CoreWeave

Phase A — Understand the business

Lens 1 · Company Overview

CoreWeave is a specialized AI "neocloud" — it builds and rents out large-scale NVIDIA-GPU compute clusters, optimized for AI training and inference, to AI labs and hyperscalers. Not a general-purpose cloud; a purpose-built GPU utility.

• Revenue trajectory (hypergrowth): $229M (FY23) → $1.9B (FY24) → $5.1B (FY25); $2.1B in Q1'26 vs $982M Q1'25, +114% YoY. ~38% of the Q1'26 increase came from existing customers (land-and-expand), the rest from new logos.
• The backlog is the story: RPO $60.7B at Dec'25 vs $15.1B at Dec'24 — a 4× jump. Revenue is contracted, multi-year, committed — not spot. This is what bulls point to: ~12× FY25 revenue already under contract.
• Contract structure: large committed take-or-pay-style agreements with significant up-front/contractual terms; the company invoices in advance of performance under some contracts.
• Customers (concentrated, diversifying): Customer A = Microsoft 67% FY25 (62% FY24, 35% FY23); easing to 45% in Q1'26 as Customer B = OpenAI ramps to 20% (top two = 65%, down from ~77% in FY24). OpenAI total deal now $22.4B; Meta $14.2B through 2031.
• The economic model — and the catch: capex-heavy (build data centers, buy GPUs), debt-financed. Interest expense, net was $536M in Q1'26 alone (vs $264M) → a $(656)M pre-tax loss for the quarter. CoreWeave is an infrastructure-arbitrage: borrow to buy GPUs, lock in contracts, earn the spread — but the debt service is brutal and growing faster than the top line is turning profitable.

Lens 2 · Supply Chain

Upstream inputs → CoreWeave → end customer, with the chokepoints marked:

• NVIDIA (GPUs) — the critical single-source-ish input. CoreWeave is NVIDIA-backed (NVIDIA is an investor and preferential-allocation partner). GPU allocation is both its edge and its dependency..
• Server OEMs (Dell, Supermicro-class) assemble the GPU systems.
• Data-center developers / colocation + "powered shell" operators — CoreWeave leases/fits out capacity rather than owning all of it (finance leases → part of the interest load).
• ⚡ POWER is the binding constraint — CoreWeave's first listed risk factor is power access: "Our business would be harmed if we were not able to access sufficient power…". This ties straight into the Frontier-Stack thesis (compute gated by energy).
• Financing is a supply input too — growth is gated by the debt markets. The ability to keep raising capital to buy GPUs is as load-bearing as the GPUs themselves.
• Downstream: Microsoft, OpenAI, Meta (hyperscalers + frontier labs).
• Chokepoint summary: GPU allocation (NVIDIA), power/data-center capacity, and capital access — any one tightening throttles growth.

Lens 3 · Competitive Advantages (moats)

• Speed-to-deploy + scale — CoreWeave stands up large GPU clusters faster and at larger scale than rivals historically managed; first-mover in the dedicated-neocloud niche.
• NVIDIA relationship — preferential GPU access + investment. Real advantage today, but a borrowed moat: it exists at NVIDIA's discretion.
• Switching costs via the RPO — multi-year committed contracts lock customers in (the $60.7B backlog is the moat, contractually).
• Specialized AI-optimized stack vs general-purpose hyperscaler cloud.
• ⚠ Moat durability is the central debate. Bargaining power is weak: 67% of FY25 revenue came from one customer (Microsoft) that is also a competing cloud. Customers are hyperscalers who can — and do — build their own capacity. The moat is "we have the GPUs and the contracts now"; it is not obviously durable if NVIDIA broadens allocation or customers in-source. Who needs whom more is not settled in CoreWeave's favor.

Lens 4 · Segments

• Effectively a single segment — AI cloud compute. The filings do not break revenue into product segments; it's one line of business. The meaningful disaggregation is by customer, not product.
• Customer "segments": Microsoft 67%→45%, OpenAI ~20% and ramping, Meta ramping (through 2031). The trend is slow de-concentration — top-two from ~77% (FY24) to 65% (Q1'26) — healthy, but still dangerously concentrated.
• Geography: predominantly US, expanding internationally (EU data-center buildout). Not yet a material geographic split.
• Trend read: top line accelerating (+114% YoY in Q1'26), backlog 4×'ing, concentration easing — the commercial trajectory is excellent. The question every later lens has to answer is whether the financial model (debt + losses) survives long enough to harvest it.

---

Phase B — Measure performance

Lens 5 · Earnings Result (Q1'26, reported May 7 2026)

• Revenue $2.08B, +111.6% YoY — a beat vs consensus $1.97B.
• Adjusted EBITDA $1.16B, 56.0% margin (down from 62% in Q1'25). But adj EBITDA is the flattering metric — it adds back the $1,147M of D&A + the $536M interest. It is not cash the equity holder sees.
• The tell — adjusted operating income collapsed to $21M (1.0% margin), from $163M (17.0%) a year ago. After depreciation, the business barely earns operationally.
• GAAP: operating loss $(144)M → interest $(536)M → pre-tax loss $(656)M. (A ~$314.6M net loss figure circulates; reconciliation vs the $(656)M pre-tax is unclear — treat the filing's pre-tax loss as authoritative, net line to verify in Lens 10.)
• FY26 guidance: $12–13B revenue (~+145% vs FY25), adj operating income $900M–$1.1B (~8% margin — thin); Q2 $2.45–2.6B.
• Market reaction: −10% after-hours on light revenue guide + a raised 2026 capex forecast. Capacity-constrained — demand exceeds deployable supply.

Lens 6 · Earnings-Call Sentiment (trend)

• Management's consistent frame: margin compression is a timing artifact — "deployment-related expenses are incurred before revenue fully ramps". The pitch: trust the scale curve; margins normalize as the fleet matures.
• Tone trajectory: "hypergrowth + 17% adj-op margin" (early '25) → "1% adj-op margin, but temporary" (Q1'26). Increasingly a promissory narrative — burden of proof has shifted onto management, and the tape isn't extending full credit (stock fell on the guide).
• Emphasised: contracted backlog, "insatiable AI demand," capacity constraints. Downplayed: the depreciation schedule and the interest load — the two things the bears actually attack.

Lens 7 · Comps

• CRWV ~7× P/S; on enterprise value it's richer — EV ~$85B / FY25 rev $5.1B ≈ 16×; / FY26 guide $12.5B ≈ ~7× EV/S fwd. With ~$25B debt, EV/S is the honest multiple, not P/S.
• Nebius (NBIS) ~62× P/S — far pricier on sales, but +684% YoY, only ~$399M quarterly revenue (⅕ CoreWeave's scale), lower debt/equity, smaller losses. Growth-adjusted, the gap narrows.
• Analyst dispersion is itself the signal: Bernstein SELL $67 ↔ Citi Buy $158 ↔ BNP $192 — a ~3× bull/bear spread. The market has no consensus on intrinsic value, the hallmark of a levered, binary infra bet.
• True peers are thin: Nebius (closest public), private neoclouds (Crusoe, Lambda), and the hyperscalers (customers and competitors). No clean comp — which is exactly why Phase D needs a DCF, not a multiple.

Lens 8 · Stock-Price Catalysts

• Correction: CoreWeave IPO'd March 2025 (~$40), not 2026 — the census note "IPO Mar 2026" is off by a year (flag to fix research/universe/datacenters.csv).
• The arc: IPO → +330% to ~$187 (AI-infra euphoria, mid-'25) → $63.80 low (Dec 17 '25) → ~$96 (Jun '26). Roughly halved off the high.
• What actually moved it: (1) lockup expiration Sept 24 '25 → insider selling; (2) GPU-depreciation controversy — CoreWeave extended GPU useful life, lowering annual depreciation and inflating adj EBITDA/net income (Michael Burry flagged it Nov '25); (3) Q3'25 earnings miss → ~15–20% plunge; (4) credit-market stress — major holder Magnetar cut its stake, CoreWeave's CDS/credit-default pricing widened.
• The tell: the stock reacts to balance-sheet, accounting-credibility, and insider-selling signals — not revenue (which keeps beating). The market prices growth as given and survival + accounting trust as the variable. That's a map straight to where the bear case lives (Lenses 10 & 13).

---

Phase C — Judge people & books

Lens 9 · Management

• Founders are traders, not technologists. Michael Intrator (CEO), Brian Venturo (CSO), Brannin McBee — all from energy-futures trading at hedge fund Hudson Ridge. Founded Atlantic Crypto (2017, ethereum mining on NVIDIA GPUs), pivoted to AI cloud after the 2018 crypto crash, renamed CoreWeave 2019. They bought distressed GPUs in the crypto-winter and rode the AI wave.
• Track record: built a ~$50B company from "GPUs on a pool table" in ~8 years — elite execution + timing. But the DNA is leveraged-arbitrage dealmaking, which is exactly what the business is (borrow → buy GPUs → lease them). They are well-suited to the trade and less obviously suited to operating durable infrastructure.
• 🚩 Insider selling is loud: founders have sold ~$2.3B since IPO — Venturo $1.1B+, McBee $426M. Still hold ~18% (Intrator 10.4%, largest holder). Sold under 10b5-1 plans (premeditated, softens the signal) — but cashing out $2.3B while pitching "trust the scale curve" is a real tension.
• Capital allocation = the strategy: maximal debt-funded capex, no buybacks/dividends (correct — every dollar goes to GPUs). Dual-class structure: founders keep control via Class B.

Lens 10 · Forensic Red Flags

1. 🔴 GPU depreciation — a SIX-YEAR life "reflecting continuous advancements in hardware…". NVIDIA ships new architectures every ~18 months (Hopper→Blackwell→Rubin); economic life is arguably 3–4 yrs. A 6-yr schedule understates depreciation → overstates adj EBITDA and net income. This is THE forensic issue (Burry-flagged). Sensitivity: shortening to 4 yrs would add roughly 50% to annual GPU depreciation — enough to erase the "path to profitability."
2. 🔴 Free cash flow is deeply negative — a debt treadmill. FY25 operating cash +$3.1B vs investing −$10.3B, plugged by +$9.3B new financing (debt); Q1'26 investing −$7.7B. The business cannot self-fund — it runs on perpetual credit access, and that access is showing stress (CDS widening, Magnetar exiting).
3. Adj EBITDA ≠ cash. The 56% adj-EBITDA margin adds back the very depreciation + interest that ARE the economics. The chasm between adj EBITDA (~$1.16B/qtr) and the $(656)M pre-tax loss is the tell.
4. Customer concentration: Microsoft 67% (FY25) / top-2 65% (Q1'26). One renegotiation = catastrophic — and Microsoft is reportedly building its own capacity.
5. Leverage: ~$25B debt + ~$10B leases vs ~$4.8B equity; total liabilities $50.8B against $55.6B assets. Thin cushion; GPU values could impair if AI capex slows.
6. Regulatory: clean. 0 SEC enforcement findings (EDGAR EFTS LR; AAER search 500'd). The risk is accounting aggressiveness + leverage, not fraud.

Phase D — Project & stress-test

Lens 11 · Forward Projection

• Revenue (RPO-backed, high confidence): FY25 $5.1B → FY26 ~$12.5B (guide midpoint, +145%) → FY27 ~$20B (+60%) → FY28 ~$27B (+35%). Top line is the easy part — it's contracted.
• Profitability (the hard part): FY26 adj operating income guide ~$1B (~8%); GAAP still a loss. The model is dominated by two inputs — GPU depreciation life and capex intensity — and it stays FCF-negative while capex runs $10–15B/yr.
• DCF sketch (why a multiple won't do): equity value is a function of terminal margin × GPU life × discount rate (high, given leverage).
• Bull (GPUs last 6 yrs, margins → ~22% adj-op by '28, $27B rev): EV ~$100–120B → ~$130–180/sh (≈ BNP $192).
• Bear (GPUs obsolete in ~3.5 yrs, margins stay thin, FCF negative, debt compounds): impaired terminal value → equity $30–50B → ~$55–90/sh (≈ Bernstein $67).
• Base: ~$90–110/sh — roughly where it trades ($96). The market is pricing the middle; the trade is a view on which tail.
• Tracked forecast (Brier): logged — "CRWV remains FCF-negative through FY27", p=0.65 (see forecast.ts).

Lens 12 · Bull vs Bear

• Bull: $60.7B contracted backlog (~12× revenue) + capacity-constrained AI demand + NVIDIA backing → a $200B+ company if margins normalize and credit stays open. The purest public play on the AI-compute land grab.
• Bear: a leveraged, FCF-negative GPU-leasing business masking 3–4-yr economics with a 6-yr depreciation schedule; 67% concentration in a customer that's also a competitor; $2.3B founder selling; credit stress already visible. Thin equity under $50B of liabilities.
• Pre-mortem (18 mo, thesis broke): AI-capex digestion → Microsoft trims commitments → growth undershoots the RPO-implied path → CoreWeave can't refinance maturing debt at workable rates → forced capex slowdown → GPU impairments expose the 6-yr-life fiction → equity craters.
• Contrarian view: the market fixates on debt + depreciation, but the true swing is counterparty quality — the RPO is only worth what Microsoft/OpenAI/Meta actually honor. This is a leveraged bet on the AI-capex cycle continuing, wearing an infrastructure-company costume.

Lens 13 · Devil's Advocate (short-seller)

Strip the accounting: a leveraged GPU-leasing co using a 6-yr depreciation life to make 3–4-yr economics look profitable; FCF ≈ −$7B/yr, funded by debt that's repricing higher (CDS widening). 67% of revenue from a competitor (Microsoft) reportedly slowing. Founders dumped $2.3B. Four things must all hold for $96: the AI-capex supercycle runs 3+ yrs uninterrupted, the top customers honor the full RPO, credit markets stay open, and GPUs stay economically useful ~6 yrs. If growth disappoints 20–30%, the leverage turns vicious. Permanent-impairment scenario: a credit freeze (recession / capex pause) forces refinancing of ~$25B debt into a glutted, fast-depreciating-GPU market → impairments wipe the thin equity. Target $67 or below.

Lens 14 · Management Questions (ordered by information value)

1. What is the economic useful life of your GPUs, and defend 6 years when NVIDIA re-architects every ~18 months?
2. At what revenue level does the business generate positive FCF without new debt issuance?
3. What % of the $60.7B RPO is strictly non-cancelable, and what are the termination terms for your top two customers?
4. Microsoft is 45–67% of revenue and building its own capacity — what is the plan for the quarter they stop growing with you?
5. What is the weighted-average maturity and rate on your ~$25B debt, and your refinancing schedule through 2028?
6. What GPU impairment would you take if you moved to a 4-year depreciation life?
7. How much of the RPO is with OpenAI, and how is OpenAI funding its ~$22.4B commitment to you?
8. What covenants exist on your debt, and how close are you to any of them?
9. Why are the founders selling $2.3B while guiding to a margin recovery you ask the market to underwrite?
10. What is your replacement capex (vs growth capex), and how does it scale as the installed base ages?
11. What happens to utilization/pricing if a wave of new neocloud + hyperscaler capacity lands in 2027?
12. How concentrated is your power/data-center supply, and what is your secured-power runway?
13. What is your contracted vs spot revenue mix, and how does pricing reset on renewal?
14. How do you think about return on invested capital per GPU cohort, net of depreciation and financing?
15. Under what scenario do you slow growth to protect the balance sheet — and what triggers it?

---