IREN — knowledge base

Overview

IREN is pivoting from a bitcoin-mining-led operator into a vertically integrated AI-infrastructure platform combining grid power, powered land, high-density data centers, NVIDIA accelerators, networking, cooling, and software/orchestration (Mirantis/Morantis, closed 2026-08-03, announced 2026-08-04; named an inaugural NVIDIA-certified hypervisor on 2026-08-27) to deliver AI compute and, longer term, token output while protecting customer data.

The recurring thesis has three layers:

Near term: Horizon One is delivered and accepted; convert sold-out 2026 capacity into recognized revenue; deliver Horizons 2–4 in the December 2026 quarter; fund a much larger FY27 build.
Medium term: use Microsoft and other creditworthy contracts to de-risk financing, then expand Sweetwater, Canada, Australia (Bundy), Spain (Nostrum), and Oklahoma while mixing enterprise, AI-native, frontier-lab, and sovereign demand.
Long term: move beyond powered shells or bare-metal GPU leasing toward AI factories, managed compute, orchestration, and potentially token-output economics.

Updated as of 2026-08-27 (FY26 results): management framed a full pivot to AI cloud. $4 billion of ARR is contracted for 2026 capacity, of which $1 billion is operating after Microsoft accepted Horizon One. 2026 capacity was described as sold out. New multi-year customers include Cohere, Prometheus (previously unnamed), Perplexity, FigurAI, FowlAI/Fal.ai, HiggsfieldAI, and a separate undisclosed frontier AI lab. Three-year contract pricing is up ~125% since November; recent deals exceed $20 million per IT megawatt, with active discussions around $25 million. Platform targets ~300 MW IT in 2026 and ~0.5 GW more in 2027 against a 5+ GW announced pipeline. This official IT-load guide supersedes earlier speaker construction figures (~730 MW toward ~1.21 GW in 2027), which were not company guidance and mixed gross-power and IT concepts. ▶ 00:04:43

Updated: Horizons 2–4 are officially targeted for the December 2026 quarter. That supersedes speaker expectations of Horizon Two in late September or ~one month from 2026-08-21. Horizon One (50 MW IT, first of four 50 MW liquid-cooled Childress phases under the ~$9.7 billion Microsoft GB300 contract) was delivered and accepted ~2026-08-13, ~nine months after signing, dual-validated by Microsoft and NVIDIA, with NVIDIA Exemplar Cloud status on GB300 NVL72. Technical difficulty of a standalone 50 MW phase is limited; speakers argue moat rises exponentially toward multi-hundred-MW and GW-scale factories. “Time-to-computer” (secured power, liquid-cooled build quality/speed, capital-to-flywheel financing) is argued to outrank peers’ short-term “time-to-revenue” optics. ▶ 00:05:07

The operating-model choice remains a mix, not a settled identity. Direct compute offers higher potential returns and control over token economics but brings utilization, GPU obsolescence, software, customer-acquisition, financing, and residual-value risk. Long-term colocation offers contracted cash flow and refinancing support while capping upside. Updated as of 2026-08-27: the disclosed model is an anchor Microsoft GPU-cloud tranche plus multi-year enterprise/AI-native/frontier contracts at neocloud-like pricing (>$20M/MW IT), with Morantis positioned for managed services—not a pure colo or pure hyperscaler book. Pre-earnings (2026-08-23), Frans Bakker was “leaning heavily towards NVIDIA for the majority or if not all of the remaining capacity”; that is speaker analysis, not guidance, and the call instead named a diversified customer set. On-demand/burst compute was discussed as an option but not preferred versus longer high-margin reserve contracts. ▶ 00:24:33

Peer economics still frame the debate. Nebius contracted prints of $20–25M/MW (mid-term/GB300), paybacks compressed toward ~1 year 10 months, and a backlog still heavily bare-metal (Meta/Microsoft) sit beside colo comps (Riot–Anthropic ~$2.4M average annual revenue per MW; Fermi–TensorWave real but speaker ~$2.7M/MW figure overstated at ~$1.95M). CoreWeave A100 contracts through 2029 undercut two-to-three-year obsolescence bears. IREN’s recent air-cooled B300 economics were discussed at ~$17.1M/MW (speaker calc; unverified in filings). Updated: company-stated recent multi-year pricing above $20M/MW IT, talks ~$25M, now supersedes that speaker calc as the relevant contracted benchmark. Valuation debate into the print cast IREN (~$14–15B range in mid-August discussion) as ~4–6 months behind Nebius on revenue at a deep discount to Nebius (~$62–70B post-Q2). ▶ 00:09:40

Financing is no longer a purely hypothetical edge. Updated: in the three months before the FY26 call IREN secured ~$6.5 billion of GPU financing, with more than 100% of associated GPU CapEx funded including prepayments; $2.8 billion required no investment-grade offtake and priced in single digits, including $2.4 billion at 9% fixed (Blue Owl/PIMCO). Microsoft-related investment-grade GPU financing was $3.6 billion at a ~6% weighted average, funding ~96% of associated GPU CapEx with prepayments. Updated: the long-standing claim that ~95–96% of all Microsoft-project capex was investment-grade debt is only confirmed for associated GPU CapEx, not buildings, cooling, or site works. FY27 CapEx is guided at ~$25–30 billion against ~$14 billion of cash and committed GPU financing/prepayments (including $7.6 billion cash at 30 June 2026, of which $1.7 billion restricted) and ~$8 billion more GPU funding targeted. All data centers are being kept unencumbered as future asset-finance capacity. NVIDIA partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR aim to mobilize >$500 billion of third-party AI-infrastructure capital; IREN has hired two senior KKR-linked experts and has said it continues to evaluate NVIDIA backstop arrangements against dilution of upside economics. ▶ 00:05:20

Power is a portfolio, not energized IT. Updated: fact-check as of 2026-08-23 confirms a secured/pipeline power portfolio of approximately 5–5.8 GW; the FY26 call cited a 5+ GW announced pipeline. Earlier “has 5.8 GW” language remains misleading if read as currently energized capacity. Sweetwater 1’s 1.4 GW substation was energized and reported 2026-05-01 (guided toward April/Q2). Sweetwater (~2 GW campus, Sweetwater 1+2) remains the expected NVIDIA DSX flagship under the up-to-5 GW partnership; no second equivalent flagship partner has been announced. NVIDIA conditional equity rights: up to 30 million shares at $70 (5 million shares per 100,000 GPUs, capped at 600,000 GPUs over five years) and a five-year $3.4 billion NVIDIA AI-cloud contract for 60 MW IT at Childress beginning 2027. Other sites: Childress 750 MW planned; Oklahoma 1.6 GW targeted ~2028 (SPP); Bundy/Bundey South Australia ~800–850 MW approaching groundbreaking (speaker intel, ahead of Spain/Kiowa); Spain/Nostrum ~490 MW; Canada ~160 MW. ▶ 00:09:57

Childress layout beyond Horizon 1–4 is still contested. Ground/satellite intel (2026-08-21) described mining halls cut to ~two-thirds height, a rumored horizontal building, and hybrid air-cooled load with liquid assist rather than full-liquid Horizon 5–6 copies—conflicting with an 2026-08-13 IR statement that H5/H6 will be liquid-cooled. Public materials emphasize liquid for H1–4 and also reference separate Childress air-cooled capacity. NVIDIA was rumored as the next Childress customer. Earnings Q&A covered liquid-cooling conversions; it did not, in captured remarks, resolve H5/H6 design or the NVIDIA-customer rumor. H2/H3 outside plant had been described as progressing at mining-era speed, with DC36 lab activity (chillers, generators, air tubes) read as secretive B300/peak-summer testing.

Open-weight models remain a possible demand driver: enterprises may reserve frontier models for complex work and shift simpler tasks to cheaper open-weight models, raising aggregate inference and private-compute demand. Open weights are not fully open-source. Lower token prices or efficiency gains could also compress infrastructure economics. A company chart claim that weekly token usage across eight OpenRouter-tracked models rose nearly 17x in eight months is unverifiable as stated. ▶ 00:01:56

Sovereign AI and data control stay central, but claims that using OpenAI or Anthropic automatically gives provider engineers access to proprietary enterprise data are misleading; handling depends on product, contract, and privacy settings. ▶ 00:13:08

Governance and communication remain overhangs even after a stronger forward narrative. Founder RSU packages (~9.1 million each to co-CEOs Daniel and William Roberts; reports value the combined package around $788 million / hundreds of millions), the Golden State Warriors sponsorship (June 2026 multi-year partnership, jersey patch, Official AI Cloud Partner; reportedly >$50 million per year), backward-looking ARR habits, limited marketplace disclosure, and opaque construction timelines have damaged sentiment. Class B voting control makes a hostile takeover unlikely without founder consent. Shareholders distinguish alignment with the multi-year strategy from agreement with every management decision. Some multi-name holders still call management reactive versus Nebius or retrofit operators such as White Fiber; a 2026-08-23 guest who owns both said the IREN strategy looks right now but was late on commercial and marketing build-out, with Nebius offering higher near-term agility and IREN the safer long-term physical-asset edge. Pre-earnings consensus was that the 2026-08-27 call had to sell the forward story because the quarter’s numbers were a sunk cost. ▶ 00:04:32

The durable common ground: power availability, construction cadence, contract structure, financing, and utilization matter more than short-term equity volatility. Demand continues to outpace supply. Near-term tape into the print was volatile and short-influenced (short interest last reported ~93 million shares, ~27–30% of float as of mid-July 2026). Sector headwinds discussed 2026-08-23 included CPI, Jackson Hole, rates, data-center moratoriums, and community pushback (including on peers such as CIFR). Nebius Vineland (NJ) work halted after a Bloom Energy fuel-cell/LNG pivot and DataOne stop-construction orders for unpermitted installs—cited as cautionary “time-to-revenue” regulatory risk versus IREN’s grid/EPC path. Sell-side often lumps neoclouds by announced MW without differentiating execution.

Key facts & figures

FY26 results and call were on Thursday, 2026-08-27. ▶ 00:01:19
$4 billion of ARR is contracted for 2026 capacity; $1 billion of that ARR is operating after Horizon One acceptance. Management described 2026 capacity as sold out. $1 billion operating ARR is a run-rate, not a quarterly revenue print. ▶ 00:04:43
Updated: July 2026 raised the year-end 2026 AI Cloud ARR target to over $4 billion, of which ~85% was under contract after $2.8 billion in new multi-year deals. The 2026-08-27 figure of $4 billion contracted is the current contracted amount. ▶ 00:17:39
June quarter (Q4 FY26, period ended 30 June, before August handover): revenue $137.2 million, including AI cloud revenue $70.5 million; net loss $684 million, largely non-cash impairments of $450.4 million as sites transition off mining. Prior quarterly AI-cloud prints were $17 million–$33.6 million. ▶ 00:21:50
New multi-year cloud contracts named on the call: Cohere, Prometheus, Perplexity, FigurAI, FowlAI/Fal.ai, HiggsfieldAI, plus a separate leading frontier AI lab. Prometheus was the previously unnamed party. Individual contract dollars were not broken out in captured remarks. ▶ 00:04:21
Three-year contract pricing up ~125% since November 2025; recent three-year contracts priced above $20 million per MW of IT load; active discussions around $25 million per MW IT. ▶ 00:09:40
Horizon One delivered to and accepted by Microsoft, announced 2026-08-13: 50 MW IT, first of four 50 MW liquid-cooled Childress phases (Horizons 1–4, ~200 MW IT) under the multi-year ~$9.7 billion Microsoft GB300 contract, plus ~$5.8 billion Dell hardware purchase agreement and ~20% prepayment. Prior guidance was Q3 calendar 2026; delivery landed before mid-Q3, ~nine months after signing (~Nov 2025). Dual Microsoft and NVIDIA validation; NVIDIA Exemplar Cloud on GB300 NVL72. Microsoft security on site post-handover. Modular liquid cooling; speakers cite a ~750-mile campus fiber network. Horizon One pro-rata ARR previously discussed at almost $500 million—part of, not equal to, the $1 billion now operating. ▶ 00:05:07
Updated: Horizons 2–4 targeted for the December 2026 quarter. Supersedes late-September / “~one month from 21 Aug” speaker delivery hopes. Ground intel that H2 B3 already had GPUs (as of 2026-08-13) and that outside plant was moving at mining-era speed is construction color, not a delivery date. ↗
Updated: Microsoft GPU financing confirmed. $3.6 billion investment-grade GPU facility at ~6% weighted average, funding ~96% of associated GPU CapEx including prepayments. This resolves the GPU-financing portion of the old unverified “95–96% of Microsoft-project capex” claim; it does not prove the same ratio for non-GPU site capex. Earlier discussion had blended ~3% on ~96% with ~45% prepayments—treat the call’s ~6% / ~96% GPU figure as the controlling disclosure. ▶ 00:17:10
~$6.5 billion of GPU financing secured in the three months before the call; more than 100% of associated GPU CapEx funded including prepayments. $2.8 billion needed no investment-grade offtake and priced in single digits, including $2.4 billion at 9% fixed (Blue Owl/PIMCO). ▶ 00:05:20
FY27 CapEx guided at approximately $25–30 billion, covering Microsoft delivery, 2026 ARR buildout, and further expansion. ▶ 00:19:09
Cash $7.6 billion at 30 June 2026 ($1.7 billion restricted), within ~$14 billion of cash plus committed GPU financing/prepayments. ~$8 billion of additional GPU funding targeted. All data centers kept unencumbered for future asset finance. Arithmetic gap versus the $25–30 billion CapEx guide remains. ▶ 00:20:33
Platform targets roughly 300 MW IT in 2026 and about 0.5 GW additional IT in 2027; 5+ GW announced pipeline. ↗
Headcount nearly tripled in FY26, including hundreds via Mirantis/Morantis and Nostrum; more than 4,000 people mobilized across active sites. Nostrum associated with ~490 MW and 50+ people, consistent with the Spanish acquisition. ▶ 00:16:18
Morantis/Mirantis k0rdent named an inaugural NVIDIA-Certified Hypervisor around 2026-08-27, extending NVIDIA validation to the software layer and, per the call, enabling managed services and new customer classes. Integration still ongoing. ▶ 00:03:32
Secured/pipeline power portfolio approximately 5–5.8 GW (grid-connected or pipeline), not all energized. Earlier “~3 GW / 4.5–4.6 GW with ~10% contracted” and “~1.8 GW with ~16% leased” are historical snapshots, superseded for current portfolio scale and for 2026 AI contracting (sold out / $4 billion ARR contracted). Claim of “nearly 3 GW added in six months” remains misleading versus disclosure timing. ▶ 00:09:57
Sweetwater 1: 1.4 GW substation energization announced 2026-05-01 after guidance toward April/Q2. Substation energization is not 1.4 GW of IT load online. Campus plan remains ~2 GW (Sweetwater 1+2); interconnection approvals secured. Expected DSX flagship under the May 2026 NVIDIA partnership for up to 5 GW; no second flagship partner announced. Ground intel (2026-08-21): major underground water/sewage, full metal-frame kit on site for ~200 MW IT, modular dry-cooler build; earlier aerials showed dry-cooler foundations implying ~125 MW IT coolers. ▶ 00:18:29
NVIDIA investment rights: up to 30 million shares at $70 (~$2.1 billion) tied to deployment milestones including up to 600,000 GPUs over five years. Separate five-year $3.4 billion NVIDIA AI-cloud contract, 60 MW IT at Childress, beginning 2027. Company continues to evaluate NVIDIA backstop structures against dilution of upside economics. ▶ 00:33:33
IREN issued roughly $1 billion of zero-interest convertible notes with ~42.5% conversion premium (historical). ▶ 00:56:38
Prince George: cloud revenue rose from $7 million to $18 million quarterly; $400 million contracted ARR targeting $500 million; 22k GPUs discussed as the $500 million ARR path. Canadian sites total ~160 MW, GPU cloud a minority. ▶ 00:09:44
Oklahoma 1.6 GW site announced, ~2028 target, Southwest Power Pool. ▶ 00:11:02
Childress planned campus 750 MW. Horizon 5–6 potentially ~100 MW; IR (8/13) said liquid-cooled—misleading/partial versus public materials that also cite air-cooled Childress capacity. Ground intel 2026-08-21: hybrid air-plus-liquid assist, halls cut to ~two-thirds height, rumored horizontal building; NVIDIA rumored next customer. John Gross hired Chief Innovation Officer (Feb 2026) for liquid-cooling/thermal standards. ▶ 00:31:26
Dell is counterparty on the ~$5.8 billion hardware agreement; Dell and Lenovo are NVIDIA OEMs working with IREN on AI-factory design. IREN owns the physical assets. Lenovo also tied to Canada GPU clusters. ▶ 00:02:20
Industry GPU pricing: 2026 reports of Nvidia GPU price increases of ~10–15% and later 15–30% on various lines amid supply constraints. ▶ 00:48:30
HBM (Samsung, Micron, SK Hynix), TSMC leading-edge (ASML EUV), advanced packaging, networking, OEM integration, cooling, and construction remain binding constraints alongside power. Nvidia Rubin/HBM architecture shifts are a monitored variable. ▶ 00:29:30
Data-center GPUs can remain economically useful well beyond two or three years, especially for inference. CoreWeave contracted Nvidia A100s through 2029 (~nine years from the 2020 launch). Unit economics still deteriorate versus newer generations. ▶ 00:05:47
CapEx to build ~1 GW of HPC commonly cited ~$50–60 billion+ (Jensen-linked range often $40–80 billion total; compute portion frequently $40–50 billion to Nvidia). Not an IREN-specific budget; IREN’s own FY27 guide is $25–30 billion, not a per-GW figure. ▶ 00:51:36
Warriors/Valkyries: June 2026 multi-year partnership, jersey badge/patch, Official AI Cloud Partner; value reportedly >$50 million per year (report-based, not a filing figure). ▶ 00:09:25
Co-CEO equity package: board approved ~9.1 million RSUs each (~18.2 million total), multi-year vesting; reports cite a combined value around $788 million (hundreds of millions). ▶ 00:04:32
Goldman Sachs 13G: ~33.64 million shares, 9.4%. JPMorgan also disclosed a stake. ▶ 00:14:53
Short interest last reported ~93 million shares (~27–30% of float, mid-July 2026 settlement). Into 2026-08-21 the stock held above $42; 2026 YTD low ~$29; 2025 ATH close $76.41 (15 Nov 2025) after ~15x from ~$5.30. No post-earnings price is in the new source material. ▶ 01:08:47
RPO weighted average remaining term previously 17–18 months; new AI contracts had averaged ~20% prepayment (July IR). ▶ 00:39:11
Applied Digital: initial ~250 MW / $7 billion expanded to ~400 MW / $11 billion 15-year CoreWeave leases. ▶ 00:09:18
Colo comps (2026-08-12): Riot ~$9.1 billion / 20-year / 191 MW Anthropic add-on at Rockdale (~241 MW total contracted) implies ~$2.4 million average annual revenue per MW (accurate). Fermi 15-year ~$6.5 billion / 222 MW TensorWave lease at Project Matador is real; speaker ~$2.7 million/MW was overstated (~$1.95 million). ▶ 00:12:57
Nebius Q2 2026: revenue $582.3 million (+454% YoY); adj. EBITDA $236.2 million vs ~$173 million estimate; AI cloud $575 million; ARR $3.0 billion; ACV >$20 million/MW (landmark deals $20–25 million); year-end 2026 contracted power target 5 GW; expects >$9 billion customer prepayments in 2026; shareholder letter said it could sell entire 2027 capacity on current terms. Post-print NBIS ~$259 (+~34%). Sites include France (near Lille, Meta references), Finland (Lappeenranta, up to 310 MW; Microsoft references), Birmingham/UK and US allocations. Vineland NJ halted on Bloom/LNG permitting. August 2026: upsized ~$5 billion convertible senior notes (e.g. 0.50% due 2030 and 4.50% due 2034). ▶ 01:05:50
CoreWeave Q2 2026 backlog ~$104 billion as of 30 June did not include >$25 billion net new customer commitments added in early Q3. ▶ 00:17:58
Texas: Gov. Abbott directed a PUCT/ERCOT audit of data-center interconnection projects (~474 GW queue, mostly data centers) versus ~70–90 GW peak grid. ERCOT Batch Zero notifications delayed from 2026-08-07; ~31 Aug 2026 status or deficiency outcomes planned. SB6 requires curtailment capability for noncritical large loads. Cipher more Batch-Zero-exposed than IREN’s advanced sites. NY (Hochul, July 2026) hyperscale moratorium hit other names. ▶ 00:59:50
xAI Colossus: ~100,000 GPUs in about four months. Musk/SpaceX/xAI plans referenced toward ~10 GW (exceed 2 GW end-2026; closer to 10 GW than 5 GW by end-2027). ▶ 00:56:07
White Fiber strategic agreement with Krambu for exclusive GPU operation on ~100 MW high-density capacity expected 2027; stock surged well over 20%. ▶ 00:04:36
MSCI index inclusion with purchases scheduled 27 Feb 2026. ▶ 00:21:30
Claims that remain inaccurate or unverifiable: IREN uniquely operating from utility grids; “only” end-to-end EPC AI-factory builder; Anthropic/OpenAI 85–90% enterprise-model share; Anthropic $100 billion ARR by end-2026; Chamath-associated token-cost doubling; BlackRock GIP/MGX $40 billion valuation benchmark; $6 billion ATM authorization; CoreWeave “40+ U.S. data centers / 250k GPUs / 3 GW / $60 billion backlog” as previously stated; PP&E liquidation at several times market cap; rear-door heat exchangers solving 100% of rack heat; TSMC manufacturing 100% of everything sold as an NVIDIA GPU; Microsoft lacking its own LLM (Phi family exists). ▶ 00:40:57

Thesis & bull case

Core edge is scarce, scalable, grid-connected land plus demonstrated construction. The market may underestimate the difficulty of converting power rights into energized, contracted AI capacity. A confirmed 5–5.8 GW secured/pipeline portfolio and a 5+ GW announced pipeline are the scale backdrop; energized IT is the scarce conversion step. ▶ 00:35:20
Vertical integration—from electricity through delivered compute, now with an NVIDIA-certified hypervisor—can improve efficiency, data security, customer flexibility, and margin capture versus bare-metal-only peers.
Horizon One acceptance is the zero-to-one proof: construction, liquid cooling, Microsoft acceptance, and NVIDIA Exemplar status landed inside the guided Q3 window. Horizons 2–4 are the refine step (December-quarter target); Horizons 5–6 are the GW-scale step. Microsoft meticulousness on the first data center is a quality-bar feature. ▶ 00:00:11
Sold-out 2026 capacity and pricing up ~125% since November, with prints above $20 million per IT MW and talks near $25 million, support scarcity pricing and a path toward Nebius-like contracted yields rather than colo-only economics. ▶ 00:09:40
Customer mix is broadening beyond one hyperscaler anchor: Cohere, Prometheus, Perplexity, FigurAI, Fal.ai, Higgsfield, and an unnamed frontier lab sit alongside Microsoft. That supports the “hyperscalers as financing anchors, not the end market” thesis. Frans’s estimate that hyperscalers could stay below 15% of long-term capacity remains outside analysis, not guidance. ▶ 00:04:21
Financing flywheel is now evidenced at both ends of the credit spectrum: ~6% investment-grade GPU debt against Microsoft and single-digit non-IG (9% on $2.4 billion) without an IG offtake, plus prepayments covering associated GPU CapEx. Keeping data centers unencumbered preserves a later asset-finance option. A speaker refinance sketch—if data centers cost ~$3 billion, refi out ~$2 billion into Sweetwater—is speculation, not a company plan, but it is directionally consistent with unencumbered shells. ▶ 00:38:31
NVIDIA’s full-stack and “AI factory” framing (tokens as physical output of electricity, chips, and data centers) makes asset-heavy enablers central. IREN is the sole announced DSX flagship partner (Sweetwater, up to 5 GW partnership). Conditional equity rights and the $3.4 billion / 60 MW cloud contract are supply and credibility catalysts. Exemplar Cloud status is achieved, not rumor. ▶ 00:04:29
Sweetwater remains the largest upside lever: phased gigawatt campus, 1.4 GW substation already energized (May 2026), dry-cooler and steel progress, multi-tenant potential, tightly interconnected training optionality. ▶ 00:18:29
Australia/APAC (Bundy ~800–850 MW nearing groundbreaking), Spain/Nostrum (~490 MW), Canada, and Oklahoma 1.6 GW are second vectors if Texas execution holds. ▶ 01:02:59
GPU longevity and payback compression (industry deals toward 1–2 years; A100s contracted through 2029) undermine obsolescence bears and leave residual-value tails after capital recovery. ▶ 00:05:47
Bitcoin-mining cash flow and a large cash balance ($7.6 billion at 30 June) reduce immediate dilution pressure even as mining is exited; impairments are the accounting cost of that exit, not evidence the AI book failed. “Why lose sleep over an impairment charge over Bitcoin mining when that’s an industry that we’re no longer participating in?” ▶ 00:12:16
Modest historical guidance is now paired with a sold-out 2026 book and named customers—bulls argue the 2026-08-27 call is where guidance and storytelling turn. Frans: “this is the earnings where things will start to turn around, at least in terms of guidance and how they are going to talk to the market.” ▶ 00:03:11
Institutional ownership (Goldman 9.4% 13G, JPMorgan) plus a cleaner operating narrative can drive the next leg if cadence holds. Extreme short interest can amplify a re-rating or a sell-the-news move; direction is not in the new sources.
Multi-winner TAM: IREN’s power/vertical path and faster commercial neoclouds can both work. Labor shortages are viewed as pay-cleared, not a structural cap. Abbott/Batch Zero noise is expected to hit developmental peers harder than already-energized or advanced IREN sites.
Valuation catch-up (speaker, 2026-08-12, not updated by a post-print price): ~4–6 months behind Nebius on revenue, higher-margin optionality from vertical integration, base case ~4x (~$160) within a year, $200+ as euphoria, private fair value argued ~$120–140. These are opinions, not guidance. ▶ 00:13:50
Frank Lee (2026-08-16): time-to-computer over time-to-revenue; post-H1 construction and financing flywheels can let IREN surpass Nebius and CoreWeave on reliable quality capacity; GW-scale complexity is an underpriced moat; near-term share-price lag framed as temporary. ↗
Frans (2026-08-21, pre-print): negatives front-loaded; AI revenue acceleration into 2027; potential $1 billion-plus AI revenue quarter by May 2027. Distinguish that quarterly revenue target from the $1 billion operating ARR already claimed—the latter does not by itself hit a $1 billion quarter. ▶ 00:53:40

Risks & bear case

Execution remains the dominant fundamental risk. Updated: Horizon One is accepted, so residual risk shifts to Horizons 2–4 (now a December-quarter target—later than bullish September hopes), billing commencement versus ARR optics, H5/H6 design, and GW-scale complexity. Cooling, piping, insurance sign-off, and customer acceptance already slipped H1 once (failed valve / fire-water audit) before acceptance. ▶ 00:28:49
FY27 CapEx of $25–30 billion dwarfs historical raises. ~$14 billion liquidity/commitments plus ~$8 billion targeted GPU funding do not obviously cover the full guide; residual equity, converts, project debt, or prepayments are still required. A weak share price makes any ATM more dilutive. Public-company storytelling matters because the stock price is a financing tool: “whether you want to be or not, you’re a salesman, and you’re selling your stock price.” ▶ 00:07:27
Non-IG GPU debt at 9% is workable but expensive versus Nebius’s ~$5 billion convertible (0.50% / 4.50% tranches). Financing advantage is proven, not cheap at every layer. ▶ 01:05:50
Nvidia GPU price hikes of ~15% and more in 2026 can inflate the $25–30 billion guide if not passed through. Pricing power above $20–25 million per MW is not guaranteed to persist for 2027–28 vintages. ▶ 00:48:30
NVIDIA backstops or equity-linked structures can secure GPUs while diluting upside economics; the company has said any structure must be weighed against that dilution. Over-association with NVIDIA remains a strategic risk if customers shift to TPUs, Trainium, or custom ASICs. ▶ 00:33:33
December-quarter bunching of Horizons 2–4 concentrates commissioning, GPU delivery, and revenue-recognition risk into one quarter after a year of slips.
Direct neocloud operation transfers utilization, hardware, software, customer-concentration, and residual-value risk to IREN. Token-output pricing would add serving, metering, and demand-forecast risk. IREN has not clearly disclosed a token commercial model; Mirantis certification narrows but does not close that gap.
Open-weight models can expand usage and also crush token prices and compute per task. ▶ 00:43:28
Colocation caps upside and can lock scarce power into fixed economics. Hybrid uncertainty at Childress (full liquid vs air-plus-liquid assist) can disappoint investors who over-read IR liquid comments.
Q4 optics are poor on purpose and still matter for sentiment: $684 million net loss and $450.4 million impairments, AI cloud only $70.5 million in the June quarter because Horizon One acceptance was in August. ARR of $1 billion operating can be mistaken for current quarterly revenue. ▶ 00:21:50
Headcount nearly tripled and >4,000 people are mobilized—scaling evidence and an SG&A risk if “SG&A baseline” proves early. ▶ 00:16:18
Governance: large founder RSUs (reports ~$788 million combined), Warriors spend reportedly >$50 million per year, crypto-heavy board optics, and prior renderings that did not fully materialize. Class B control blocks hostile discipline. ▶ 00:09:25
Commercial agility critique persists: reactive versus Nebius’s backlog machine and White Fiber’s retrofit speed, even if the August customer list helps. Delaying contracts for price can idle GPUs. The question “why no 2026 GPU buys if shells are ready?” was a stated reason H1 did not fully re-rate the stock. ▶ 00:58:23
Childress construction culture (competitive subcontractors, need-to-know silos, heavy outsourcing) can speed builds and still create late insurance/acceptance snags.
Regulatory: Abbott audit and Batch Zero (~31 Aug 2026 outcomes), SB6 curtailment, local opposition, export controls, and state moratoriums (NY). Nebius Vineland shows generator/fuel-cell paths can be stopped; IREN’s grid path is cleaner but not immune. Canadian processes are slow. ▶ 00:09:54
Competitive supply: hyperscaler proprietary silicon; xAI/SpaceX multi-GW plans; Nebius and CoreWeave winning the time-to-revenue narrative and cheap capital. ▶ 00:56:07
Macro: rates, CPI, Jackson Hole, credit conditions, and TSMC/ASML/HBM surprises can tighten financing even if IREN executes. ↗
Retail price targets (~4x in a year, $120–140 “private fair value”) are speaker opinion. High short interest sustains penny-stock volatility. H1 news was expected to floor rather than rocket the stock; whether the FY26 call changes that was not in the source material.
AI revenue does not automatically mean superior margins after power, depreciation, 9% debt, networking, staffing, and utilization. Gigawatt HPC build-cost analogies (~$50–60 billion+/GW) show how fast capital intensity scales; they are not IREN budgets.
Unverifiable rumors that should not be underwritten: Anthropic ~1.4 GW Bundy interest; NVIDIA already contracted as the next Childress customer; specific Sweetwater customer timing before 2027.

Timeline of developments

2025-08-24 - Expected AI-cloud or 50 MW Horizon news; Canadian air-cooled Blackwell retrofits and mining cash flow as a dilution buffer. ↗
2025-10-12 - Applied Digital–CoreWeave lease expansion cited as power-scarcity validation; $1 billion zero-interest convertible at 42.5% premium. ↗
2025-12-14 - Selloff toward ~$40 framed as accumulation; Microsoft as strategic validator; ~1.8 GW portfolio with ~16% leased. ↗
2026-01-12 - Triple-digit valuation debate on further hyperscaler deals; 30–60% drawdown warning. ↗
2026-01-18 - Horizon viewed ahead of schedule; xAI/Google speculation; PJM, EPA, and local opposition flagged. ↗
2026-01-27 - Miners and neoclouds framed as hyperscaler-power beneficiaries; dilution and repeated drawdowns expected. ↗
2026-02-03 - $3.6 billion GPU financing discussed at under 6% covering ~90–95% of Microsoft GPU needs; Oklahoma 1.6 GW; Sweetwater interconnections secured; Prince George cloud revenue $18 million, $400 million contracted ARR. ↗
2026-02-15 - $17 million quarterly cloud revenue, $2.8 billion cash, 4.5 GW portfolio, MSCI inclusion (27 Feb), GPU delays, ERCOT batch-zero watch. ↗
2026-04-14 - Early supercycle framing; debate on neocloud versus colocation for the next agreement. ↗
2026-04-26 - Pipeline estimated 10–12 GW through ~2035; Horizon 5–6 ~100 MW liquid-cooled opportunity discussed; Vera Rubin substitution clause. ↗
2026-05-02 - Earnings preview: soft quarter expected, liquid-cooling benefits, project-finance hires. ↗
2026-05-07 - NVIDIA equity rights (up to 30 million shares at $70 / 600,000 GPUs), $3.4 billion / 60 MW NVIDIA cloud contract, $33.6 million quarterly AI revenue, Horizon delays, Mirantis deal, Spain ~490 MW, $3.7–4.4 billion ARR ambition; up-to-5 GW DSX partnership, Sweetwater flagship. ↗
2026-05-09 - Mirantis Kubernetes, NVIDIA supply alignment, rotation chatter from AI equities into Bitcoin. ↗
2026-05-24 - Financing, construction pace, and marketing-as-rebrand defended; Horizon still seen on track for Microsoft. ↗
2026-05-26 - Motivair/cooling image overinterpreted; sequencing and site readiness prioritized. ↗
2026-05-27 - NVIDIA DSX/DGX alignment and Sweetwater as long-term catalyst. ↗
2026-06-07 - Australia, GPU pricing, and whether delayed contracting was intentional value maximization. ↗
2026-06-14 - Export-control and Texas scrutiny; diversification across Canada, Europe, Australia. ↗
2026-06-19 - Narrative shift toward sovereign/enterprise factories, hyperscalers as financing anchors. ↗
2026-06-28 - Alignment versus agreement with management; sponsorship and disclosure criticized; Horizon 1 read as near handover. ↗
2026-07-05 - Warriors sponsorship tied to NVIDIA full-stack rebrand; Microsoft framed as a learning step. ↗
2026-07-06 - Compensation, weak August optics, costs front-loaded into 2027 AI contribution. ↗
2026-07-08 - RAISE Summit: sovereign AI and tokenization; unofficial customer and exemplar chatter, much unverifiable. ↗
2026-07-12 - Token-factory thesis; hyperscalers estimated <15% of eventual capacity; Sweetwater 2 groundbreaking reported; financing menu and disclosure criticism. ↗
2026-07-16 - RPO term 17–18 months, ~20% prepayments, Prince George 22k GPUs toward $500 million ARR; B300 costs up ~30%. ↗
2026-07-27 - Horizon One handover expected within one to two weeks. ↗
2026-08-02 - Internal completion celebrated 19 July; handover hoped ~3 Aug; GPU debt discussed at blended 3% on 96% with ~45% prepayments. ↗
2026-08-07 - Horizon 1 in Microsoft acceptance (badge+PIN); SEC 5-day acceptance window; Mirantis closed 3 Aug / announced 4 Aug; ~$9.7 billion Microsoft and ~$5.8 billion Dell confirmed in discussion; Abbott audit; short interest ~93 million; kitchen-sink earnings preview. ↗
2026-08-09 - H1 handover called imminent; H2 late-September hope; Batch Zero delay framed as political theater; ARR talk above $4 billion. ↗
2026-08-12 - Nebius/CoreWeave prints as sentiment turn; A100s through 2029; colo comps (Riot ~$2.4 million/MW accurate; Fermi pricing overstated); IREN praised on power, criticized as reactive. ↗
2026-08-12 - Separate valuation space: Nebius +34% to ~$65 billion market-cap discussion; IREN ~$15 billion framed 4–6 months behind; Goldman 9.4% 13G; speaker base case ~4x. ↗
2026-08-13 - H1 slipped on fire/water-line valve (community intel); H5/H6 liquid claim only partly supported; ~$17.1 million/MW speaker calc on air-cooled B300; stock ~$43. ↗
2026-08-16 - Horizon One officially delivered ~nine months post-signing; Exemplar Cloud GB300 NVL72; time-to-computer thesis; six-institution >$500 billion NVIDIA capital platform; KKR-linked hires; Nebius Vineland halt. ↗
2026-08-21 - Stock >$42; H1 accepted with Microsoft security on site; hybrid Childress layout rumor versus IR liquid framing; H2 hoped ~one month out, H3 ~October; Sweetwater steel and utilities; Bundy nearing groundbreaking; earnings ~7 days out expected as kitchen-sink. ↗
2026-08-23 - Pre-earnings: miss and knee-jerk volatility expected, turnaround via guidance. Confirmed Thursday 27 Aug print, H1 delivered, ~5–5.8 GW secured portfolio, Sweetwater 1 energization 1 May (1.4 GW substation), ~85% of >$4 billion ARR target contracted, Warriors partnership, large founder package (reports ~$788 million). Frans leaned NVIDIA for most remaining capacity; company evaluating NVIDIA backstops versus dilution. On-demand compute not preferred versus long reserve contracts. Nebius ~$5 billion convert and GPU price hikes noted. Macro and moratoriums weighing on AI/DC names. ↗
2026-08-27 - FY26 call: $4 billion ARR contracted for 2026 capacity, $1 billion operating post-Horizon One; sold-out 2026 book; customers Cohere, Prometheus, Perplexity, FigurAI, FowlAI/Fal.ai, HiggsfieldAI, plus undisclosed frontier lab. Pricing +~125% since November, recent deals >$20 million/MW IT, talks ~$25 million. Horizons 2–4 guided to the December quarter. ~$6.5 billion GPU financing in three months (Microsoft $3.6 billion IG at ~6% funding ~96% of associated GPU CapEx; $2.8 billion non-IG including $2.4 billion at 9%). FY27 CapEx $25–30 billion; $7.6 billion cash at 30 June ($1.7 billion restricted) inside ~$14 billion liquidity/commitments; ~$8 billion more GPU funding targeted; data centers unencumbered. Q4 revenue $137.2 million, AI cloud $70.5 million, net loss $684 million, impairments $450.4 million. Targets ~300 MW IT in 2026 and ~0.5 GW more in 2027; 5+ GW pipeline. Morantis inaugural NVIDIA-certified hypervisor. Headcount nearly tripled; >4,000 mobilized. ↗

Open questions

When does Horizon One billing commence, and how fast does the $1 billion operating ARR convert into recognized quarterly AI-cloud revenue versus the June-quarter $70.5 million base?
Will Horizons 2–4 actually deliver in the December 2026 quarter, and what are phase-level acceptance and billing dates? Prior late-September hopes are superseded, not merely slipped in commentary.
How is the ~300 MW IT 2026 target split among Microsoft’s 200 MW (Horizons 1–4), Prince George/Canada, and other sites?
What is contracted versus still open for 2027’s additional ~0.5 GW IT, and what mix of hyperscaler, frontier lab, enterprise, AI-native, and sovereign is intended?
Are Horizons 5–6 liquid, hybrid air-plus-liquid, or air-cooled, and is any of that capacity contracted to NVIDIA, Microsoft, or others?
What commercial benefits follow NVIDIA Exemplar Cloud and the inaugural hypervisor certification—allocation, reference routing, or marketing only?
Will IREN pursue an NVIDIA backstop, and on what dilution, pricing, and volume terms relative to the existing 600,000-GPU equity-rights deal?
How is FY27’s $25–30 billion CapEx split among GPUs, shells, power, and software, and what funds the gap above ~$14 billion liquidity plus ~$8 billion targeted GPU financing?
What advance rates, covenants, and tenors apply to unencumbered data-center asset finance, and will Apollo/BlackRock/Blackstone/Brookfield/Goldman/KKR platforms participate?
What are contract tenors, escalators, prepayment rates, and termination rights on the new Cohere, Prometheus, Perplexity, FigurAI, Fal.ai, Higgsfield, and frontier-lab deals versus Microsoft?
Does management expect hyperscalers below 15% of long-term capacity, or was that only Frans Bakker’s estimate? Does the NVIDIA-heavy remaining-capacity lean survive the diversified August customer list?
Will IREN charge by token, and what Morantis productization timeline supports managed or token services versus GPU rental?
How do ~$20–25 million/MW IT cloud prices reconcile with colo comps (~$2.0–2.4 million average annual per MW) when choosing between direct operation and leases?
What utilization, depreciation life, residual value, power cost, and PUE assumptions support returns? Industry signal (A100s through 2029; paybacks 1–2 years) is not IREN’s accounting policy.
What is Sweetwater’s customer, density, and first-IT date now that the 1.4 GW substation is energized? Does DSX flagship status change scope or timing versus a ~2027 contribution?
How much Australian (Bundy ~800–850 MW), Spanish, Canadian, and Oklahoma capacity is energized versus prospective, and what regulatory gates remain?
Can governance confidence be restored around ~9.1 million RSUs per co-CEO and the Warriors deal (reportedly >$50 million/year)?
How will Batch Zero outcomes and the Texas data-center audit affect Sweetwater and later Childress interconnection in practice?
What is DC36 lab activity—B300 testing, customer proof, or something else—and what is the Mackenzie/McKenzie scope?
Will AI revenue exceed mining revenue on a reported basis in FY27, and can a $1 billion-plus AI revenue quarter arrive by May 2027?
How reproducible are xAI/SpaceX multi-GW plans, and do they erode IREN’s time-to-power edge?

Notable predictions to track

AI-infrastructure cycle remains early, with competition for power, land, accelerators, HBM, networking, cooling, construction, and capital. ↗
IREN narrows its valuation and operating gap with larger neoclouds if it energizes contracted capacity without excessive dilution. ↗
Usage-based inference growth creates room for independent GPU-cloud operators; open-weight models expand aggregate demand faster than efficiency cuts compute per task. ▶ 00:25:38
IREN becomes a vertically integrated token factory. Hyperscalers stay below 15% of eventual capacity—Frans Bakker’s estimate, not guidance.
Scarce compute lets IREN pass through higher GPU, HBM, and construction costs. Company evidence now includes three-year pricing +~125% since November and deals above $20 million/MW IT. Durability into 2027–28 is the open test. ▶ 00:09:40
Horizon/Microsoft zero-to-one delivery: achieved ~2026-08-13 with Exemplar Cloud status. ▶ 00:00:11
Updated: subsequent Horizons at progressively shorter intervals. Official prediction to track is now Horizons 2–4 in the December 2026 quarter—not late September. Speaker hopes of H2 ~one month from 2026-08-21 and H3 in October are superseded unless December-quarter guidance is beaten. ↗
Sweetwater is the defining long-term catalyst and NVIDIA DSX flagship, with commercial contribution discussed around 2027. Substation energization (1 May 2026) is done; IT contribution is not. ▶ 00:18:29
Guidance re-rates as backlog converts. Frans (2026-08-23) predicted this earnings starts the guidance turnaround. The call delivered sold-out 2026 capacity, named customers, pricing, and a financing model; market reaction was not in the source. ▶ 00:03:11
AI revenue exceeds bitcoin-mining revenue around the FY27 transition. A $1 billion-plus AI revenue quarter by May 2027 (Frans, 2026-08-21) is still ahead; $1 billion operating ARR is not that quarterly print. ▶ 00:53:40
August results look weak on bonuses, RSUs, depreciation, and mining impairments, with the call focused on forward guidance. Realized: Q4 revenue $137.2 million (inside the ~$120–150 million speaker band), AI cloud $70.5 million, impairments $450.4 million, net loss $684 million. ▶ 00:21:50
Mirantis becomes a higher-margin enterprise/sovereign layer. Updated: acquisition closed 2026-08-03; inaugural NVIDIA-certified hypervisor 2026-08-27. Commercial productization still to prove. ▶ 00:03:32
NVIDIA GB300 exemplar status: achieved. Rubin samples or milestones by year-end 2026 remain unofficial. ▶ 00:00:11
Asset-backed lending, prepayments, and private credit reduce—but do not eliminate—ATM and hyperscaler-credit dependence. Updated: GPU financing proof is in; data-center asset finance is the next predicted leg; >$500 billion third-party platforms may participate. ▶ 00:05:20
Margin expansion undershoots revenue growth once depreciation, 9% non-IG debt, power, and utilization are fully in the P&L.
Contracted ARR around $3.7–4.4 billion by end-2026. Updated: $4 billion ARR is already contracted for 2026 capacity, ~$1 billion operating; remaining conversion and any upside above $4 billion are the track. ▶ 00:04:43
Platform delivers ~300 MW IT in 2026 and ~0.5 GW more in 2027. ↗
FY27 CapEx of $25–30 billion is funded without a disorderly equity raise. ▶ 00:19:09
H1 acceptance floors rather than re-rates the stock; fuller catch-up versus Nebius is more a 2027 event. Still open post-call. ↗
Abbott/Batch Zero noise has limited near-term impact on IREN versus developmental peers. Batch Zero status/deficiency outcomes were due ~31 Aug 2026. ▶ 00:59:50
Longer GPU lives and 1–2 year paybacks continue to weaken two-to-three-year obsolescence bears. ▶ 00:05:47
Colo pricing at or above ~$2 million average annual per MW holds for creditworthy labs; neocloud/contracted yields in the high teens to mid-$20 millions per MW hold for IREN’s direct book. Hybrid lower-CapEx Childress layouts improve returns if validated. ▶ 00:12:57
Skilled-trades labor does not hard-cap Texas greenfield builds if pay clears the market. ↗
BitcoinAIGuy/Jeff Ward (2026-08-12): IREN a “complete steal”; ~4x (~$160) within a year, $200+ euphoria, private fair value ~$120–140; institutions drive the next leg; multi-year hold over FOMO trim. Speaker opinion. ▶ 00:13:50
Frank Lee (2026-08-16): post-H1 construction and financing flywheels let IREN surpass Nebius and CoreWeave on quality capacity; near-term lag versus time-to-revenue peers is temporary. ↗
Frans Bakker (2026-08-21/23): kitchen-sink quarter then narrative turn; H2 celebration soon after H1 (timing now in tension with December-quarter guide); Bundy breaks ground ahead of Spain/Kiowa; IREN trades as an AI-cloud name, not a miner. ▶ 00:03:11
Ghosty: possible medium-term decline toward a chart gap near $25 while remaining long-term bullish. Unresolved. ↗
Oklahoma 1.6 GW online ~2028, diversifying beyond ERCOT. ▶ 00:11:02
Canadian capacity 180–200 MW with low-cost dense retrofits; Microsoft relationship lifts GPU pricing and Canadian monetization. ↗
Commercial deal disclosure before end-August 2026: largely realized via the 27 Aug customer list, without full dollar disclosure per name. ▶ 00:19:18