and the stock is down ~40% in a month. Meanwhile TeraWulf, still essentially a pure Bitcoin miner doing $34M a quarter, re-rated 13–19% intraday on July 6 on the Anthropic lease announcement: zero mix change, zero revenue, immediate re-rate. IREN's own biggest moves were November 2025 (Microsoft signed) and May 2026 (NVIDIA). Markets re-rate on contracts and cash-flow prints, not on mix percentages.

But the steelman is real, and it's about plumbing rather than perception. There are four mechanisms where mix genuinely bites, and they're worth separating from the loose version of the argument:

1. Classification and index baskets. IREN is still bucketed as a crypto miner — Finviz has it under Financial — Capital Markets. GICS sub-industry reclassification toward IT/data-centers is a mechanical flow event, and classification committees key off revenue source. This is the strongest form of the argument precisely because it isn't a restatement of AI revenue: it's a threshold effect, with a committee, on a date. Russell 1000 (June 26) was step one. 2. The Bitcoin-beta discount loses its numerator. Part of the miner multiple is rational commodity exposure. At 77% mining revenue, co-moving with BTC is correct pricing. Below 5% there is nothing left to co-move with. That's de-rating the discount, not re-rating the asset — a different and more defensible claim. 3. Coverage migrates desks. IREN is covered by crypto analysts. When coverage moves to data-center and infrastructure desks, the comp set in the models changes from MARA and CLSK to DLR and EQIX. The multiple partly is the comp set. 4. The impairment overhang ends. Q3 FY26's $247.8M net loss was largely $140.4M of mining decommissioning. When the ASICs are gone, the headline GAAP loss stops being a wind-down artifact — which matters for the screens that never read past the first line. So: mix is a gate, not a trigger. Contracted cash flow is the cause. The mix collapse is a symptom of that same cause — and it happens to unlock the classification and comp-set changes that let the multiple actually move. Necessary, not sufficient. The causation runs AI revenue → mix, never the reverse. That the crossover lands on the November print isn't a coincidence or a confirmation; it's the same event described twice.
⚠ The Trap This Chart Sets — Read Before September
The most useful thing in this model is not the crossover. It's the trough.

Total revenue bottoms at roughly $137M in Q4 FY26 — the print landing ~September 16. That quarter should look genuinely bad: revenue down for a third straight quarter, GAAP loss still fat with decommissioning charges, mining collapsing faster than Horizon 1 can bill. This is structural, not a failure. You are watching a company switch off a cash-generating asset before its replacement starts invoicing. There is no version of this transition without that valley, and the valley is deepest one print before the inflection.

So pre-commit now, while it's cheap. A weak revenue headline in September is the model working, not a break condition. What matters in that print is exactly one thing: did Horizon 1 complete, and does the DDTL draw cadence corroborate it? If H1 lands and the draws support it, a $137M quarter is noise. If H1 slips — or billable revenue misses with no supporting draw evidence — that is thesis damage, and no amount of mix-shift charting offsets it.

The failure mode is selling a structurally ugly print that your own model predicted. Decide the reading before the tape provides the emotion.
Catalyst Roadmap

WHAT TO WATCH

The thesis resolves on a schedule now. September tests credibility; November is the BIP-analogue inflection; May 2027 is the first clean run-rate quarter. Everything between those dates is beta. Q4 FY26 confirmed for ~September 16, after close.

Completed
May 7, 2026
NVIDIA — $3.4B Contract + 5GW Partnership
Second anchor customer. 30M warrants at $70 strike. Validating — but CoreWeave, Lambda, Nebius and Crusoe have similar NVIDIA ties. Not a moat.
Completed
May 14, 2026
$3.0B Convertible Notes Closed
1.00% coupon, 32.5% premium, $201.3M capped call → ~$110.30 effective cap. Arb selling pressure has long since cleared.
Completed
June 1, 2026
$3.65B Investment-Grade GPU Financing Closed
The most important event of the year. Rated A / A(low). Credit markets re-rated the Microsoft cash flows a year before the equity market will.
Completed
June 3 / June 15, 2026
Bundey 800MW + Nostrum Closed
Two new continents in twelve days. APAC via a signed 330kV transmission agreement; Europe via 490MW of secured Spanish power and a 50-person team.
Completed
June 26, 2026
Added to Russell 1000
Index inclusion widens the passive holder base. Mechanical, not fundamental — but it changes who owns the float.
Resolved Against Us
July 6, 2026
Anthropic → TeraWulf, Not IREN
20-year, 401MW, ~$19B lease at Hawesville KY. The highest-probability next-customer hypothesis went elsewhere. Mitigant: Anthropic reportedly signed 12+ LOIs with developers and already runs Amazon, Google, Microsoft, Fluidstack and SpaceX in parallel. Multi-vendor by design — but IREN was not first in the queue.
Completed
July 20, 2026
$2.8B Signed — Eight AI Developers
Year-end ARR target raised from $3.7B to $4B+, ~85% under contract. ~45% of associated GPU capex prepaid; ~4-year weighted average term. A partial add-trigger fire — fundamental, and it weakens the concentration critique, but not the hyperscaler anchor the rule specifies. The pop reversed inside a week.
Completed
August 3, 2026
Mirantis Acquisition Closed
13,673,894 shares aggregate (~$625M at the $45.66 May 1 close) plus ~$40M cash and RSUs. Adds orchestration, monitoring, support and 1,500+ enterprise customers. The move up the stack is now executed, not announced — which is the axis the whole re-rating case runs on.
Horizon 1 — What Is Actually Known, August 6
A claim circulated in late July that Horizon 1 was handed to Microsoft on July 16, 2026. It traces to a single social-media post, recirculated by aggregators. There is no IREN press release and no 8-K.

The evidence against treating it as fact is IREN’s own. On July 20 — four days later — the company published a promotional release announcing $2.8B of new contracts and a raised ARR target. It does not mention a Horizon 1 handover. It describes IREN as allocating capacity ahead of commissioning. A company that PRs a CISO appointment does not sit on the first delivery under a $9.7B contract for the sake of modesty. Satellite-imagery posts — fresh paving, a Microsoft security booth — are consistent with an imminent handover. They are not the handover. Nothing has slipped: guidance is Q3 CY2026, which runs to September 30.

The disclosure that actually matters. Footnote 2 of that same release states the $4B+ target assumes 480MW gross by year-end, with revenue ramping subject to “commissioning, testing and customer acceptance” of GPUs — and it puts that process in the months following each data center’s delivery.

That is the three-step gate in the company’s own words: delivery, then commissioning and testing, then customer acceptance, then revenue. This page previously carried that sequence as an inference. It is now disclosed. Two consequences. First, a confirmed Horizon 1 handover is not a revenue event and should not be traded as one — the gap is months, stated by the company. Second, it independently supports the calendar below: a Q3 delivery produces very little in the September print and lands as real ARR in November and February. The delivery date is the question everyone is asking. The acceptance date is the one that pays.
The Execution Test
~Sept 16, 2026
Q4 FY26 — Horizon 1 Credibility
First partial ARR visibility. Working view: H1 completion lands nearer September than management's earlier signalling implied, and the company has now disclosed that acceptance trails delivery by months. Expect thin billing here even on a clean delivery — judge this print on the handover itself and the DDTL draw, not the revenue line. A slip past September 30 with no supporting draw evidence is genuinely thesis-damaging. H2–H4 track ahead of H1.
Anchor — Thesis Validation
November 2026
Q1 FY27 — First Real ARR Step-Up
The BIP 2012–13 moment. First meaningful ARR recognition from a partial Horizon 1 billing quarter. This is where contracted cash flows stop being a slide and start being a line item. Anchor metric: FCF after maintenance capex — not GAAP net income, which GPU depreciation renders uninformative.
Confirm
February 2027
Q2 FY27 — First Full Horizon 1 Quarter
H2–H4 ramping alongside. The shape of the curve becomes visible rather than inferred.
Run-Rate Inflection
May 2027
Q3 FY27 — All Four Horizons + NVIDIA Live
First clean full quarter with everything billing. The primary run-rate inflection — and the point at which a 6–8x EBITDA multiple becomes indefensible on the numbers rather than the narrative.
The Add Trigger
Undated
Third Anchor Customer
Sweetwater is uncontracted and Bundey is unbuilt. A new hyperscaler deal is the add condition — a fundamental catalyst, not a sentiment dip. Meta and Oracle now rank above Anthropic. OpenAI uncertain (MNPI dynamics, Musk litigation noise). Status: unfired. The July 20 cohort is logged as a partial fire — right category, wrong counterparty scale.
Governance
2026 AGM
Say-on-Pay Vote — Proxy Filed July 9
Watch ISS and Glass Lewis. The founder grants consumed ~71% of the 25.5M-share 2025 incentive pool. A significant against-vote would not change cash flows, but it would price the board's judgement.
The Uncomfortable Part

GOVERNANCE — THE RSU GRANTS

A bull thesis that doesn't carry its own strongest counter-argument isn't a thesis, it's marketing. So: here is the thing the board did that a lot of shareholders hated, what defends it, and what doesn't.

What Happened
18,198,656
RSUs — 9,099,328 to each Founder-CEO
Approved by the board June 30, granted July 1, 2026. At the June 29 close of $45.91 that was ~$418M per CEO — ~$836M combined. Four equal annual tranches, each locked a further two years after vesting; the final tranche is unsellable until fiscal 2033. No further equity grants until fiscal 2031. Across the 2025 and 2026 awards together, each Founder-CEO holds RSUs worth roughly a 3% stake.

The grant is ~71% of the 25.5M shares in the 2025 Omnibus Incentive Plan pool — 698,656 units more than the 17.5M pool shareholders originally backed in November. The stock sold off hard in the two sessions that followed.
The Board's Defence — Chair's Letter, July 8
Independent chair David Bartholomew wrote to shareholders directly. His argument, fairly stated:

1. The headline number is a share-price artefact. The process began mid-2025, when the market cap was under $4B. It's now over $16B. The board deliberately granted less and deferred the rest until performance was proven — which put the risk on the founders, not shareholders. The dollar figure is large precisely because the deferral worked.

2. Hurdles were tried and failed in both directions. Earlier awards with share-price hurdles failed to vest during a weak sector tape even as IREN outperformed peers. Restructured in 2025 to time-based. Then half the 2025 award's hurdles were satisfied far earlier than the multi-year alignment intended, because the stock ripped. Both failure modes, same instrument.

3. There is no floor. Fixed share count, no reset, no repricing. "The value of the award is determined by the same share price that determines shareholder returns, rising and falling exactly as your shares do." Locked to 2033.
Where This Page Lands
The critique survives — but narrowly, and on design, not size.

The size objection is weak and this page won't make it. If you hold a re-rating thesis, you are explicitly forecasting a large number here; complaining when the number arrives is incoherent. Bartholomew's point that the headline reflects a $4B→$16B move is correct, and the 2033 lock is genuinely long by any standard. The grants have zero FCF impact — they are ASC 718 expense, and this page's anchor metric is free cash flow after maintenance capex.

What doesn't survive is the leap from "hurdles are hard to set" to "therefore, no hurdles." That's a non-sequitur. The board's own evidence — hurdles missed in a bad tape, hurdles cleared too easily in a good one — is an argument against absolute share-price hurdles. It is not an argument against relative TSR against a neocloud basket, which is immune to exactly the sector-beta problem the letter describes. Nor is it an argument against operational hurdles: Horizon delivery dates, contracted ARR, MW energized. IREN has a beta near 3.4. A relative-TSR structure was the obvious instrument, it was available, and the board didn't reach for it.

So: time-vesting was a choice, and the letter defends the outcome rather than the instrument. That's the critique. It is narrow, it is real, and it does not move a single dollar of the November 2026 cash flow this thesis rests on.
The Delivery Discount

EVERYONE HAS MEGAWATTS. ALMOST NOBODY HAS HANDOVERS.

The sector does not trade on contracted power. It trades on a discount for the gap between announced capacity and capacity that actually bills — and that discount is the largest single component of the re-rating case.

Every neocloud has a gigawatt slide. The market has learned to discount them, and it has learned for cause. Core Scientific guided powered infrastructure for CoreWeave by mid-2025, hit permitting friction and GB200 redesign issues, and re-guided to 250MW by year-end. Nebius’ New Jersey delays have been persistent. A STACK–Oracle permitting impasse stalled until Bloom Energy stepped in. Applied Digital delivered Building 2 at Polaris Forge 1 in November 2025 — then in March 2026 amended the lease to suspend the term on two of its four data halls. Delivered is not the same as billing, and the filings say so.

This is not a claim that the sector is failing. SemiAnalysis, which called the Core Scientific and Nebius slips early, also correctly modelled CoreWeave’s 1.7GW year-end target as achievable. The point is narrower and more useful: delivery is the variable the market cannot verify from a press release, so it applies a blanket haircut and waits. Liquid cooling made the haircut worse. OpenAI’s Abilene expansion stalled on liquid-cooling reliability. Hyperscalers run unforgiving acceptance testing, and every rack brings firmware qualification against new GPU SKUs.
Structural — Not Cultural
IREN designs, engineers and builds in-house. No third-party colocation developer sits between the contract and the concrete. Most of the sector’s documented slippage is counterparty slippage — a developer missing a date for a neocloud tenant. IREN structurally cannot have that failure mode, because there is no counterparty to miss.
The Record
5 EH/s to 50 EH/s on the schedule it set. Sweetwater 1 energized on schedule. Approximately 3MW of self-built AI Cloud capacity to 480MW being delivered this year, per Roberts on July 20, with 1.2GW targeted for 2027. No missed public milestone to date.
The Unproven Part
Every milestone above is one IREN set for itself. Horizon 1 is the first date a hyperscaler gets to grade. Microsoft’s acceptance testing is not IREN’s energization checklist. CoreWeave and Nebius have completed hyperscaler-standard handovers; IREN has not yet completed one.
Why This Is A Gate, Not The Trigger
The tempting version of this argument is that Horizon 1 lands, the delivery discount collapses, and the stock re-rates. That skips a step. Delivery proof is a precondition — the same class of thing as the AI-versus-Bitcoin revenue mix. It has to be true before GICS reclassification, coverage-desk migration and the GPU-impairment overhang can do their work. It is not the mechanism itself.

What clearing it buys is specific and worth having: it removes the largest input to the discount and it does so on a name whose credit has already re-rated. That is the BIP sequence — credit first, then proof of operations, then equity. An investment-grade GPU facility at ~3.31% blended is the market already saying the cash flows are good if they arrive. Horizon 1 is the arrival.

The honest limit. Clearing H1 makes IREN a proven executor. It does not make IREN the best executor in the sector — that claim needs completions IREN does not yet have, against peers who do. And note the tape’s own verdict: if a first hyperscaler handover were the re-rating trigger, a confirmed one would not leave this stock at roughly half its high. The discount is real. The catalyst that collapses it is more likely the November print showing acceptance converting to billed revenue than the handover headline itself.
Valuation Framework

WHAT IS IREN WORTH?

Market prices IREN at construction-phase multiples on a company with $15.9B of signed contracts and an investment-grade rating on the cash flows. The re-rating gap is the thesis. Prices as at August 6, 2026 — IREN's ~3.4 beta means this moves 10%+ on sector days.

Current Market — Aug 6, 2026
~$14B market cap
~$39
~50% below the $76.87 high (Nov 5, 2025)
Traditional Infrastructure
12x EBITDA
~$52
+30% from here
Street Consensus — 16 Analysts
Buy · Range $41–$131
~$82
Bernstein $100 · H.C. Wainwright $90 · JPM at Sell
Digital Realty Comp
20x EBITDA
~$96
+140% from here
Equinix-Style
25x EBITDA
~$124
+210% from here
Agrippa Base Case
Full build fair value
>$150
Bundey, Spain and Kiowa not yet modelled
The Brookfield Infrastructure Partners Parallel

BIP spun off into the 2008 financial crisis, trading to $5–6. Management acquired distressed assets. The stock drifted until 2012–13 when contracted cash flows became undeniable — then re-rated hard. 18% annualised total returns since inception. Same model: own the infrastructure, bring institutional co-investors in at the asset level, retain the promoted interest.

The sequencing is what matters, and it is now visible. Credit markets re-rate before equity markets. The $3.65B facility being stamped A / A(low) in June 2026 is the analogue of BIP's cash flows becoming bankable in 2011 — the rating agencies got there first because their job is to underwrite contracts, not narratives. Equity follows when the cash flow shows up in a print. That print is November 2026. The investors who sit through the construction grind capture the bulk of the lifetime return — and the grind is the whole point, because it's what keeps the multiple low while the contracts are already signed.

The World Agrees

DEMAND IS INFINITE

"We cannot build datacenters fast enough. The demand is just so enormous."
Brad Smith
President, Microsoft
"I honestly believe the AI revolution is underhyped. US datacenters will need +29 GW by 2027 and +67 GW by 2030."
Eric Schmidt
Former CEO, Google — Congress Testimony
"GPU demand is infinite."
Kevin Weil
Chief Product Officer, OpenAI
"We have better models, and we just can't offer them because we don't have the capacity."
Sam Altman
CEO, OpenAI
"Time to data center is the key decision point in many commercial discussions."
Kent Draper
Chief Commercial Officer, IREN
"We don't have a demand problem — we have a power and capacity problem."
Digital Bridge
Global Infrastructure Panel
Current Rating — August 2026
HOLD & WATCH

Thesis materially strengthened since April on the things that matter: investment-grade credit rating, a second anchor contract, a $4B+ year-end ARR target with ~85% under contract, and two new continents. The enterprise-cloud gap vs. CoreWeave and Nebius was the strongest bear argument on this page in June; the July 20 cohort and the closed Mirantis deal have narrowed it, not closed it. Anthropic still chose TeraWulf.

The add trigger has not fired. A third anchor customer is the condition, and eight AI developers do not meet it. The drawdown since July is sector beta, a governance tantrum and a JPMorgan short call — none of them a fundamental catalyst, and treating a dip as an add-trigger would be a quiet rule change. September is the execution test; November is the thesis.

Break Conditions — None Currently Triggered
⚠️
Horizon 1 slips past September 2026, or a billable-revenue miss with no supporting DDTL-draw evidence
⚠️
No third anchor customer by mid-2027 — Sweetwater and Bundey remain uncontracted capacity
⚠️
GPU rental pricing compresses structurally — JPMorgan's Sell case; Meta selling excess capacity is the live version
⚠️
Capital structure deterioration forcing distressed equity issuance
Signal vs. Noise Signal: DDTL draw cadence · hyperscaler announcements · Horizon delivery dates · FCF after maintenance capex.
Noise: analyst PT dispersion ($41 to $131 on the same company) · convertible-arb selling · 3.4-beta sector days · GAAP net income, which GPU depreciation makes uninformative · satellite and land-record inference, which produces confirmation loops without touching the variables that set the price · 13G filings from broker-dealer parents — BofA's 21.03M-share 5.8% passive stake, disclosed August 3, read as validation and moved the stock 8%. But a BofA Corp 13G aggregates BofA Securities and Merrill market-making and hedging inventory, and IREN has $3.0B of converts outstanding. Same category as the Susquehanna filing, 12.6M of whose 14M shares were options.
Short Interest — Read It Correctly FINRA reported 94M shares short as of July 15 — 26.22% of outstanding. That number gets quoted as a coiled spring. It is not. Borrow costs 0.64% with 2.1M shares available and 2.33 days to cover. Cheap borrow with open inventory is what a crowded short does not look like. A material share is convertible delta hedging against the $3.0B notes and ATM-related mechanical flow — positions that are short the stock without being bearish on it. Rising short interest alongside a price drawdown is logged here as noise, in both directions: it is not confirmation of the bear case, and it is not a squeeze setup.