CoreWeave (CRWV) vs Nebius (NBIS) vs IREN (IREN) — Comparative Investment Analysis

As of mid-October 2026 (latest reported quarters: CRWV/NBIS Q2 2026, IREN FY2026/Q4 FY2026)

Summary

All three are neocloud / AI-infrastructure builders renting GPU capacity, all are growing very fast, all are GAAP-unprofitable in the latest quarter, and all are burning billions in free cash flow to build capacity ahead of contracted demand. They differ in scale, funding model, and binding constraint:

• CoreWeave is the largest and most proven at scale, with the biggest backlog ($104B) — but also the most leveraged balance sheet, and in October 2026 its bonds were pricing real distress (2032s below 80c, yield ~15%).
• Nebius has the fastest growth, the best AI-cloud unit margins, and a prepayment-funded model that softens its debt load — but it is priced for near-perfect delivery of a $7–9B year-end ARR target, and its backlog rests on two customers.
• IREN has the most optionality (power portfolio, Microsoft + Nvidia anchor contracts, a mining business to wind down) and the lowest valuation against contracted ARR — but the weakest near-term execution proof: only ~$1B of its $4B contracted ARR was operational at FY-end, and reliability questions hit its legacy sites in October.

Framework verdict (comparative, not a personal recommendation): For a growth-oriented investor who can tolerate volatility, Nebius looks strongest overall on the combination of growth, unit economics, and funding model — provided H2 2026 delivery lands. CoreWeave is the scale/valuation alternative if credit markets stabilize; IREN is the higher-variance optionality play if acceptance cadence and reliability improve. Specific view-changing triggers are listed at the end.

Market context on the measurement date: on Oct 14, 2026, all three fell sharply (CRWV −15.5%, NBIS −10.1%, IREN −12.2%) after a Gartner warning that up to 80% of neoclouds could cease independent operations by 2030, plus hotter-than-expected September CPI (3.7%) raising rate-hike odds. This was a sector repricing, not a company-specific earnings event. Sources: https://www.fool.com/coverage/stock-market-today/2026/10/14/stock-market-today-oct-14-coreweave-plummets-on-gartner-neocloud-warning-and-bond-yield-surge/ ; https://www.abcmoney.co.uk/2026/10/coreweave-stock-plunge-shares-sink-15-on-gartner-warning/

1. What each business actually does

Sources: CRWV — https://www.theregister.com/off-prem/2026/08/12/coreweave-revenue-doubles-as-debt-pile-reaches-356b/5286832 ; NBIS — https://financefeeds.com/nebius-nbis-q2-2026-earnings-preview-capex/ ; IREN — https://www.coindesk.com/markets/2026/08/28/iren-shares-fall-8-as-costly-ai-transition-weighs-on-earnings ; https://kalkine.com/news/artificial-intelligence/why-iren-nasdaqiren-reported-a-fiscal-2026-net-loss-despite-higher-revenue-and-ai-cloud-growth

2. CoreWeave (CRWV)

Latest financials — Q2 2026 (reported Aug 11, 2026)

• Revenue $2.575B, +112% YoY, +24% QoQ; H1 2026 revenue $4.653B.
• Gross profit $1,696M, 65.9% gross margin (74.2% a year earlier). Note: other vendors quote a 71.7% FY2025 gross margin excluding server/network depreciation — definitions differ, so compare GAAP gross margins with care.
• GAAP operating loss −$49M (−1.9% operating margin); adjusted operating income $128M (~5%); adjusted EBITDA $1.51B, ~59% margin.
• Net loss −$626M (vs −$290M a year earlier); net interest expense $640M = 24.9% of revenue, up ~140% YoY.
• Operating cash flow $679M; cash purchases of PP&E/software $6,422M → OCF-minus-cash-capex −$5,743M. On management's broader capex definition ($9.4B in Q2), FCF is closer to −$7.9B. Either way: deeply negative.
• Cash $5,524M at Jun 30. Principal debt $35,551M; net debt ~$29.5B; debt + lease obligations ~$52B per secondary tally. $10B available under revolver / delayed-draw facilities at Jun 30.
• Guidance: Q3 revenue $3.45–3.6B; FY2026 revenue raised to $12.4–13.2B, capex $35–39B, year-end annualized run-rate revenue $18.5–19.5B, adjusted operating income $960M–$1.15B.

Sources: https://www.cestriancapitalresearch.com/coreweave-nasdaq-crwv-stock-rating-do-nothing-too-much-balance-sheet-risk/ ; https://www.indmoney.com/blog/us-stocks/why-is-coreweave-stock-falling ; https://www.stocktitan.net/sec-filings/CRWV/10-q-core-weave-inc-quarterly-earnings-report-d3b596a1f167.html

Debt raises / refinancing

• April 2026: $3B converts upsized to $4B at close, 1.75% coupon, due 2031/2032.
• September 2026: announced $3B converts due 2033, upsized to $3.7B priced / $4.2B closed, 2.875%, conversion $97.85 (22.5% premium), ~$499–566M spent on capped calls; plus ATM for up to 35M Class A shares. Also June $3.5B senior notes due 2032 and August $2.6B delayed-draw term loan at SOFR+550bp.
• October stress: 2032 unsecured bonds traded below 80c, yield ~15% (Bloomberg via Fool/Stocktwits); management said it is seeking investment-grade treatment for more chip-backed debt to access insurer capital.

Sources: https://www.reuters.com/legal/transactional/coreweave-launches-3-billion-convertible-debt-sale-2026-09-17/ ; https://aistockwire.com/blog/coreweave-crwv-insurance-capital-unlimited-mcbee-bond-yields-october-2026 ; https://www.aiaffairs.com/markets/coreweave-closes-upsized-4-2-billion-convertible-bond-offering/

Customer concentration & backlog

• Q2: top 3 unnamed customers 36% + 26% + 10% = 72% of revenue; H1: Customer A 40%, Customer B 23%. In 2025, Microsoft was ~67% of revenue (62% in 2024) — concentration is falling in percentage terms but still extreme.
• Revenue backlog $104.2B at Jun 30 (+246% YoY); RPO $103.7B, ~41% expected in next 24 months; plus >$25B net new commitments early in Q3 (excluded from the $104B). >50% of backlog already attached to contracts where delivery has begun; expected >2/3 by year-end.
• Named: Meta up to ~$21B new deal (total ~$35B cited in secondary), OpenAI order form up to ~$6.5B through May 2031 (~6.3% of RPO), Jane Street ~$6B, Anthropic multi-year, Nvidia ~9% stake + $6.3B take-or-pay + Vera Rubin launch privileges (secondary).

Sources: https://www.fool.com/earnings/call-transcripts/2026/08/18/coreweave-crwv-q2-2026-earnings-call-transcript/ ; https://www.theregister.com/off-prem/2026/08/12/coreweave-revenue-doubles-as-debt-pile-reaches-356b/5286832

Capacity

• 1.5 GW active power across 51 data centers at Jun 30 (+~500 MW in Q2); 3.7 GW contracted at Jun 30, 4.2 GW by Aug 11; target >8 GW active by 2030 (and >1.85 GW active by end-2026 per secondary). Indonesia (360 MW, online 2028) and India/AdaniConneX (240 MW, mid-2028) extend the footprint. Nvidia partnership targets >5 GW of AI factories by 2030.
• Source: https://blockspace.media/insight/coreweave-convertible-notes-equity-sales-2033/

Valuation (Oct 14 close)

• Price $65.78 (finance data, data/crwv.json, fetched 2026-10-15T10:10:22Z); 52-week $60.55–$153.20; 1-yr return −45.1% to Oct 13 close.
• Market cap: Finnhub quoted $45.27B (https://finnhub.io/?q=%22CRWV%22). Caution: the Q2 10-Q disclosed 551.5M shares (458.9M Class A + 92.7M Class B) at Jul 31, which at $65.78 implies ~$36.3B. Vendor market caps differ by share-count/dilution treatment — treat $36–45B as the range, not a point.
• Multiples (using the range): TTM revenue ~$7.6B → ~4.8–6.0× TTM sales; FY2026 guidance midpoint $12.8B → ~2.8–3.5× 2026 sales; year-end ARR midpoint $19B → ~1.9–2.4× exit ARR. Cheapest of the three on forward sales/ARR. Analyst consensus cited as Moderate Buy, avg target ~$140.64 (secondary, stale relative to Oct selloff) — do not anchor on it.

Biggest risks (CRWV)

1. Balance sheet / refinancing: $35.6B principal debt, interest already 25% of revenue, junk-bond yields ~15% in October; $35–39B of 2026 capex must be funded in a rising-rate market. If credit access tightens or backlog conversion slips, the model breaks first here.
2. Customer concentration + customer-competition: top 3 = 72% of revenue; Microsoft, Meta, OpenAI are also building their own compute. A lost, delayed, or repriced anchor contract would hit revenue, backlog, and financing simultaneously.
3. GPU economics / execution: rental-rate compression (H100 rates cited down 30–40% from peak in secondary), depreciation outrunning EBITDA, and the physical challenge of converting 4.2 GW contracted power into revenue-generating active power on schedule.

3. Nebius Group (NBIS)

Latest financials — Q2 2026 (reported Aug 12, 2026)

• Revenue $582.3M, +454% YoY (from $105.1M); AI Cloud $574.9M, +514%; H1 revenue $981M, +529% YoY. ARR $3.0B (+598% YoY, +56% QoQ).
• Gross margin ~77% (cost of revenue 23% of revenue, −6pp YoY). Group adjusted EBITDA $236.2M, 41% margin (vs −$21M a year earlier); AI Cloud adjusted EBITDA $285.7M, ~50% margin.
• GAAP operating loss −$175.9M (−30.2% operating margin) after D&A of $259.7M (+245% YoY) — D&A alone exceeded adjusted EBITDA. GAAP net loss −$190.4M (prior-year +$502.5M included a $597.4M revaluation gain). Depreciation lives were extended from 4 to 5 years in Jan 2026 — a key bear-case flag (Burry thesis).
• Operating cash flow $2,246M in Q2 — but this is prepayment-driven (deferred revenue rose to ~$6.0B). Capex purchases $5,657M in Q2 ($8,130M in H1) → OCF-minus-capex −$3.41B in Q2.
• Cash $8,042M at Jun 30 (vs $3,678M at Dec 31, 2025); non-current debt $8,499M; total debt ~$10.06B (TIKR); deferred revenue $5,975M.
• Guidance reaffirmed: FY2026 revenue $3.0–3.4B, year-end ARR $7–9B, group adjusted EBITDA margin ~40%, capex $20–25B.

Sources: https://financefeeds.com/nebius-nbis-q2-2026-earnings-preview-capex/ ; https://www.appeconomyinsights.com/p/neocloud-economics ; https://btw.media/en/nebius-counts-power-in-gigawatts-revenue-arrives-in-megawatts ; https://ts2.tech/en/nebius-stock-gains-2-1-despite-an-84-sell-target-as-analysts-split-by-326/

Debt raises / refinancing

• July 2026: first senior secured facility ~$775M, SOFR+2.50%, due Oct 2030, backed by deployed GPUs + contracted cash flows.
• August 2026: $5.75B converts closed — $3.45B 0.50% due 2030 + $2.3B 4.50% due 2034. Earlier converts (Jun 2025 $1B; Mar 2026 ~$4.3B at 1.25%/2.625%) convert at ~$51–$183, well below the then-share price, so they are economically equity-like if the stock holds — at the cost of dilution.
• Q2 ATM: 12.7M Class A shares at ~$223.6 avg = ~$2.8B raised; 12.3M shares remained available under the program at quarter-end. Shares outstanding were 253.9M at Mar 31 (220.4M Class A + 33.5M Class B) — the ATM alone added ~5% to the count in one quarter.

Sources: https://nebius.com/newsroom/nebius-group-announces-closing-of-private-offering-of-convertible-senior-notes-with-aggregate-gross-proceeds-of-approximately-5-75-billion ; https://www.stocktitan.net/sec-filings/NBIS/6-k-nebius-group-n-v-current-report-foreign-issuer-a9d248e84219.html

Customer concentration & backlog

• RPO ~$37.5B at Jun 30 (from ~$21B at end-2025, ~$34B at Mar 31). Contracted revenue >$40B including options; firm portion ~$32B = ~80% from two customers: Microsoft $17.4B (up to $19.4B) through 2031 (Vineland, NJ), Meta $12B firm + $15B option = up to $27B (first $3B deal Nov 2025, expanded Mar 2026, Vera Rubin).
• Q2 revenue: 3 unnamed customers 24% + 21% + 14% = 59%. Q2 signed 4 deals averaging >$1B each, including Reflection (>$1B) and Cohere — the list is widening, slowly.
• Prepayments: 70% of Q2 deals included prepayments covering 50–60% of associated capex; >$9B prepayments expected in 2026; estimated payback on Q2 deals 1 yr 10 mo (vs 2–3 yrs historically). Annual contract value rose from ~$12M/MW (2026 base) to >$20M/MW on Q2 deals; short-term deals at $40–50M/MW; on-demand prices raised ~20% on Oct 1.

Sources: https://www.fool.com/investing/2026/10/05/nebius-has-usd40-billion-of-customer-commitments-most-of-it-comes-from-2-customers/ ; https://www.fool.com/investing/2026/09/27/nebius-is-raising-the-price-of-its-ai-compute-on-oct-1-here-s-what-that-says-about-the-shortage/

Capacity

• From 170 MW at end-2025 to 800 MW–1 GW by end-2026; contracted power target raised to 5 GW (from 3–4 GW). Vineland, NJ: 300 MW phase 1 for Microsoft (community opposition / expansion hearings reported Aug–Sep 2026; planning decision pending at last report). Finland Mäntsälä scaling to 75 MW. Oct 1: +50 MW deal with AIB Data Centers (SE US); acquired Inferize for inference utilization.
• Nvidia: $2B investment Mar 2026, 9.3% stake (13G incl warrant), Exemplar Cloud status on GB300; Vera Rubin deployment via Meta deal.

Sources: https://www.fool.com/investing/2026/10/09/heres-what-1000-invested-in-nebius-could-be-worth/ ; https://en.infomaxai.com/news/articleView.html?idxno=134002

Valuation (Oct 14 close)

• Price $186.01 (data/nbis.json, fetched 2026-10-15T10:10:22Z); 52-week $73.52–$299.86; 1-yr return +52.7% to Oct 13 close — the only one of the three up over 12 months.
• Market cap: Finnhub $60.10B (https://finnhub.io/?q=%22NBIS%22). Share-count check: 253.9M shares at Mar 31 × $186 = ~$47.2B before Q2 ATM (+12.7M) and convert dilution — again, treat $47–60B as the range.
• Multiples: TTM revenue ~$1.36–1.4B → ~34–44× TTM sales (by far the most expensive on trailing sales); FY2026 guidance midpoint $3.2B → ~15–19× 2026 sales; year-end ARR midpoint $8B → ~5.9–7.5× ARR. Analyst 2027–28 estimates cited (Fool/TIKR): revenue $12.3B in 2027, $23B in 2028; at those numbers the stock would be ~4–5× 2027 sales — i.e., the valuation only normalizes if the delivery ramp lands. Rosenblatt initiated Buy at $304 on Oct 7–8; Northland cited at $415; bears at ~$83–122 (wide dispersion = high uncertainty, not a consensus).

Biggest risks (NBIS)

1. Two-customer backlog: ~80% of firm contracted revenue from Microsoft + Meta. A cancellation, delay, or failure to exercise the Meta $15B option would cut the growth story at the knees.
2. Dilution as a funding model: converts + ATM are the funding strategy. If the stock falls, the convert overhang becomes real debt and the ATM raises less per share — a reflexive loop in reverse.
3. Execution / delivery + depreciation: contracted power (5 GW) is not connected power (800 MW–1 GW by end-2026). Vineland permitting/community risk, GPU supply, and the Burry critique (5-yr lives vs 8-yr residual-value claims) all sit between backlog and GAAP profit. Q2 D&A already exceeded adjusted EBITDA.

4. IREN (IREN)

Latest financials — FY2026 (year ended Jun 30, 2026, reported Aug 27, 2026)

• Total revenue $707.0M, +41.1% (from $501.0M). AI Cloud $128.8M (from $16.4M, +687%); mining $578.2M (from $484.6M).
• Q4 FY2026: revenue $137.2M, −27% YoY, −5% QoQ; AI Cloud $70.5M vs mining $66.7M; adjusted EBITDA $19.2M, 14% margin (−68% QoQ); net loss −$684M incl $450.4M impairment.
• FY2026: net loss −$702.6M (vs +$86.9M) driven by $638.8M non-cash impairments on decommissioned mining hardware; adjusted EBITDA $245.7M, 35% margin (vs $269.7M / 54% a year earlier). Gross profit $487.3M implies ~68.9% gross margin on a pre-depreciation/impairment definition — do not compare directly to CRWV/NBIS GAAP operating margins, which use different cost treatments.
• OCF $2,100.4M (including $1,841.7M deferred-revenue/prepayment inflow); investing −$4,723.0M; financing +$9,680.1M. Approx FCF −$2.4 to −$2.6B depending on capex definition.
• Cash + restricted cash $7,619.5M ($5,895.6M cash + $1,670.3M restricted); debt $7,593.0M; deferred revenue $1,842.6M; equity $4,185.6M. Net debt on cash-only basis ~$1.7B.
• SG&A +229% in FY2026; stock-based comp ~29% of revenue (secondary) — the pivot is expensive in overhead as well as capex.

Sources: https://kalkine.com/news/artificial-intelligence/why-iren-nasdaqiren-reported-a-fiscal-2026-net-loss-despite-higher-revenue-and-ai-cloud-growth ; https://www.coindesk.com/markets/2026/08/28/iren-shares-fall-8-as-costly-ai-transition-weighs-on-earnings ; https://investwithcolby.substack.com/p/iren-limited-iren-q4-fy2026-earnings

Debt raises / refinancing

• Microsoft GPU package: $3.65B investment-grade facility (6.0% avg; $1.5B delayed-draw loan + $2.1B 5.96% senior notes due 2031, rated A / A(low)) — with prepayments covers 96% of the $5.81B GPU capex for Microsoft. This is the cleanest financing of the three on a per-contract basis.
• Second package: $2.8B for non-investment-grade customers, incl $2.4B at 9.0% fixed (Blue Owl/PIMCO, Mackenzie air-cooled expansion) — a 300bp spread that prices counterparty quality.
• Converts: Dec 2025 $2.3B (0.25%/1.00%, conversion ~$51.40) + equity to retire old low-strike converts; May 2026 $3.0B due 2033, 1.00%, conversion ~$73.07, capped call to $110.30; Sep 2026 $875M 0% due 2031 (+$125M option).
• Nvidia: right to buy up to 30M shares at $70 over 5 years, vesting only as Nvidia GPUs deploy (full vesting at 600k GPUs). $2.1B is the full-exercise value, not cash invested.

Sources: https://respectinvestment.com/investor-strategy/iren-limited-iren-stock-prediction-65-bull-26-bear/ ; https://www.stocktitan.net/sec-filings/IREN/8-k-iren-ltd-reports-material-event-7bbd0ce1dd0b.html

Customer concentration & backlog

• No customer-percentage revenue breakdown was disclosed in the sources reviewed — a disclosure gap vs CRWV/NBIS. Economically, the book is dominated by two contracts:
• Microsoft: ~$9.7B, 5-yr, GB300 capacity, 20% prepayment, $5.8B Dell GPU order; 4×50 MW Horizon at Childress; Horizon 1 (50 MW) accepted Aug 2026, Exemplar Cloud status; remaining Horizon 2–4 due later in 2026.
• Nvidia: $3.4B, 5-yr AI Cloud contract for Nvidia's own workloads (Blackwell, ~60 MW Childress) + up to 5 GW DSX partnership.
• Contracted ARR $4.0B for 2026 capacity, but only ~$1.0B operational at FY-end. Target: $3.7B ARR by end-2026, rising to $4.4B with Dell Blackwell systems (Microsoft ~$1.9B avg/yr + Nvidia ~$0.7B avg/yr + ~$1.8B planned deployments). An unconfirmed Anthropic report circulated in Oct 2026 — treat as unverified.
• Secondary tally: ~$13.8B capital commitments vs ~$7.6B cash+restricted — a funding gap to close via debt/equity/prepayments.

Sources: https://www.datacenterdynamics.com/en/news/microsoft-signs-97bn-cloud-capacity-deal-with-iren/ ; https://www.theblock.co/news/business/2026-08-17-iren-delivers-first-four-ai-cloud-deployments-microsoft-under-9-7-billion-deal-412016 ; https://www.stocktitan.net/sec-filings/IREN/8-k-iren-ltd-reports-material-event-81e2aabaad95.html ; https://bytevyte.com/iren-shares-rise-then-fade-on-unconfirmed-anthropic-cloud-deal-report/

Capacity

• 2026 target: 480 MW AI Cloud, 150,000 GPUs. 2027: 1,210 MW (730 MW under construction across BC + Texas, incl Childress + Sweetwater 1). Beyond 2027: 5 GW global power portfolio (N. America, Spain, APAC/Australia). Sweetwater: 2 GW total (1.4 GW + 0.6 GW), first 300 MW expected Q4 2027. Childress campus: 750 MW grid capacity. ~23,000 GPUs in fleet at last report (Oct), contracts secured for ~11,000 of them — the 150k target requires a steep H2 ramp.
• Note the acceptance gate: energized MW ≠ revenue. Only customer-accepted MW converts contracted ARR into recognized revenue (Horizon 1 = first 50 MW accepted).

Sources: https://www.stocktitan.net/sec-filings/IREN/8-k-iren-ltd-reports-material-event-81e2aabaad95.html ; https://WWW.FOOL.COM/investing/2026/09/10/irens-sweetwater-announcement-hints-at-a-larger-gi/ ; https://www.ainvest.com/news/iren-energized-megawatts-clear-gate-revenue-2609/

Valuation (Oct 14 close)

• Price $30.56 (data/iren.json, fetched 2026-10-15T10:10:22Z); 52-week $28.93–$76.87; 1-yr return −45.7% to Oct 13 close.
• Market cap: Finnhub $13.85B (https://finnhub.io/?q=%22IREN%22); Barchart 357.4M shares implies ~$10.9B; FinanceCharts annual 380.2M shares implies ~$11.6B. Range ~$11–14B.
• Multiples: FY2026 revenue $707M → ~15–20× TTM sales (expensive on trailing sales, but trailing sales are dominated by the shrinking mining business); contracted ARR $4.0B → ~2.7–3.5× contracted ARR (cheapest vs contracted ARR alongside CRWV); if only the ~$1B operational ARR counts, ~11–14× operational ARR. Consensus cited as Moderate Buy, avg target ~$79 (secondary) — implies >100% upside, which reflects how much of the bull case is still in future delivery.

Biggest risks (IREN)

1. Execution / reliability: SemiAnalysis (Oct 7–8) cited user reports of multi-day outages, storage/network failures at Prince George & Mackenzie, BC, calling them the worst sites by users, and challenged the "managed cloud services" claim; newer Childress/Sweetwater said to look much better. No IREN response appeared in coverage reviewed. For a company selling uptime to Microsoft/Nvidia, this is the most company-specific risk of the three.
2. Transition economics: Q4 total revenue fell 27% YoY while the pivot burned cash; adjusted EBITDA margin collapsed to 14% in Q4; $639M of impairments show how fast mining hardware value can be written off. If AI acceptance slips, the company is left with shrinking mining revenue and a large fixed build.
3. Funding / dilution + concentration: debt $7.6B, commitments ~$13.8B vs $7.6B cash+restricted, and effectively two anchor customers (Microsoft + Nvidia) account for most contracted ARR. The 9.0% non-investment-grade package shows what the market charges for non-hyperscaler risk.

5. Head-to-head comparison

Cross-cutting risks for all three: (a) GPU depreciation / residual-value debate (Burry: 5-yr lives vs 8-yr DCF residual claims; Nvidia circular-financing critique); (b) Gartner-style neocloud consolidation risk — hyperscaler customers are also competitors and could insource or squeeze pricing; (c) power/interconnection as the true bottleneck — contracted GW ≠ energized GW ≠ accepted/revenue GW; (d) rate sensitivity — all three fund multi-billion capex in debt/convert markets, so the Oct 14 CPI + bond-yield shock is a live stress test, not a hypothetical.

Sources for cross-cutting: https://stocktwits.com/news-articles/markets/equity/michael-burry-warns-nvidia-ai-infrastructure-expansion-mirroring-1960s-computer-leasing-bubble/cZD01nERB0L ; https://www.ainvest.com/news/nvidia-big-answer-burry-short-rides-number-skeptics-watch-2609/

6. Framework: which looks best overall, and what would change the view

For a growth-oriented investor who can tolerate volatility (comparative framing, not a personal recommendation):

• Best overall: Nebius. It pairs the fastest growth with the best AI-cloud margins (50%), the clearest repricing power (shorter contracts, Oct 1 price rises, $20–25M/MW on new deals vs $12M base), and the most self-funding build (prepayments covering half or more of capex, deferred revenue $6B). The valuation is demanding, but it is the only one of the three where unit economics are visibly improving and the funding model does not depend on junk-bond market access week to week.
• Runner-up: CoreWeave. If the question is "who has the most contracted demand and the cheapest price for it," CRWV wins. The $104B backlog + $25B early-Q3 commitments + 4.2 GW contracted power is a scale moat none of the others match. The reason it is not ranked first is that the binding constraint is not demand but credit: October's ~15% bond yields show what happens when the market doubts the funding model, and interest already consumes a quarter of revenue.
• Highest variance: IREN. The power portfolio (5 GW) + Microsoft/Nvidia anchors + per-contract investment-grade financing for Microsoft give it real assets and real customers. But the market is being asked to pay for $4B of contracted ARR of which only ~$1B is operational, while digesting a Q4 margin collapse and reliability allegations at legacy sites. The upside if Horizon 2–4 + Sweetwater land on schedule is large; the downside if acceptance slips is a return to a shrinking mining business with a large debt load.

What would change this view — specific developments to watch:

Next catalysts: CRWV Q3 earnings (guided $3.45–3.6B revenue) and bond-market tone; NBIS Q3 / year-end ARR print vs $7–9B guide; IREN Horizon 2–4 acceptance + Q1 FY2027 (quarter ending Sep 2026) update on operational ARR and BC-site response.

Could not verify / open questions

• Share counts / market caps: vendors disagree materially (see ranges in each section). A fully diluted share count including convertibles, warrants, ATM shares, and Class B shares was not reconciled to a single authoritative figure for any name as of Oct 15. Multiples above are therefore ranges.
• Gross-margin comparability: each company treats depreciation, stock-based comp, and (for IREN) impairments differently. GAAP operating margins are not like-for-like; adjusted EBITDA margins are the cleaner cross-company profit proxy, but they exclude the D&A that is the core bear-case issue.
• IREN customer % concentration: no revenue % by customer was found in reviewed sources — flagged as a disclosure gap.
• CoreWeave customer identities: 10-Q uses Customer A/B/C letters; mapping letters to Microsoft/OpenAI/Meta is inferred from secondary sources, not confirmed in the filing excerpts reviewed.
• Analyst price targets: cited targets ($140 CRWV, $304/$415 NBIS, $79 IREN) come from secondary aggregators and pre-date or straddle the Oct 14 selloff; they are dispersion indicators, not forecasts to rely on.
• Unconfirmed items: Anthropic–IREN contract (unverified press report); IREN BC outage claims are user reports compiled by SemiAnalysis, not audited outage data or a customer termination notice.

Sources

• Finance data (live, Oct 15, 2026): data/crwv.json, data/nbis.json, data/iren.json, data/voo.json — fetched 2026-10-15T10:10:22Z. Prices quoted are the Oct 14 closes as carried in that feed ($65.78 / $186.01 / $30.56); VOO $707.43 for market context.
• Per-company source notes with URLs and read dates: notes/source-crwv.md, notes/source-nbis.md, notes/source-iren.md.
• Key primary/near-primary sources: CoreWeave Q2 earnings call transcript (Fool); CoreWeave 10-Q via StockTitan; Nebius $5.75B converts close (nebius.com newsroom); IREN 8-K / Q3 FY26 transcript via StockTitan; Microsoft–IREN deal (Data Center Dynamics); IREN Horizon 1 delivery (The Block).

==== END DOCUMENT 1 ====