CoreWeave, Inc. (NASDAQ: CRWV) — A Toll Bridge Built on a Credit Card

Report date: 2026-06-10 Company: CoreWeave, Inc. | Exchange/Ticker: NASDAQ: CRWV | Sector: Information Technology — Internet Services & Infrastructure (AI Cloud / “neocloud”) CIK: 0001769628 | Fiscal year end: December | IPO: 2025-03-28 Market data (2026-06-10): Price ~$98.45 | Market cap ~$59.8B | Enterprise value ~$92.6B | ~447.6M shares (Class A + B) | Short interest ~17% of float | 52-wk range $63.80–$187.00

⚡ Claude’s Take

This block is the author’s own independent, subjective opinion and general information only — not investment advice. The analysis that follows is deliberately position-free and carries no recommendation or price target; this opening block is the single place a view is expressed.

Verdict: AVOID the equity here — and resist the temptation to short it outright. A genuinely extraordinary business strapped to a balance sheet that converts a great demand story into a leveraged bet on cheap capital and a 6-year depreciation assumption. Not ownable for a quality-and-returns investor at ~$98; only worth a hard look deep in the $50s–60s and with the funding/depreciation risks visibly contained.

CoreWeave is the best-executed independent AI-cloud operator in the world — the only SemiAnalysis “Platinum” neocloud, ~5× larger than its nearest public peer, NVIDIA-blessed, with a $99B take-or-pay backlog and customers (OpenAI, Microsoft, Meta, Anthropic) that are the whole AI economy. That is not the problem. The problem is the return structure. This is a leveraged spread business: borrow at 9–15%, buy GPUs depreciated over an aggressive 6 years, and rent them under multi-year contracts whose customer prepayments make operating cash flow look positive while free cash flow runs −$7.3B (FY25) and −$4.7B in a single quarter (Q1’26). On GAAP and on cash, it loses money and destroys capital today; the entire bull case rests on a future GAAP operating margin the company has never posted at scale, and on a depreciation schedule that is the single largest cost line and the one most exposed to GPU obsolescence. By the author’s standards — durable moat, returns on capital, capital allocation as the bridge to shareholder value — it fails the first two outright (no customer captivity at the dominant revenue layer; negative ROIC) and the third is a near-monthly debt raise with ~$3.4B of insider selling and zero open-market buys behind it.

What stops this from being a clean short is that the valuation is not euphoric — the stock is already ~47% off its high, trades at ~0.93× contracted backlog and near the cheap end of its own (brief) history, ~17% of the float is already short, and the near-term backlog is contractually real and sold out. The asymmetry is simply poor at this price: my illustrative base case (~$80–110B EV) brackets today’s ~$92.6B, so the buyer is paying roughly fair value for a flawless execution-and-funding path while wearing a bear case (~$30–50B EV, i.e. ~50–65% downside) that only needs one of {pricing rolls over, a 4-year depreciation reality, a funding-market hiccup, a Microsoft renewal trim} to land. Framing: a falling knife / cyclical-at-the-peak, not a value opportunity. Conviction: medium. The single datum that flips me bullish: GAAP operating margin inflecting to low-double-digits by Q4’26 with A100/H100 pricing still firm and funding still mostly non-dilutive debt at a falling rate. The single datum that flips me decisively bearish: any quarter where the ATM/equity becomes a primary funding source, or the GPU depreciable life is lengthened to manufacture margin. Tag: “A toll bridge built on a credit card.”

1. Executive Summary

CoreWeave is the scaled leader of the independent “neocloud” layer — a purpose-built AI cloud that procures NVIDIA GPUs at the frontier, builds and operates data centers to house them, wraps them in genuinely best-in-class orchestration software (Mission Control, SUNK, Weights & Biases), and rents the resulting compute under long-dated, take-or-pay contracts. The operational record is exceptional: revenue grew $229M (FY23) → $1.92B (FY24) → $5.13B (FY25) → a $2.08B quarter in Q1’26 (+112% YoY), the contracted backlog (RPO) reached $99.4B, and SemiAnalysis rates CoreWeave the sole Platinum provider. Management guides FY26 revenue of $12–13B and exit-2026 ARR of $18–19B, with >75% of a >$30B exit-2027 ARR already contracted.

The financial structure, however, is the opposite of the operational story. The widely-cited “~72% gross margin” is an artifact of an accounting policy that books GPU depreciation outside cost of revenue; loaded back in, the business posts a −$46M GAAP operating loss (FY25) and a −$1,167M net loss, the gap driven by $1,229M of interest expense on ~$25–28B of debt priced at 9–15%. Free cash flow was −$7.25B in FY25 and −$4.7B in Q1’26 alone; the only positive cash-flow metric — operating cash flow — is fully manufactured by +$4.2B of customer prepayments (deferred revenue). ROE and ROIC are negative. The business does not self-fund and is structurally dependent on perpetual, cheap capital-markets access.

The competitive verdict is that CoreWeave is a best-in-class operator without a durable moat. Its advantages — privileged NVIDIA allocation, a superior software/cost stack, and scale — map in Greenwald’s taxonomy to a supply + cost advantage without customer captivity, the weakest and least durable of the genuine advantage types. Switching costs are low precisely where the revenue is (bare-metal training compute is fungible), customer concentration is extreme (Microsoft ~67% of FY25 revenue), and the structural threat is not other neoclouds but the company’s own hyperscaler customers in-sourcing — sharpened by the May-2026 Google/Blackstone TPU-based AI-cloud JV. The capital-cycle (Marathon) read is textbook mid-euphoria: rising prices on six-year-old GPUs and sold-out capacity are exactly the conditions that precede supply-driven mean reversion.

This memo takes no position and sets no price target. It frames valuation as embedded expectations: at ~$92.6B EV (≈0.93× backlog, ≈7× FY26 revenue), the market is underwriting on-schedule conversion of the contracted backlog at a true, depreciation-bearing operating margin in the mid-20s%, funded largely by debt at a falling cost — a demanding conjunction, though notably not one that requires heroics beyond what is already signed. The decisive open questions are the honesty of the 6-year GPU depreciable life and the terms of the next ~$20B of financing.

2. Business Overview

What CoreWeave is. CoreWeave is an AI-compute rental business with a software optimization layer — a “neocloud,” not a diversified cloud. It procures NVIDIA GPUs at scale, builds/leases and operates the data centers, and rents the compute. The 10-K (filed 2026-03-02) describes three delivery layers: Infrastructure Services (the GPU compute itself — the overwhelming majority of revenue), Managed Software Services, and Application Software Services. (FACT)

The product stack is the genuine differentiator management leans on:

• CoreWeave Mission Control — proprietary orchestration, automation, observability; provisions GPUs, schedules workloads, predicts and routes around hardware failure. NVIDIA qualified CoreWeave’s software as a reference architecture in Q1’26 — a meaningful third-party validation. (FACT — Q1’26 call, 2026-05-07)
• SUNK (Slurm-on-Kubernetes) for large-scale training; a Kubernetes-native control plane.
• LOTA (Local Object Transport Accelerator) — node-local storage caching that cuts egress cost and raises throughput.
• Weights & Biases (acquired May 2025) — MLOps / experiment-tracking, the highest-stickiness asset in the portfolio.
• Newer surface area: CoreWeave Interconnect (with Google Cloud), SUNK Anywhere, LOTA Cross-Cloud, Flex Reservation/Spot pricing (introduced Q1’26, “immediately oversubscribed”), and CoreWeave Omni (the full stack deployed inside a customer’s own data center). (FACT)

Revenue model — take-or-pay. The core contract commits a customer to a fixed price per GPU-hour for 4–6 years (weighted-average ~5 years) whether or not the capacity is used. (FACT — 10-K; BofA conference 2026-06-03) These contracts are also CoreWeave’s financing instrument: they are pledged to the asset-backed debt market to borrow at rates “way closer to our customers than our own.” Revenue is recognized straight-line over the contract life only once capacity is deployed and handed to the customer — which is why revenue steps up as data halls go live and why margins are timing-distorted during ramp. The company supplements with higher-ASP on-demand and spot, and is deliberately “twisting the dial” toward shorter-dated exposure on already-financed gear to capture pricing upside. (FACT)

Segmentation & geography. Effectively single-segment AI cloud, no reportable segment breakout beyond the service-line description. Overwhelmingly US — 1,967 of 2,189 employees are US-based. (FACT) Revenue is contracted but not “recurring” in a SaaS sense — long-dated infrastructure rental to a handful of counterparties, closer to a leasing book than a subscription annuity. The software products (storage, CPU, networking, W&B) are each guided to exceed $100M ARR by end-2026 — small against a $12–13B revenue base, but the genuinely repeatable/sticky slice. (FACT)

Customer concentration — quantified and severe.

• Microsoft (Customer A) = ~67% of FY2025 revenue (62% in FY24). (FACT — 10-K Customer Concentration note)
• Top two customers ≈ 77% of FY2024 revenue. (FACT)
• OpenAI signed an MSA (2025) committing up to ~$11.9B through Oct 2030, plus a Sept-2025 order form of ~$6.5B through May 2031. (FACT)
• Meta added a $21B agreement (announced April 2026); Anthropic was added as a customer in Q1’26; Jane Street added $6B of capacity in Q1’26. (FACT)

Management’s counter-narrative is rapid diversification: Microsoft is “not even our second largest customer” today (was 85% of backlog at IPO), 10 customers are committed at ≥$1B each, and the financial-services vertical is “approaching $10 billion” of backlog. (FACT — BofA 2026-06-03) This is real, but the FY2025 recognized-revenue base is still 67% one customer — the forward backlog is diversifying faster than the income statement. (INTERPRETATION)

Backlog. RPO was $60.7B at 2025-12-31 (vs. $15.1B a year earlier) and $98.8B at Q1’26; the broader “revenue backlog” reached $99.4B after >$40B of new Q1 commitments. Backlog is near-term-weighted: 36% expected within 24 months, 75% within four years. (FACT)

The crypto-mining origin. Founded September 2017 as “Atlantic Crypto,” CoreWeave derived most pre-2022 revenue from (now-discontinued) crypto mining, launched the CoreWeave Cloud platform in 2020, and pivoted hard into AI/GPU cloud. (FACT — 10-K) The relevance: a company barely six years removed from a discarded business model now carries ~$25–28B of debt and has never traversed a full demand cycle in its current form. (INTERPRETATION)

Verdict. A capital-intensive, take-or-pay GPU-rental business with a genuinely strong software/orchestration wrapper and the deepest blue-chip roster among independents — but with extreme single-customer concentration, single-segment exposure to one input (NVIDIA GPUs) and one bottleneck (power), and revenue that is contracted rather than recurring. The full-stack story is real as a cost-and-reliability advantage; it does not yet convert the business into a software-margin franchise.

3. Industry Dynamics

The build-out. AI compute is a once-in-a-generation capex super-cycle: combined Big-Tech 2026 capex guidance runs ~$635–725B (more than 2× 2024); Alphabet alone guides up to ~$185B for 2026; CoreWeave itself guides $31–35B of 2026 capex after $14.9B in 2025. The value chain runs NVIDIA (silicon) → neoclouds/hyperscalers (capacity) → AI labs (models) → enterprises (applications). (FACT)

Where the profit pool sits — overwhelmingly at NVIDIA, which sells GPUs to everyone at ~75% gross margin and controls allocation and product cadence. The rental layer earns a spread — transiently fat while supply is scarce, structurally thin once it normalizes. CoreWeave’s reported ~72% “gross margin” is misleading on its own: the real cost lives in “Technology & Infrastructure” ($2,929M, 57% of FY25 revenue), which carries depreciation, lease, and power. After it, GAAP operating loss was −$46M and net loss −$1,167M — the gap being interest. (FACT/INTERPRETATION) This is the tell that the economics are a leveraged real-estate/leasing model, not a software-margin business.

Power is the true bottleneck, not GPUs. New grid-connected capacity takes 3–5 years; transformer, switchgear, and turbine shortages add 24–72-month delays. Management is explicit the constraint “isn’t just power, it’s labor, it’s memory, it’s storage, it’s our ability to bring up infrastructure.” (FACT — Q1’26 call) Power scarcity is what currently props up GPU rental pricing and could extend the favorable phase 2–4 years — a cyclical, not structural, support. (INTERPRETATION)

Barriers to entry are capital/power barriers, not franchise barriers. Entry requires NVIDIA allocation, multi-GW power, land, and tens of billions of financing — which deter sub-scale entrants but emphatically not the hyperscalers (AWS, Azure, GCP, Oracle OCI), who are vastly better capitalized and who in-source. Management’s own count: only six companies in the Western world deliver AI cloud at scale — Microsoft, Meta, Google, Amazon, Oracle, and CoreWeave — and five of them have “gigantic other beautiful businesses that obscure away the economics.” CoreWeave is the one pure-play; that is both its appeal and its fragility. (FACT/INTERPRETATION)

Capital-cycle read (Marathon lens). Textbook mid-euphoria: extraordinary current returns (Q1’26 56% adjusted-EBITDA margin) plus abundant cheap capital drawing a flood of supply. Marathon’s framework predicts mean reversion — high returns attract capital, capital builds capacity, capacity compresses pricing and returns. The closest historical analogues are the 1999–2001 telecom/fiber build-out and the crypto-mining cycles: demand was real, the technology transformative, the early operators earned huge returns — and the capacity-owning middle layer was where capital was destroyed when supply overshot. The GPU case is arguably worse than fiber, because GPUs depreciate and must be perpetually replaced — a reinvestment treadmill dark fiber never had. (INTERPRETATION)

Hyperscaler in-sourcing — the structural threat. Management rebuts the “temporary, we’ll bring it in-house” framing with two arguments: hyperscalers have never historically shrunk a revenue-generating data-center footprint, and since AI-cloud capex is ~2/3 “what goes inside” vs. 1/3 shell, the leased portion is more attractive to keep outsourced. (INTERPRETATION — management’s, and coherent.) But it remains a hypothesis that cuts against the plain fact that every major customer (Microsoft, Meta, Google, Amazon) is simultaneously a potential in-sourcer with a structural cost-of-capital advantage.

The Google/Blackstone AI-cloud JV (May 2026) — targeting ~500MW by 2027, running heavily on Google’s own TPUs (bypassing NVIDIA), aimed at a market estimated at ~$400B by 2031 — adds a deep-pocketed, vertically-integrated, TPU-armed supplier precisely to the layer where CoreWeave’s pricing power is least defensible. Bernstein’s read: it “will compress CoreWeave’s pricing power and margins on new contracts.” (reported in the financial press, May 2026) Not a near-term displacement (CRWV has a multi-year head start), but a medium-term margin headwind. (INTERPRETATION)

Verdict. Structurally mediocre-to-poor for the rental layer. High capital intensity, price-taking on the two key inputs (GPUs and power), a commoditizing core service, and a supply wave that should compress ASPs. The industry is attractive only while power scarcity persists and demand outruns supply — true today, but a cyclical condition, not a structural moat. The durable value most likely accrues to the toll-takers (NVIDIA, power/grid) and the application layer — not to the capacity renters in the middle. CoreWeave is the best-positioned player in a structurally challenged layer.

4. Competitive Position

Scale reality — CRWV is the clear independent leader. It is ~5× Nebius’s revenue (Q1’26 ~$2.1B vs ~$399M), with a $99.4B backlog / $60.7B RPO vs Nebius’s $21.3B; it is the sole Platinum ClusterMAX provider (SemiAnalysis) and the only independent serving OpenAI, Anthropic, Meta, Google, Microsoft, and NVIDIA plus the next tier (Cursor, Cognition, Perplexity, Cohere, Mistral). (FACT — BofA 2026-06-03; SemiAnalysis via NBIS report ) The flip side: it carries ~$25–28B of debt and a >$2.1B interest run-rate — the opposite balance-sheet posture from net-cash Nebius. CoreWeave bought its scale lead with leverage.

Greenwald moat-type analysis. The candidate advantages map to:

1. Supply advantage (NVIDIA allocation/early access). First-to-market on GB200/GB300, NVIDIA’s $2B equity investment, reference-architecture qualification, 5GW line-of-sight. Genuine and valuable while GPUs are scarce — but non-exclusive: NVIDIA also backs Nebius, Crusoe, Lambda, and Nscale. Management itself downplays it: “We are capable of securing infrastructure at gargantuan scale, independent of any support from NVIDIA… [the 5GW] gives us the ability on an opportunistic basis to accelerate.” (FACT — Q1’26 call) A rentable supply edge, not a proprietary moat.

2. Cost/efficiency advantage (purpose-built stack). The strongest operational claim — power-shell-to-live in ~6 weeks vs. 3–6 months for rivals, no virtualization overhead, Mission Control raising utilization and GPU longevity. This shows up in the Platinum rating and the mid-20s% contribution margins. (FACT/INTERPRETATION) But a cost advantage in a commodity is a margin edge, not a franchise — replicable by hyperscalers and well-funded entrants who match the vertical integration. In Greenwald’s terms, a cost advantage without captivity erodes as competitors copy the cost structure; it does not compound.

3. Switching costs / customer captivity — the crux. At the bare-metal/training layer (most revenue) switching costs are low — GPU compute is fungible and sophisticated buyers multi-source. CoreWeave’s stickiness evidence: 90% of reserved-instance customers use ≥2 products, 75% use ≥3, and Weights & Biases / storage / networking embed workflows. (FACT — Q1’26 call) This is moderate and rising stickiness, concentrated in the smaller software/managed layers — exactly where the hyperscalers (Bedrock, Vertex, Azure AI) are racing to commoditize.

Best-fit classification: a supply + cost advantage without durable customer captivity — the weakest and least-durable of the three genuine advantage types, and the one that does not survive scale competition from better-capitalized rivals. There is no economies-of-scale-plus-captivity (the only durable combination), because there is no captivity at the dominant revenue layer. (INTERPRETATION)

Pressure-testing the model. Management’s own residual-value framing reveals the structure: a 5-year take-or-pay contract “is going to pay for all the financing costs and all the CapEx… and pay for another 5 years of data center expense on top of that,” after which “you’re going to own infrastructure that is your own to monetize” at spot. (FACT — BofA 2026-06-03) This is a capacity-arbitrage model — buy financed by the customer, then re-rent the depreciated asset — not a captive-customer model. The customer is locked in for the term by contract, not by switching cost; at renewal they are free to leave, and the entire model depends on a healthy resale/spot market for aging GPUs. That is exactly what oversupply (or a Google-TPU price war) would break.

The moat-to-financial-outcome test (the author standard). If CoreWeave’s software stack vanished tomorrow, could its bare-metal customers rent equivalent NVIDIA compute from Nebius, Lambda, Oracle, or Azure with limited disruption? At the training layer, yes — outcomes would not durably deteriorate. The claim survives only at the W&B/MLOps/inference-tooling layer, which is small (sub-$100M ARR products) and contested. By the the author rule, a “moat” that fails this test at the dominant revenue layer is not a moat. (INTERPRETATION)

Where the moat could become real. Three paths: (1) persistent multi-year power scarcity that caps supply; (2) the software/inference/MLOps layer growing into a sticky, franchise-margin share of revenue before hyperscalers commoditize it; (3) self-build data centers (first online “later this year”) converting a lease-cost edge into an owned-asset cost advantage. All three are plausible and in progress; none is proven through a downturn.

Verdict. No durable competitive advantage in the Greenwald sense — a best-in-class operator without customer captivity in a commoditizing rental layer. CoreWeave has the strongest operational and scale position among independents (Platinum ClusterMAX, fastest deployment, deepest backlog, broadest blue-chip roster, improving cost of capital). But the differentiation is a margin advantage, not a franchise: NVIDIA supply is shared, cost edges are replicable, and switching costs are low precisely where the revenue is. On current evidence, CRWV is the best house on a structurally cyclical street, owning a leveraged capacity-arbitrage book whose returns are real today and contingent tomorrow.

5. Growth History and Forward Opportunities

The trajectory is almost without precedent for an infrastructure business: revenue $229M (FY23) → $1.92B (FY24, +738%) → $5.13B (FY25, +168%) → $2.08B in Q1’26 (+112% YoY, +32% QoQ) — “the fastest cloud in history to reach $5 billion in annual revenue.” (FACT — Q4’25 call)

Which lever is binding — capacity, not demand or price. This is the most important growth-quality datum:

• Management states CoreWeave is “largely sold out for near-term capacity across our fleet” and “remain sold out of our 2026 capacity,” now extending into 2027. (FACT — Q1’26 call)
• Pricing is rising, even on older silicon: “Average pricing for the A100s, H100s and H200s and L40s all increased quarter-over-quarter”; “average H100 pricing in Q4 was within 10% of where it started the year, while average A100 pricing increased in 2025.” (FACT — Q1’26 / Q4’25 calls)
• Growth is therefore supply-gated, paced by megawatts delivered. A rental business raising prices on six-year-old GPUs is the cleanest possible signal of a capital-cycle peak: good for near-term growth, a flag for durability. (INTERPRETATION)

Capacity: active power crossed 1 GW in Q1’26 (from ~50 MW historically), contracted power >3.5 GW, target >1.7 GW active by end-2026 and >8 GW by 2030. (FACT — Q1’26 call)

The RPO → revenue conversion bet is the whole thesis in one number. Backlog $66.8B (end-FY25) → $99.4B (Q1’26), >$40B of new commitments in Q1 alone; 36% recognizable within 24 months, 75% within four years. (FACT) Unlike a SaaS pipeline, this backlog is contractually committed, take-or-pay, and multi-year — which is why management can guide ARR years out. But converting it requires CoreWeave to (a) physically build ~$30B/yr of data centers on schedule, (b) finance that build, and © keep every counterparty solvent and willing across 4–6 years. The backlog is demand-derisked; it is not execution- or funding-derisked. (INTERPRETATION)

FY26 guidance (exact). Revenue $12–13B (~140% growth at midpoint); adjusted operating income $900M–$1.1B; capex $31–35B (raised); exit-2026 ARR $18–19B; exit-2027 ARR >$30B, >75% already contracted. Q2’26: revenue $2.45–2.6B, adj. op income $30–90M, interest expense $650–730M, capex $7–9B. (FACT — Q1’26 call)

The margin J-curve is the most misunderstood number. Q1’26 adjusted operating margin was 1% — explicitly the trough — even as adjusted EBITDA margin was 56%. The gap is depreciation + lease/power on newly-deployed capacity not yet generating revenue: 1–2 months of “costs but no revenue” at negative contribution margins, then “contribution margins normalizing in the mid-20s” by month 3. Management guides adj. op margin to low-double-digits by Q4’26 and a long-run 25–30%. (FACT — Q1’26 call) The “timing not economic” claim is credible at the contract level but unproven at the consolidated GAAP level — the company has never posted GAAP operating profit at scale. (INTERPRETATION — rule 8)

Forward levers. (1) OpenAI/Meta/Anthropic ramp (largely 2027-weighted). (2) Inference inflection — management estimates >50% of compute is now inference, which is stickier, spreads demand across GPU generations (supporting the older-SKU pricing), and underwrites the depreciation defense. (3) Software/platform ARR — storage already >$100M ARR with 80% attach in large customers; software, CPU, networking each >$100M ARR by end-2026; CoreWeave Omni and SUNK licensing expand TAM beyond owned capacity (excluded from guidance — optionality). (4) Self-build sites (first online later this year). (5) Enterprise / financial-services / international — the FS vertical (~$10B backlog) is the antidote to the concentration bear case. (FACT — Q1’26 / Q4’25 calls)

Verdict. High-quality on visibility and demand-backing; low-quality on returns and self-funding. The growth is genuinely contracted, diversifying away from the Microsoft anchor, and accelerating — rare and real. But it is capital-cycle-dependent (the sold-out, rising-price condition is what Marathon flags as mean-reverting), execution-fragile (every dollar requires a megawatt delivered on time amid a global component shortage), and not self-funding (FY25 FCF −$7.25B against $31–35B of 2026 capex). The label that fits: demand-backed but balance-sheet-financed hyper-growth — a high-quality top line resting on an unproven return structure.

6. Financial Quality

Multi-year income statement (FACT — 10-K Statement of Operations; Q1’26 10-Q; revenue XBRL-confirmed):

The decisive feature: FY2024 briefly produced +$324M operating income; FY2025 swung back to a −$46M operating loss — not because revenue stalled (+168%) but because depreciation tripled as the GPU build caught up. Below the operating line, $1,229M of interest expense turns the small operating loss into a −$1,167M net loss. This is the signature of a leveraged spread business, not an operating-leverage story. (INTERPRETATION)

The “72% gross margin” is a misleading subtotal. CoreWeave reports no gross-profit line. Per its own accounting policy, GPU/server/networking depreciation is booked in “Technology & Infrastructure” ($2,929M, 57% of FY25 revenue), while “cost of revenue” carries only facility rent, power, and data-center personnel. (FACT — 10-K cost-classification note) The single largest economic cost of delivering compute — depreciation of a fleet that loses ~$2.4B of book value a year — sits outside the line analysts use to compute “gross margin.” Load it back in and the business is loss-making at every level below revenue.

The depreciation wall and the 6-year GPU life (FACT — PP&E note):

Technology equipment is depreciated straight-line over 6 years (raised from 5 years in 2023) — the most aggressive in the peer set (Nebius uses 4 years). (FACT) Two flattering effects: (1) NVIDIA ships a new flagship roughly annually; a 6-year book life on assets whose economic competitiveness may erode in 2–3 years understates depreciation — a 4-year life would add >$1B of annual depreciation at the current ~$21B equipment base, deepening the net loss materially. (2) $9.4B (28%) of gross P&E was construction-in-progress at FY25 — built and capitalized but not yet depreciating — so reported D&A understates the eventual run-rate as that capacity energizes. The depreciation wall is ahead, not behind. (INTERPRETATION)

Cash-flow quality — “operating cash flow” is a prepayment artifact (FACT — Cash Flow Statement):

Three QoE red flags: (1) OCF is manufactured by deferred revenue — FY25 OCF of $3,058M is more than fully explained by the +$4,174M swing in customer prepayments; strip it and underlying operating cash is negative. (2) FCF is catastrophically negative and worsening — −$7.25B (FY25), −$4.7B in Q1’26 alone. (3) True capex exceeds the cash figure — FY25 also carried $11,151M of non-cash, OEM-financed equipment additions on top of the $10.3B cash capex, so economic capex was ~$21B, financed largely by vendor financing that then appears as debt. The $10.3B cash-capex line understates real capital intensity by roughly half. (FACT/INTERPRETATION)

Returns. ROE and ROIC are negative — no positive return on capital. Stockholders’ equity was negative (−$414M) at FY24-end and is positive ($3,335M) at FY25 only because of IPO proceeds and preferred conversion, not retained earnings (accumulated deficit −$2,643M). (FACT)

Unit-economics sketch (INTERPRETATION/ASSUMPTION): with ~$21B of technology equipment on a 6-year life, annual GPU depreciation alone is ~$3.5B run-rate, and annual interest is >$2.1B run-rate (Q1’26 $536M ×4). Against ~$8.3B annualized revenue, depreciation + interest consume ~$5.6B — roughly two-thirds of revenue before facility, power, and personnel. The cluster can be a good investment at peak pricing and a poor one at trough pricing; adjusted EBITDA (which sits above both depreciation and interest) is structurally silent on which regime you are funding into. This is why anchoring on GAAP operating losses and cash burn, not adjusted EBITDA, matters.

Verdict. Economics do not improve with scale — they deteriorate as depreciation and interest compound. Revenue scaled 22× in three years yet the business moved from a small operating profit back to operating and net losses, FCF deepened to −$7B+, and the only positive cash metric is borrowed-forward customer prepayments. The reported margin structure is flattered by excluding GPU depreciation from “gross margin,” by a 6-year GPU life, and by a large not-yet-depreciating CIP balance. Anchored on GAAP and cash, this is a loss-making, capital-consuming enterprise at present, with a plausible but unproven path to operating leverage if ARR scales into the fixed-cost base and pricing holds.

7. Capital Allocation

The model is a near-monthly debt raise. Debt by instrument (FACT — Debt Note; figures Q1’26 unless noted):

Plus $2.8B of 9.75% Senior Notes issued April 2026 (post-quarter). MD&A cites total indebtedness ~$25.1B at 3/31/26 with ~$8.8B undrawn. (FACT) Effective rates of 9–15% (DDTL 1.0 at ~15%, newest notes at 9.75%) are high-yield/venture pricing; the facilities are collateralized by the GPUs and pledged contractual cash flows, with DDTL 4.0 a non-recourse SPV carrying project-level debt-service-coverage tests and mandatory rate/power hedging. The GPUs and the customer contracts are encumbered. (FACT/INTERPRETATION)

The interest burden. Net interest expense $349M (FY24) → $1,148M (FY25); cash interest paid $869M (FY25); Q1’26 P&L interest $536M, annualizing >$2.1B. CoreWeave also capitalizes interest (~$182M FY25) into PP&E, which both understates reported interest and inflates the depreciable base — a second QoE flatter. (FACT)

Off-balance-sheet leases — a large parallel obligation. Operating lease liabilities (PV) ~$8.2B, finance leases ~$0.25B; undiscounted operating lease payments ~$13.6B; and $38.5B of leases not yet commenced (commencing 2026–2029), off the balance sheet today — including a single site with $13.5–14.4B of rent over 16 years. Total contractual obligations dwarf the equity base and cash. (FACT — Lease Note)

The funding gap is the central capital-allocation question. Cash fell from $3.13B (FY25) to $2.24B (Q1’26); Q1’26 FCF burn was −$4.7B in a single quarter. Against ~$2.2B unrestricted cash + ~$8.8B undrawn facilities, the company has roughly two-to-three quarters of runway at the Q1’26 burn rate before needing fresh capital — which it has raised more or less continuously ($2.0B NVIDIA equity Jan’26; $2.8B notes Apr’26). With FY26 capex guided to $31–35B, the model is predicated on permanent, cheap capital-markets access. Any tightening, or a downgrade-driven spread widening, is an existential risk, not a margin risk. (INTERPRETATION)

M&A — the failed vertical-integration bet. The marquee capital-allocation event of the period was the all-stock acquisition of Core Scientific (CORZ) — signed July 2025 to bring power/data-center real estate in-house — which was terminated on 2025-10-31 after CORZ shareholders rejected the deal. (FACT — 8-K) CoreWeave did not secure the captive power it sought, leaving it reliant on the $38.5B of leased capacity. The Weights & Biases acquisition (May 2025) was a smaller, sensible capability buy. (INTERPRETATION)

Dilution and SBC. SBC surged to $630M (FY25) from $31M (FY24) — ~12% of revenue, real and recurring. The share base grew sharply (WA diluted shares 218M → 435M → 532M), reflecting IPO, preferred conversion, the NVIDIA placement, and convert overhang. (FACT)

Insider signal — strongly negative. A direct parse of all 357 post-listing Form 4s (4,138 transactions) shows ~$9.0B of open-market sales since Nov-2024; excluding sponsor Magnetar (~$5.5B), insiders/officers/directors sold ~$3.44B, of which the three founders alone sold ~$2.58B (Venturo ~$1.24B, McBee ~$0.79B, CEO Intrator ~$0.53B). (FACT) This validates and exceeds the press “$2.3B insider selling” figure. There were zero post-IPO open-market purchases by management — the only code-P buys in the corpus were pre-IPO private placements at ~$47/share. Sales were largely Rule 10b5-1-planned and routed through estate-planning trusts (softening the signal), and insiders retain large stakes — but the magnitude and complete one-sidedness as the stock ran from ~$100 to ~$187 is a strong negative. (INTERPRETATION)

Verdict. A coherent but pro-cyclical, serially-leveraged, serially-dilutive capital program whose shareholder outcome depends as much on the terms of the next ~$20B as on operational execution. Management is clearly capable operationally, but on the the author standard — capital allocation as the bridge between business value and shareholder value — the bias is cautious-to-negative: negative returns on capital, a failed vertical-integration bet, encumbered assets, heavy SBC, and ~$3.4B of insider selling with zero buys.

8. Changes and Headwinds — Last Two Years

CoreWeave has compressed an entire corporate lifetime into ~24 months — IPO, hyper-scaling, and a near-monthly capital-raising cadence.

• IPO (2025-03-28). Listed on NASDAQ; first public quarter Q1’25.
• Anchor contracts. OpenAI MSA (~$11.9B) + Sept-2025 order form (~$6.5B); NVIDIA residual-capacity MSA (~$6.3B initial); Meta $21B (Apr 2026); Anthropic and Jane Street ($6B) added Q1’26. (FACT — 8-Ks / calls)
• NVIDIA $2.0B equity placement (Jan 2026, 23M shares at $87.20) — supplier becomes a significant shareholder; carries a circular-financing taint. (FACT)
• Core Scientific (CORZ) acquisition signed (Jul 2025) and terminated (Oct 2025) — the vertical-integration-into-power bet failed when CORZ holders rejected the all-stock deal. (FACT)
• Weights & Biases acquired (May 2025) — MLOps capability.
• Continuous debt issuance — DDTL 2.1/3.0/4.0, senior notes (incl. $2.8B 9.75% in Apr 2026), converts; first IG-rated (A-) HPC-backed term loan at <6% (a genuine cost-of-capital improvement). (FACT)
• Governance / controls / litigation. An un-remediated material weakness in internal control over fixed assets; a securities class action (Masaitis, filed 2026-01-12) and related derivative suits. (FACT — 10-K / 10-Q) A fixed-asset control weakness in a company mid-$40B-PP&E build is precisely where capitalization, useful-life, and impairment errors would hide.
• Capex guidance raised to $31–35B for 2026 (from $30–35B); the stock fell ~47% from its $187 high into mid-2026 on debt-issuance, Chanos-short, and insider-selling concerns. (FACT)

Verdict. On balance these changes strengthen the growth and optionality case (deeper roster, NVIDIA equity, falling cost of capital, software optionality) while sharply raising the financing and execution bar. The thesis now lives or dies on funding $30B+/yr of capex without ruinous dilution — and the failed CORZ deal removed the cleanest path to controlling its own power cost.

9. Risk Analysis

The two risks that actually move the thesis are the financing/dilution gap (High/High) and the depreciation/obsolescence question (the 6-year life is the single largest cost line and the entire margin story rests on it). The capital-cycle/pricing and demand-cycle risks are the medium-term governors; everything else modulates severity.

10. Valuation Discussion (Embedded Expectations)

No price target. Scenario and embedded-expectations framing only.

Peer comps (trailing multiples are near-meaningless at +100% growth; the honest lenses are forward revenue, forward ARR, and EV/backlog) (reconciled to market data 2026-06-10 and the NBIS report):

On every forward metric CRWV is cheaper than NBIS — and the most striking number is EV/backlog ~0.93×: the market pays less than one times the $99.4B contracted RPO. Two readings: the bull says the market is mispricing contracted, take-or-pay cash flows; the bear says the discount is deserved — it prices in the ~$25–28B debt (NBIS is net-cash), the ~67% Microsoft concentration, the 6-year depreciation tail risk, and the funding/dilution overhang. The valuation gap to NBIS is the market’s explicit penalty for leverage + concentration + depreciation risk, partly offset by CRWV’s far larger scale and 4–5× longer backlog. (INTERPRETATION)

Own-history percentiles (own-history valuation percentiles, n=2 components — caution: ~14-month history): composite ~17.6th percentile (cheaper than ~82% of its own record), P/S ~7.7×, P/B ~10.9×. Consistent with the ~47% drawdown from $187. Own-history only, not cross-sectional, and too short to be load-bearing.

Reverse-DCF / embedded expectations. To justify $92.6B EV at ~8× mature EV/EBITDA, CRWV needs ~$11.6B of durable EBITDA; at ~6×, ~$15.4B. At the long-run adjusted-operating margin target (25–30%) and EBITDA running ~45–55%, that implies ~$21–34B of sustained revenue — roughly the exit-2027 >$30B ARR level, held and grown, at the high end of the margin band, after the depreciation wall is fully in GAAP results. So the embedded expectation is that CoreWeave (1) converts >75% of its already-contracted exit-2027 ARR on schedule, (2) holds a true, depreciation-bearing operating margin in the mid-to-high 20s%, (3) funds the ~$30B+/yr build largely with debt at a falling cost of capital, and (4) earns a return on incremental capital above its ~6–9% cost of debt. A demanding conjunction — but, unlike NBIS at ~17.7× forward revenue, the CRWV price does not require heroics beyond the contracted backlog; it largely requires delivering what is already signed at the promised margins. “Cheap if you believe the conversion and the depreciation schedule; expensive if you don’t.” (INTERPRETATION)

Scenario analysis (illustrative EV ranges; NOT price targets):

• Bear (~$30–50B EV): GPU economic life proves shorter than 6 years and/or pricing rolls over as supply floods; one large counterparty trims or delays; the funding gap forces a dilutive equity raise at a depressed price; mature margin compresses. ARR converts but at lower realized economics. ~50–65% below current EV.
• Base (~$80–110B EV): Backlog converts roughly on schedule to ~$30B exit-27 ARR; op margin inflects to low-double-digits in Q4’26 and grinds toward mid-20s; cost of capital keeps falling; dilution contained. ≈ current EV.
• Bull (~$140–220B+ EV): CRWV becomes a durable AI-cloud utility; ARR blows through $30B toward 8 GW by 2030; software/Omni/self-build lift blended margins >30%; WACD hits genuine IG; residual GPU value proves real. A re-rate toward NBIS-style forward multiples on a far larger base.

Right vs. wrong. The market is right that demand is real now, the backlog is contracted, the customer base is diversifying, and the cost of capital is measurably falling. It is possibly wrong in extrapolating sold-out, rising-price conditions through a full capital cycle; in treating the 6-year depreciation life as settled; and in under-weighting the dilution/leverage arithmetic of self-funding $30B+/yr with negative FCF. The ~0.93× EV/backlog says the market is already skeptical — this is a contested valuation, not a euphoric one.

11. Variant Perception

Consensus. A genuinely split sell-side book — roughly 5 strong-buy / 2 buy / 9 hold / 1 sell / 1 strong-sell, median “Hold,” target ~$139 vs. ~$98 spot — plus ~17% short interest (51.4M shares) and a ~47% drawdown. The division is the signal: extraordinary contracted growth and real demand, but a balance sheet and a short thesis that cap conviction.

Strongest bull case. CoreWeave is the scaled, NVIDIA-blessed leader of a generational build-out, with unmatched visibility: a $99.4B take-or-pay backlog (4× NBIS), >75% of >$30B exit-27 ARR already signed; NVIDIA as investor and supplier; a measurably improving cost of capital (first IG-rated HPC term loan <6%, >$20B raised YTD all oversubscribed, no maturities until 2029); diversification away from Microsoft (~$10B FS vertical, 10 customers >$1B); and pricing power on old silicon (A100/H100 prices rising) that validates inference-driven, multi-generation demand and the 6-year depreciation life.

Strongest bear case (the Chanos short). (1) Circular financing — NVIDIA sells to, invests equity in, and benefits from CoreWeave’s buying, a self-reinforcing loop that can flatter demand. (2) Depreciation understatement — a 6-year life on annually-refreshed silicon; if economic life is shorter, the entire “mid-20s contribution margin” overstates true economics. (3) Customer concentration — ~67% Microsoft. (4) Funding gap / dilution — $31–35B capex, −$4.7B/qtr FCF, viable only if capital markets stay open at falling rates. (5) Commoditization — weak intrinsic moat; the software/Omni defense is unproven through a downturn. (6) Insider selling — ~$3.4B sold, zero buys.

The 3–5 assumptions that matter most: (1) GPU economic life ≥ 6 years (depreciation honest); (2) capacity stays sold-out at firm/rising prices through the cycle; (3) funding stays mostly non-dilutive debt at a falling cost; (4) Microsoft renews and concentration keeps falling; (5) mid-20s contract contribution margins convert to GAAP operating profit at scale.

Falsification tests. Bull breaks if: GPU contract/spot pricing rolls over for two consecutive quarters; equity/ATM becomes a primary funding source (visible dilution); a major counterparty trims/delays/non-renews; GAAP operating margin fails to inflect to low-double-digits by Q4’26; or the depreciation life is lengthened to manufacture margin. Bear breaks if: ARR tracks toward the high end of the $18–19B exit-26 / >$30B exit-27 guide on schedule; cost of debt keeps compressing toward genuine IG with deals oversubscribed; backlog keeps diversifying with Microsoft below ~40% of revenue; and Q4’26 adj. op margin lands in low-double-digits with mature-cohort contribution margins demonstrably holding mid-20s. Continued rising prices on A100/H100 through 2026 would directly falsify the depreciation-understatement claim.

12. Fact vs. Interpretation Table

13. Open Questions

1. Is GPU economic life truly ≥ 6 years? No disclosed cohort-level residual value, utilization decay, or impairment data. This single assumption drives the entire margin story.
2. What are the terms of the next ~$20B of financing? Debt vs. equity mix, rate, dilution. The model assumes continuous cheap access.
3. Microsoft renewal economics. Does the ~67% anchor renew, at what price, and how fast does the recognized-revenue base actually diversify (vs. the backlog)?
4. Re-contract / spot economics at end of take-or-pay term. The capacity-arbitrage model depends on a healthy resale market for depreciated GPUs — untested through a downturn.
5. Material-weakness remediation. When is the fixed-asset ICFR weakness remediated, and does it surface any capitalization/useful-life restatement risk?
6. How independent is “demand” from NVIDIA? The circular-financing loop (supplier = investor) complicates the demand read.

14. What Must Be True

Bull case — what must be true: GPU economic life holds at/above 6 years; capacity stays sold out at firm-to-rising prices through the capital cycle; ~$30B+/yr of capex is funded largely by debt at a falling cost toward genuine investment grade with minimal equity dilution; Microsoft renews and concentration keeps dropping; and the mid-20s contract contribution margins convert into low-double-digit GAAP operating margin by Q4’26 and toward 25–30% thereafter.

Falsification test: the bull case is broken if GPU contract/spot pricing declines for two consecutive quarters, or equity/ATM becomes a primary funding source (visible dilution), or GAAP operating margin fails to inflect to low-double-digits by Q4’26 as guided.

Bear case — what must be true: GPU economic life proves materially shorter than 6 years (depreciation understated), and/or the capital cycle turns — supply (hyperscaler in-sourcing + Google/TPU JV + neocloud build) floods in and ASPs compress before the backlog is recognized, and/or capital markets tighten and force dilutive equity at a depressed price. Any one of these converts the leveraged spread bet into permanent capital impairment.

Falsification test: the bear case is broken if A100/H100 pricing continues rising through 2026, and exit-26/exit-27 ARR tracks toward the high end of guidance on schedule, and cost of debt keeps compressing toward IG with oversubscribed deals — i.e., the depreciation and funding fears are empirically refuted.

15. Source Appendix

See the separate Source Appendix (below) for the full list of primary filings (10-K FY2025, Q1’26 10-Q and prior 10-Qs, 8-Ks, DEF 14A, Form 4 corpus, S-1), management transcripts (5 earnings calls + 7 conference presentations, Q1 2025 through June 2026), and secondary industry sources, with URLs and access dates.

The analysis above is deliberately position-free and contains no buy/sell recommendation and no price target; the only stated position is the clearly-labeled “Claude’s Take” block at the top, which is the author’s own independent opinion and general information, not investment advice.

APPENDIX A — Standard Diligence Questionnaire

CoreWeave, Inc. (NASDAQ: CRWV) — Standard Diligence Questionnaire

Supplemental to the main analysis. Fact / Interpretation / Assumption labels applied where it matters. As of 2026-06-10.

General

What thoughtful questions have other investors asked about this company? The dominant investor debates are: (1) Is the 6-year GPU depreciable life honest? — the single largest cost line and the linchpin of the margin story. (2) Can it fund $31–35B/yr of capex without ruinous dilution? — the bull/bear pivot. (3) Is NVIDIA’s demand signal “real” or circular? — NVIDIA is supplier, investor, and a residual-capacity backstop customer. (4) Does the take-or-pay backlog survive a downturn? — counterparty solvency/renewal risk. (5) What is the residual/spot value of GPUs at end of contract? — the capacity-arbitrage model depends on it. (6) Jim Chanos’s short thesis (circular financing + depreciation + insider selling) frames much of the bear discourse.

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? There are no earnings — FY25 GAAP net loss −$1,167M. On the cycle, the business is at a clear peak operating environment: sold-out capacity and rising GPU prices (even on 6-year-old A100s) are textbook late-capital-cycle conditions (Marathon lens). (Interpretation)

Driven by external environment or internal actions? Both. The demand environment (hyperscaler AI capex ~$635–725B in 2026) is external and dominant; CoreWeave’s deployment speed, software, and NVIDIA allocation are internal execution that lets it capture more of it.

How stable are revenues? Contractually stable (98% of FY25 revenue from committed take-or-pay contracts, ~5-year weighted term), but concentrated (Microsoft ~67%) and dependent on capacity going live on schedule. Stable on paper; fragile on concentration and execution. (Fact/Interpretation)

Outlook for products/services. Demand outruns supply through 2026–2027; the question is durability of pricing as ~$700B of industry capex floods supply and Google’s TPU-based JV enters.

How big will this market be? Very large and growing — AI cloud is a multi-hundred-billion-dollar market by 2031 (one estimate ~$400B). The issue is not market size but which layer captures the profit (NVIDIA and power, not necessarily the rental layer). International exposure is currently small (overwhelmingly US revenue).

Business Quality & Competitive Moat

Is the industry getting more or less competitive? More. Hyperscalers in-source; neoclouds (Nebius, Crusoe, Lambda, Nscale) build; Google/Blackstone launched a TPU-based AI-cloud JV (May 2026). Capital is flooding the layer.

How profitable is the business (ROIC, ROE)? Negative. FY25 net loss −$1,167M; negative NOPAT; accumulated deficit −$2,643M. No positive return on capital today. (Fact)

How profitable is the industry / barriers to entry? The profit pool sits at NVIDIA (~75% GM). Barriers are capital + power + NVIDIA allocation — high enough to deter sub-scale entrants, not high enough to deter better-capitalized hyperscalers. Capital/power barriers, not franchise barriers. (Interpretation)

Can the business be easily understood? Conceptually yes (buy GPUs, build data centers, rent compute), but the financials are opaque — the “gross margin” excludes the GPUs’ own depreciation, OCF is prepayment-driven, and ~$11B of capex is non-cash OEM-financed. Understanding the economics requires looking past the headline metrics.

Undermined by foreign low-cost labor? Not directly — it’s a capital/power business, not a labor-arbitrage one. The relevant “low-cost” threat is cheaper capital and cheaper power (hyperscalers, sovereign clouds, TPU economics).

Do brands matter? Modestly — the SemiAnalysis “Platinum” ClusterMAX rating and the NVIDIA reference-architecture qualification are reputational assets, but customers ultimately buy GPU-hours on price/performance/availability.

Nature of competition / switching costs? Competition is on price, performance, deployment speed, and GPU availability. Switching costs are low at the bare-metal/training layer (most revenue) and only moderate at the software/MLOps layer (W&B, storage). 90% of reserved customers use ≥2 products — real but shallow stickiness. (Fact/Interpretation)

Financial Condition & Balance Sheet

Assets not fully recognized on the balance sheet? Minority — the W&B IP and the software stack are partly internally-developed. More importantly, $38.5B of leases not yet commenced are off the balance sheet today. (Fact)

Off-balance-sheet liabilities? Yes and material — the $38.5B of not-yet-commenced lease payments plus ~$1.1–1.7B of equipment-procurement commitments at leased sites. (Fact)

How conservative is the accounting? Aggressive on the items that matter: 6-year GPU life (vs. peers’ 4), GPU depreciation booked outside “cost of revenue,” interest capitalized into PP&E, and a large not-yet-depreciating CIP balance — all flatter near-term margins. Compounded by an un-remediated material weakness in fixed-asset internal controls. (Interpretation)

How CapEx-hungry? Extremely — the most capital-intensive model imaginable. FY26 capex guided $31–35B (cash), with FY25 economic capex ~$21B (incl. OEM financing). FCF −$7.25B (FY25). (Fact)

Capital Allocation & Management

How much FCF, and how is it used? FCF is deeply negative (−$7.25B FY25). There is no FCF to allocate; the company consumes capital, funded by continuous debt and equity issuance.

Significant acquisitions? Weights & Biases (May 2025, capability buy). The marquee deal — the all-stock Core Scientific (CORZ) acquisition to vertically integrate power — was signed July 2025 and terminated October 2025 when CORZ holders rejected it. (Fact)

Buying back shares? No — issuing them. Share base grew from 218M to 532M WA diluted shares (IPO, preferred conversion, NVIDIA $2.0B placement, convert overhang). SBC was $630M (FY25, ~12% of revenue). (Fact)

Issuing shares to insiders / insider behavior? Heavy SBC; and insiders have sold ~$3.4B of stock since listing (founders $2.58B) with zero open-market buys (mostly 10b5-1, via estate-planning trusts). (Fact)

Compensation / incentive alignment. Dual-class structure concentrates founder control; the founders retain large economic stakes but have monetized heavily. The DEF 14A (2026-04-22) is the source of record for the metric structure. (See proxy.)

Motivations of management. Operationally mission-driven and clearly capable; but the scale and one-sidedness of insider selling against widening GAAP losses is a governance caution. (Interpretation)

Valuation & Market Data

ADR / MLP / K-1? No — a US C-corp common stock (Class A). No K-1. (Fact)

Dividend policy? None, and none expected — all capital is reinvested. (Fact)

How profitable? Not profitable on GAAP or cash. “Adjusted EBITDA” (56% Q1’26) sits above both depreciation and interest and overstates economic profitability. (Fact/Interpretation)

Is net income diverging from cash from operations? Yes, sharply — net loss −$1,167M vs. OCF +$3,058M (FY25), but the OCF is entirely manufactured by +$4,174M of customer prepayments (deferred revenue). The “good” cash-flow number is borrowed forward. (Fact)

Risks & Downside

What would cause the stock to decline? A GPU pricing rollover; a depreciation-life reality check (4 vs. 6 years); a funding-market tightening forcing dilutive equity; a Microsoft renewal trim; an AI-capex pause; or a confirmed accounting/control issue.

Risk of catastrophic loss? Yes, real. The leverage (~$25–28B debt at 9–15%, encumbered assets, non-recourse SPVs) plus negative FCF plus depreciation/obsolescence risk create a genuine path to permanent capital impairment if the capital cycle turns before the backlog converts. (Interpretation)

Chance of total loss? Low but non-trivial — the backlog and asset base provide downside support, but a financing freeze coinciding with a pricing collapse and a depreciation cliff is a tail that exists here in a way it does not for net-cash peers. (Interpretation)

Recent News & Events

Has the business environment changed recently? Yes — capex guidance raised to $31–35B; OpenAI/Meta/Anthropic contracts added; NVIDIA took a $2.0B equity stake (Jan’26); the Google/Blackstone TPU JV emerged (May’26); the stock fell ~47% from its $187 high on debt-issuance and short-thesis concerns (Chanos, June 2026). (Fact)

Significant acquisitions? Core Scientific (terminated); Weights & Biases (closed). (Fact)

Change in accounting policies? GPU depreciable life raised to 6 years (2023); an indicated/ongoing watch item given peer practice. Un-remediated material weakness in fixed-asset controls. (Fact)

Recent changes — new markets, facilities, management? Rapid capacity expansion (1 GW active, >3.5 GW contracted, first self-build site online “later this year”); new financial-services vertical (~$10B backlog); continuous executive build-out as a young public company. (Fact)

APPENDIX B — Source Appendix

CoreWeave, Inc. (NASDAQ: CRWV) — Source Appendix

All sources accessed 2026-06-10 unless noted. CIK 0001769628. Primary filings are public SEC documents (EDGAR); transcripts are public company-event records.

Primary — SEC filings (EDGAR, CIK 0001769628)

Reconciliation: revenue XBRL-confirmed via us-gaap:RevenueFromContractWithCustomerExcludingAssessedTax; debt, PP&E, lease, and cash-flow figures read directly from the 10-K/10-Q notes.

Primary — Management transcripts (public company events)

Secondary — industry, competition, valuation

• SemiAnalysis ClusterMAX 2.0 GPU-cloud ratings (CoreWeave = sole Platinum; Nebius = Gold) (independent GPU-cloud benchmarking).
• Bernstein (Madison Rezaei) on the Google/Blackstone AI-cloud JV and CRWV pricing pressure — reported May 2026.
• Big-Tech 2026 capex aggregation (~$635–725B) and Alphabet ~$185B 2026 capex guide — industry compilation.
• Jim Chanos short thesis on CoreWeave (circular financing, depreciation, insider selling) — June 2026 press.

Market data (reconciled to filings)

• Price / market cap / enterprise value / short interest / comps: public market-data providers (e.g. Yahoo Finance) for CRWV and peers (NBIS, ORCL, NVDA), 2026-06-10. Treated as third-party; reconciled to filings.
• Recent-news flow (insider-selling, debt-issuance, the Google JV, the Nebius comparison): financial press, June 2026 — used as a triage signal and validated against primary sources.

Analytical frameworks applied

• Greenwald & Kahn, Competition Demystified — moat-type taxonomy (supply/cost advantage without captivity); moat-to-financial-outcome test.
• Marathon / Chancellor, Capital Returns — supply-side capital-cycle analysis; asset-growth anomaly; mean-reversion of high returns.