IREN

• FY2025 revenue $501.0M (10-K, FYE 6/30/2025) vs $187.2M FY24 — a Bitcoin-era base.
• FY2026 9-month revenue $569.8M vs $313.7M prior (+81.6% YoY).
• Q3 FY26 (ended 3/31/26) revenue $144.8M = Bitcoin mining $111.2M + AI Cloud Services $33.6M.
• AI Cloud Services: Q3 $33.6M vs $3.6M prior (+839% YoY); 9-mo $58.3M vs $9.4M (+6.2×). This is the thesis segment inflecting.
• Q3 FY26 operating loss –$233.5M (9-mo –$426.3M); net loss Q3 –$247.8M.
• Balance sheet (3/31/26): cash & equivalents $2.213B; PP&E net $4.37B (vs $1.93B at 6/30/25 — buildout capex); total assets $7.27B; total liabilities $4.60B; equity $2.66B.
• D&A 9-mo $305.6M (vs $117.3M prior) — the depreciation drag that makes GAAP income deeply negative even as revenue scales.
• Debt: 2029 and 2030 convertible senior notes (two tranches) — the 2024/2025 capital raises that funded the buildout.

Vertically integrated AI Cloud operator: owns grid-connected land/power, builds liquid-cooled data centers, deploys NVIDIA GB300 NVL72 (and mixed GPU) clusters, and (post-Mirantis) sells orchestration/managed services. FY26 still majority Bitcoin-mining on a GAAP P&L, with AI Cloud now the growth and contracted-revenue engine.[1][2]

• monetization model: Bitcoin self-mining (block rewards + transaction fees) + AI Cloud GPU services (contract/usage). No hardware/consumables/license. Source: IREN FY2025 10-K, Item 1 Business; sec.gov ; accession UNKNOWN.
• revenue mix: FY2025 total $501.0M: Bitcoin mining ~90.5% ($453.7M); AI Cloud ~9.3% ($46.4M); other ~0.2% ($0.9M). Source: FY2025 10-K, segment/revenue notes; same URL; accession UNKNOWN.
• who pays: Bitcoin mining pools/BTC market for mining; AI/ML enterprises and hyperscalers for AI Cloud. Source: FY2025 10-K, Business/Revenue; same URL; accession UNKNOWN.
• margin_value_capture: Value accrues to low-cost power, owned data centers, GPU utilization; gross/operating margin UNKNOWN (not verified from filing). Source: FY2025 10-K, MD&A/Cost of revenue; same URL; accession UNKNOWN.

• customer classes: Hyperscalers, AI/ML enterprises, Bitcoin mining pools. Source: IREN FY2025 10-K, Business/Customer concentration; same URL; accession UNKNOWN.

IREN is a vertically integrated owner and operator of data centers. Its principal activities are:

1. Bitcoin mining — operating ASIC fleets that perform SHA-256 hashing and earn bitcoin block rewards and transaction fees.
2. AI / HPC cloud services — operating GPU-based data center capacity and selling compute to customers needing AI training or inference.

The company presents itself as a data-center operator rather than a pure asset-light mining pool participant. The critical business-model issue is whether revenue remains dominated by bitcoin mining or whether AI/HPC becomes a material second engine.

The exact FY2024 revenue split must be read from the IREN 20-F. I will not assert it as fact here. A working-memory estimate, marked unverified, is:

IREN’s monetization has two distinct mechanics:

1. Bitcoin mining: IREN contributes compute to the Bitcoin network and receives bitcoin. Revenue is ultimately realized when bitcoin is sold, but the financial statement treatment depends on IREN’s accounting policy — whether mined bitcoin is measured at fair value less costs to sell, as inventory, or as an intangible asset. This is a critical accounting-policy item that must be quoted directly from the 20-F.

1. AI cloud: IREN receives cash or accrued receivables from customers for GPU compute, typically under service agreements. Pricing may be reserved, on-demand, or committed-contract, but the exact structure is disclosed only in the filing.

IREN sells mined bitcoin into the spot market or to liquidity providers/exchanges. The effective buyer classes are:

• Digital asset exchanges;
• OTC market makers;
• Principal trading firms;
• Institutional digital asset funds.

No named counterparty can be cited from filings here. If IREN’s 20-F names a top Bitcoin sales counterparty, that must be pulled from the filing. Without that, the claim is not verifiable in this session.

Potential customer classes, based on the nature of the product, are:

• AI labs training foundation models;
• Enterprises running fine-tuning or inference;
• Startups needing burst or dedicated GPU capacity;
• Cloud resellers or platform partners.

But no customer name, contract size, or concentration percentage is verified here. Any named customer should be sourced from IREN’s 20-F or 6-K only. If no customer is named, the correct research conclusion is: IREN’s filings do not disclose top AI cloud customers in the material reviewed; this is a disclosure limitation. That statement itself needs EDGAR verification.

FY26 (ended 2026-06-30) total revenue $707.0M vs FY25 $501.0M (+41.1%); AI Cloud Services $128.8M vs $16.4M (~7.9×); Bitcoin mining $578.2M vs $484.6M.[2][3] Q4 FY26 revenue $137.2M (AI Cloud $70.5M, mining $66.7M) — first quarter AI Cloud > mining.[2] Gross profit derived: revenue $707.007M − CoR $219.706M = $487.3M (68.9% GM); AI Cloud CoR only $16.932M on $128.795M (13.2% CoR / 86.8% contribution before D&A).[1][3] GAAP operating loss −$1,046.7M; net loss −$702.6M; D&A $417.7M; impairment $638.8M (mining-hardware decommission + retrofit).[3] Cash & equivalents $5,895.6M + restricted cash $1,723.9M; total assets $15.790B; liabilities $11.604B; equity $4.186B; long-term debt $7.593B; PP&E net $6.753B.[1][3] Deferred revenue $1.843B; contract liability $219.1M; unsatisfied RPO $5.1B ($0.9B in next 12 months) plus ASC 842 lease contracted value $11.4B → combined remaining contracted $16.6B.[1]

IREN monetizes the power-interconnection + liquid-cooled rack-density bottleneck that GB300 NVL72 physically requires. It is a capacity-cycle / qualification neocloud, not a chip or IP toll-road.

1) Rack-physics recalibration (NVIDIA measured, not assumed).

Prior note used ~120 kW/rack → ~416 racks / ~30k GPUs per 50 MW Horizon. NVIDIA ERA: 142 kW per GB300 NVL72 rack.[5]

• 50,000 kW IT ÷ 142 kW/rack = 352 racks per Horizon
• 352 × 72 GPUs = 25,344 GPUs per 50 MW
• Four Horizons (200 MW IT) = 1,408 racks ≈ 101,376 GPUs

Sensitivity: if IT load is derated 10% for PUE/headroom (management has described 200 MW IT on 300 MW gross as the conservative envelope on the FY26 call[13]), racks fall to ~317 / Horizon. The physics gate does not move: 142 kW >> air-cooled ~20–40 kW/rack.

2) Microsoft $ / MW vs new-contract $ / MW.

Microsoft Agreement: ~$9.7B / 5-year average term / four 50 MW (IT) tranches = 200 MW IT → $9.7B ÷ 200 MW ÷ 5 yr = $9.7M per MW-year (derived).[1] 20% of each tranche prepaid before delivery.[1]

EX-99.1 (Aug 27): recent 3-year contracts >$20m revenue per MW (IT) with ~2-year GPU payback; active discussions ~$25m per MW (IT); recent prepayments 45–55% of GPU capex.[2] EX-99.1 does not state whether $20m is per-year or term-total; we do not annualize it. Directionally management presents it as higher value per MW than the installed Microsoft book.

Payback sanity (if $20m were annual, which is assumed and flagged): AI Cloud contribution ~86.8% before D&A → ~2 × $20m × 0.87 ≈ ~$35M GPU+ancillary capex per MW would be implied. That is a model, not a filing number.

3) Contracted stack vs operating P&L (the GAAP/non-GAAP gap).

The $4bn ARR is GPU-hour × 8,760, explicitly not a substitute for GAAP revenue (EX-99.1 notes 1–2).[2] GAAP RPO of $5.1B is the conservative recognized-revenue ceiling over 60 months; $0.9B of it lands by Jun 30, 2027.[1]

4) Funding vs FY27 capex (the real time-to-monetize risk).

Management: FY27 capex $25–30B; ~$14B existing cash + committed GPU financing + prepayments (incl. $7.6B cash of which $1.7B restricted, mostly Microsoft GPU capex); targeting additional ~$8B GPU financing/prepayments.[13][1] 10-K: principal outstanding indebtedness $7,976.0M ($7,705.6M debt + $270.4M finance leases) plus $1,825.4M trade payables.[1] GPU financing for Microsoft: $3.6bn IG at ~6.0% WAC; DDTL $413M funded / $1,132M undrawn at Jun 30; USPP notes 5.96% fixed.[1][2] Non-IG: $2.8bn (incl. $2.4bn Blue Owl / PIMCO at 9.0% for Mackenzie).[2]

If the $8B incremental GPU financing does not close, Horizons 2–4 and 2027 air-cooled conversions slip — that is the binding capital, not demand, constraint.

Toll-road classification: 4 = WEAK_PROXY (bypass 2 / concentration 1 / hyperscaler 1). Unchanged vs 2026-08-22.

• Bypass (2): 20% Microsoft prepay + 45–55% GPU-capex prepay on recent deals is revealed-preference that buyers cannot source the same MW on the same clock, but 10-K lists AWS/GCP/Azure/OCI plus CoreWeave/Nebius/Crusoe/Lambda as competitors and notes customers can also be competitors.[1]
• Concentration (1): MSFT $9.7B + NVDA $3.4B = "substantial majority of contracted revenue."[1] Named diversification (Prometheus/Perplexity/Together AI $2.8B; Cohere, Figure, Fal, Higgsfield, frontier lab) is real but not yet majority.[1][2]
• Hyperscaler (1): the book is still selling capacity to Microsoft and NVIDIA. IREN is an arms dealer, not a gate the hyperscaler must buy through.

Hypothesis verdict: STRENGTHENED — FY26 10-K + Horizon 1 acceptance + NVIDIA $3.4B customer contract + $16.6B remaining contracted + Q4 AI Cloud crossing 50% of quarterly revenue all confirm execution. Sweetwater Batch Zero (Sep 8) is a positive but conditional interconnection catalyst, not a completed toll-road. The moat remains a CAPACITY_CYCLE / WEAK_PROXY neocloud: being wrong is a slipped Horizon 2–4 / failed $8B financing / Microsoft tranche termination; being right is $4bn ARR converting into GAAP revenue through CY2027. Conviction stays conservative — this is not an irreplaceable bottleneck.

Power draw at 25 J/TH:

\[

\text{Power (MW)} = 10{,}000{,}000 \text{ TH/s} \times 25 \text{ J/TH} / 1{,}000{,}000 = 250 \text{ MW}

\]

Daily electricity consumption:

\[

250 \text{ MW} \times 24 \text{ h} = 6{,}000 \text{ MWh/day}

\]

Daily electricity cost:

\[

6{,}000{,}000 \text{ kWh} \times \$0.045 = \$270{,}000/day

\]

Daily revenue at $0.06/TH/day:

\[

10 \text{ EH/s} \times 1{,}000{,}000 \text{ TH/EH} \times \$0.06 = \$600{,}000/day

\]

Implied electricity cost per TH/day:

\[

\$270{,}000 / 10{,}000{,}000 = \$0.027/TH/day

\]

Daily BTC mined at 10 EH/s out of 600 EH/s network:

\[

\frac{10}{600} \times 450 = 7.5 \text{ BTC/day}

\]

Cash electricity cost per BTC:

\[

\$270{,}000 / 7.5 = \$36{,}000/BTC

\]

At $60,000 BTC, the illustrative cash gross margin after electricity alone is:

\[

\$60{,}000 - \$36{,}000 = \$24{,}000/BTC

\]

That is a 40% cash gross margin before all other operating costs, depreciation, financing, and corporate overhead.

Break-even hashprice against electricity only:

\[

\$0.027/TH/day

\]

This calculation shows why IREN’s mining economics are fragile if hashprice falls below its electricity cost. The actual breakeven is higher after overhead, depreciation, and financing.

These are not IREN-verified numbers. Plug in actual deployed hashrate, actual efficiency, actual power price, and current hashprice from IREN’s 20-F/6-K and live network data before use.

\[

\text{Revenue} = 1{,}000 \times 0.75 \times 8{,}760 \times \$2.50 = \$16.425 \text{ million/year}

\]

\[

\text{Power cost} = 1{,}000 \times 0.75 \times 8{,}760 \times 0.7 \times \$0.045 = \$206{,}955/year

\]

This shows a very high electricity-only contribution before depreciation, networking, software, staffing, and capital costs.

If IREN can sign committed AI contracts at high utilization and high rates, the contribution per MW can exceed bitcoin mining at certain hashprices. But if spot GPU pricing falls or utilization drops, the economics deteriorate quickly.

If Riot has lower disclosed cost per bitcoin or larger cash liquidity, then IREN’s cost advantage must be proven against Riot’s primary filings, not news summaries.