CRWVCoreWeave, Inc.

CRWV — dossier, 2026-09-23

Slot names-ai-capex-4, on a Tier 0 promotion. Drafted by Claude in an unattended cloud routine. Nothing here is human-verified. Every figure carries the document it was read from; anything from background knowledge is marked [background]. Quotes were located by substring search in the source and copied, never re-typed. Never edited after today.

Price reference: $86.76 close, 2026-09-22, volume 26,428,459 (ROIC.ai NASDAQ:CRWV). The 2026-09-22 8-K gives an earlier fix: $79.88 on 2026-09-17.

---

The one-line version

CoreWeave burned $10.45 billion of free cash flow in six months, carries $35.55 billion of debt against $5.02 billion of book equity, has $10.6 billion of principal falling due inside eighteen months at effective rates running to 15% — and in the week before this pass it raised ~$7.2bn across a 2.875% convertible and a 35-million-share ATM. The coupon is the news: the existing stack pays 8.5% to 9.75%. Either the credit has been repriced, or the equity optionality is paying the coupon. Everything else — $103.7bn of RPO, 4.2GW contracted, inference from $1m to $100m in a quarter — is the reason someone was willing to write it.

---

S1 — Source

watchlist (owner, Apple Stocks, 2026-09-20), row: {ticker: "CRWV", label: "CoreWeave, Inc.", coverage: sec_domestic, cluster: ai_datacenter, cik: 1769628, note: "recent IPO; S-1 is the primer"}.

Tier 0 promotion. queue.json promoted CRWV at priority 2 on 2026-09-23 — "8-K filed 2026-09-22 (current report)". Under SKILL §0 a promoted name with no scorecard.md replaces the last member of the batch; CRWV was already the last member of names-ai-capex-4, so the batch is unchanged and the promotion was handled in place. It is removed from promoted by this slot.

S2 — Kill test

From XBRL company facts (data.sec.gov, User-Agent sent) and the Q2 2026 10-Q.

Q2 2026 alone: revenue $2,575m, operating loss $(49)m, net loss $(626)m.

Balance sheet at 2026-06-30: cash $5,524m; total assets $77,070m; total stockholders' equity $5,024m; deferred revenue (current + non-current) $9.7bn; shares outstanding 457m Class A + 94m Class B = 551m.

The debt, in management's own table (Note 10, effective interest rates as filed):

Total future principal $35,551m, of which $4,413m falls in the remainder of 2026 and $6,184m in 2027. Net interest expense rose 140% year over year in Q2 2026 and 121% in H1, "primarily attributable to increased borrowing levels and total debt obligations."

Survives-or-dies. CoreWeave survives on access to capital and nothing else. Operating cash flow of $3.66bn against $14.1bn of capex is not a business that funds itself, and $5.02bn of book equity under $35.55bn of debt is roughly 7x. The counterweight is real and also documented: $103.7bn of RPO, two named counterparties committing ~$6.5bn (OpenAI, to 2031) and ~$21.0bn (Meta, to 2032), 4.2GW contracted, and — decisively for this quarter — a capital market that has just written $4.2bn at 2.875% and stands ready to take 35m shares. The kill test is passed conditionally: it survives as long as that window stays open, and the RPO is what keeps it open.

S3 — Filings, gate-minimum

Read: the Q2 2026 10-Q in full (Notes 2, 3, 10; Part I Item 4; Part II Item 1A; MD&A); the 2026-09-22 8-K (the promoting filing); the 2026-09-17 424B5; the Q2 2026 call (turns 9–43 of 43). Not read: the FY2025 10-K (2026-03-02), the DEF 14A (2026-04-22), the S-1/424B4, any earlier transcript, and the efts.sec.gov auditor/restatement sweep. The promotion put the September filings and the latest 10-Q in front of the annual report; that is a deliberate ordering and it is also this pass's largest gap.

The three biggest risks management itself discloses

1. "A substantial portion of our revenue is driven by a limited number of customers." — Q2 2026 10-Q, Part II Item 1A. The numbers in the same paragraph: top three at "approximately 36%, 26%, and 10%" of Q2 2026 revenue; top customer alone at ~71% of Q2 2025; "OpenAI committed to pay us up to approximately $6.5 billion through May 31, 2031"; "Meta Platforms, Inc. ("Meta") initially committed to pay us up to approximately $21.0 billion". And the forward statement: "We expect that our customer concentration with a limited number of top customers is likely to continue in future years because of the long-term nature of contracts with those customers."
2. "Our substantial indebtedness could materially adversely affect our financial condition, our ability to raise additional capital to fund our operations, our ability to operate our business, our ability to react to changes in the economy or our industry, our ability to meet our obligations under our outstanding indebtedness and could divert our cash flow from operations for debt payments, and we may still incur substantially more indebtedness in the future." — Q2 2026 10-Q, Part II Item 1A, heading, immediately followed by "We have a substantial amount of debt, which requires significant interest and principal payments." The body names DDTL lenders "instituting foreclosure proceedings against their collateral" as a specific consequence of default.
3. Reliance on third-party data centres CoreWeave does not control, with a precedent already on the record: "the delays that we announced in November 2025 with respect to the delivery of certain data centers to be provided by a third-party data center provider". CoreWeave leases or licenses all of its space.

Controls: not effective

Part I Item 4: "our disclosure controls and procedures were not effective at the reasonable assurance level as of June 30, 2026 due to the material weaknesses in our internal control over financial reporting described below." The three weaknesses are IT general controls over the applications supporting financial reporting, insufficient segregation of duties, and "lack of sufficient number of qualified personnel within our accounting, finance, and operations functions". "We have also concluded that these material weaknesses continued to exist as of June 30, 2026." CoreWeave adds that they "did not result in a material misstatement to our financial statements."

This is the same disclosure names-ai-capex-3 recorded at CBRS. Two of the cluster's recent IPOs are running unremediated material weaknesses while raising billions in the capital markets.

The RPO, and the shape it shares with Oracle

"As of June 30, 2026, the Company had $ 103.7 billion of unsatisfied RPO, of which 41 % was expected to be recognized over the initial 24 months ending June 30, 2028, 39 % between months 25 and 48, and the remaining balance recognized between months 49 and 78." (Note 2.)

Against annualised Q2 revenue of ~$10.3bn that is ~10 years of sales booked and undelivered — the same ratio ORCL shows at $664bn against $67.4bn. Two companies at opposite ends of the size range, funding very different balance sheets, both now valued on a backlog roughly ten times revenue whose counterparties they will not name. CoreWeave at least names two of them.

One nuance the definition carries and Oracle's does not: CoreWeave's RPO is explicitly net of estimated variable consideration, including "amounts that may not be recognized as revenue due to delivery delays" and "estimates of committed cloud computing capacity that the Company has the right to resell". The $103.7bn is already after management's own haircut for the delivery risk the November 2025 disclosure demonstrated.

Off-balance-sheet development capital

Note 3: CoreWeave committed up to $1.7bn in H1 2026 (about $500m of it in Q2) to acquire equity in two joint ventures each holding a data centre development project. It is not the primary beneficiary, so the VIEs are unconsolidated; carrying value $479m at 2026-06-30, maximum exposure $1.7bn. That is development capital which does not show in the $14.1bn capex line.

Transcript — Q2 2026 call, 2026-08-11 (read Q&A-first, SKILL §3.1)

Turns 9–43 of 43. Prepared remarks not read. Six analysts: J.P. Morgan, Deutsche Bank, Evercore ISI, Barclays, Wells Fargo, Cantor Fitzgerald.

Guidance and changes.

• Exit ARR guidance raised to $18.5–19.5bn. The exchange is worth quoting carefully: the analyst asked about "the $18 to $19 billion in ARR as kind of the exit target for 2027" and the CFO answered "we increased the exit ARR number that we provided in guidance to you folks right now at 18.5 to 19.5 for 2026." The two years do not match. This transcript is machine-produced and the year in either sentence could be a transcription artefact; the number is recorded here as the CFO said it, and the year is flagged as unresolved rather than silently corrected. The Q3 press release settles it.
• Power: 4.2GW contracted, plus ~1.5GW of powered-land options and executed LOIs — "close to about six gigawatts already" — against a stated goal of >8GW of active power by end-2030.
• 500MW of active power added in Q2, 300MW of it in June alone, "higher than any amount of power that we've added in any history prior quarter".
• Contribution margins on recent deals 5–10 percentage points higher than recent quarters.
• Managed inference went "from $1 million to $100 million inside of a single quarter."
• DDTL 5.5 closed (announced 2026-08-10) — a facility that finances shorter-duration customer contracts, which the CEO says opens the enterprise market: "These clients want to buy compute for two years or three years, and that's not a market that was easily accessible to us until we were able to bring DDTL 5.5 to market."
• Vera Rubin is seeing margin expansion "right from the start", and is where much of the 5–10 point step function sits.
• On data-centre moratoriums: "we feel like moratoriums are not going to impact the demand for this infrastructure. They are going to impact where this infrastructure gets built" and "none of those numbers will be impacted by the regulatory pushback as of today."

What the analysts asked about. Supply chain security; managed inference and capacity allocation; the drivers of the margin uplift; regulatory pushback against data centres; fleet evolution for inference; the gap between power added and revenue added; and recontracting GPUs coming off contract. Four of the seven were, underneath, the same question: is the margin improvement structural or is it term mix.

Dodges.

• The margin uplift. Zelnick: "can you help unpack the drivers? How much is a function of shorter duration deals versus strong competitive differentiation or other factors?" Intrator: "So it's a combination of a lot of things, And it's difficult to deconstruct it." He then gave three qualitative reasons — platform quality, customers monetising their own products, premium pricing of a premium product — and no attribution. This matters because the CFO had just credited DDTL 5.5 with opening short-duration contracts, and short-duration pricing is exactly the "other factor" the question was trying to isolate.
• Power-to-revenue linearity. Turrin observed that 500MW went in while sequential revenue adds were "fairly consistent with last quarter". The CFO answered that 300MW landed in June so the effect arrives in Q3/Q4; the CEO then moved to Vera Rubin. The gap itself was never quantified.
• Long-term supply agreements. Chatterjee asked whether CoreWeave needs broader long-term supply agreements. The CEO answered on relationship quality — "it's part of what we do every single day is nurture these relationships" — and named no agreement, term or volume. The analyst's question referenced a counterparty the transcript renders as "Solidine"; that name is not resolvable against anything read in this pass and is not guessed here or in relationships.yaml.

S5 — Ownership

13F register at 2026-06-30 (3spread, CUSIP 21873S108; 676 filers against 792 the prior quarter — ingestion incomplete, no exit read from an absence; 406.4m of 551m shares, the balance being founder/insider Class B):

The largest shareholder is also a lender. Magnetar Financial holds 52.1m shares, and the 10-Q's debt table carries a "Magnetar Loan" maturing January 2029 at a 12% effective rate ($189m at 2026-06-30, down from $273m). Equity and 12% paper in the same hands, with the equity being trimmed by 16.5m shares while the loan amortises. Whatever the intent, the position is paid on either resolution of the capital structure, and that is a different kind of holder from an index fund.

NVIDIA holds 47.2m shares and did not move them. Read this beside the INTC pass in the same slot, where NVIDIA is the fourth largest 13F holder with 214.8m shares, also +0 QoQ. The routine's standing cross-cutting finding has been about contractual support — guarantees, warrants, take-or-pay, vendor loans. This is the equity version, and it is passive: NVIDIA sits on the register of the complex's most leveraged buildout and of its most capital-hungry manufacturer, and traded neither in the quarter.

Goldman trimmed 14.1m shares in the quarter and is, since September, a sales agent and forward purchaser under the equity distribution agreement and a capped-call counterparty on the converts. No inference is drawn from the sequence; both facts are in the record.

• Short interest (FINRA, via Equibles): 51,444,662 (2026-05-15) → 80,963,200 (2026-06-30, the peak) → 64,362,421 (2026-07-15) → 55,568,683 (2026-08-31). A 31% decline off the peak, and the covering ran through the 2026-08-11 results rather than starting after them. Days to cover 2.6 → 2.7 on thinner volume. Equibles also returns a model estimate for 2026-09-15 (~56.5m, +1.7%, P10–P90 49.7–64.7m); it is an estimate, not a FINRA figure, and is not used here.
• Shelf / ATM / buyback: the S-3ASR of 2026-06-05 (333-296553) is live and was drawn on by the 2026-09-17 424B5 for up to 35,000,000 Class A shares with collared forward sales. No buyback. Dilution instruments in issue: two existing convertible series, the new 2033 series (max 52,578,540 shares), the ATM (35,000,000), plus the equity plan. CoreWeave is an issuer of shares from every direction, with $566.2m spent on capped calls to slow one of them down.

S8 — Valuation and sizing ⏳ OPEN (user)

Today. $86.76 (2026-09-22). 551m shares → ~$47.8bn of market value. Net debt at 2026-06-30 was $35,551m − $5,524m = $30.0bn; the September convert adds $4.2bn of principal against $4,137m of proceeds less $566m of capped-call cost, so net debt is roughly unchanged at ~$30.4bn. EV ≈ $78bn, on annualised Q2 revenue of ~$10.3bn — about 7.6x EV/revenue. There is no earnings multiple.

Out-year model. This one is cruder than the other two names in this slot, and the reason is structural: the equity is a $47.8bn sliver on top of $30bn of net debt, so small changes in the enterprise value swing the equity a long way. Assumptions: revenue $19bn in 2027 (the exit-ARR guide read as a run rate) and $30bn in 2028; EBITDA margin 45%, which is roughly what a contract-compute business with the disclosed contribution margins should carry before D&A; net debt at $50bn by 2028, because the build continues; and ~600m shares after the ATM but before any convert conversion.

What that says. Base lands ~12% above the traded price — the only name in this slot where it does — and the bear case takes the equity to roughly nothing, because at 5x EBITDA the enterprise is worth less than the debt. That asymmetry is the investment case and the objection at once: this is a levered equity on a backlog, and the distribution of outcomes is wide in both directions rather than centred near the price. The September financing package is evidence about the left tail specifically: a market that writes 2.875% paper is not pricing the bear case, and a company that pays $566.2m for capped calls is not either.

Buy price and size are owner-only (skill §5). Liquidity: 26.4m shares on 2026-09-22, ~$2.3bn — ample, though the float is 457m Class A rather than the full 551m.

S9 — Monitoring

• Next earnings: Q3 2026, on the 10-Q cadence (Q1 filed 2026-05-08, Q2 filed 2026-08-12) — early November 2026. A precise date was not obtained; ROIC's calendar needs a paid plan.
• Feed: Tier 0 EDGAR sweep covers cik 1769628, and it is what promoted this name.
• Leading indicator between quarters: the RPO line in Note 2 together with its 24-month percentage — $103.7bn and 41% at 2026-06-30. Unlike revenue, it moves before delivery does, and unlike ARR guidance it is audited-adjacent and net of management's own delivery-risk haircut. Second: the debt table's next-twelve-months principal column, which the September convert should visibly reduce.
• Trigger that converts this to a buy: the 5–10 point contribution-margin uplift showing up in reported gross margin for two consecutive quarters while RPO grows and the 24-month share holds at or above 41% — i.e. the margin being structural rather than term mix, confirmed in the financial statements rather than in a qualitative answer management declined to decompose.
• Trigger in the other direction: any disclosed delay on a contracted gigawatt, a top-three customer renegotiation, or the material weaknesses still open in the FY2026 10-K.

---

Access and evidence limits, this run

• The FY2025 10-K and the DEF 14A were not read. The promotion put the September filings and the Q2 10-Q first; the annual risk factors, the auditor's report, auditor tenure and the Class A/Class B voting arithmetic are all unread. Largest gap; it is the scorecard's next action.
• One transcript, not two, and prepared remarks not read — so the 5–10 point margin figure is reported as management restated it in Q&A, not as it was first framed.
• The exit-ARR year is unresolved (analyst said 2027, CFO answered 2026). Recorded as said, flagged, not silently reconciled.
• "Solidine" not resolved. The analyst named a supply counterparty the transcript renders that way; no filing read here corroborates it and no name is guessed.
• No efts.sec.gov full-text sweep for auditor change or restatement language.
• No GetCustomerConcentration / GetGoingConcernStatus / GetAtmPrograms calls — Equibles budget. Customer concentration came from the 10-Q's own risk factor, which is the better source; the ATM came from the 424B5.
• Form 4 / Form 144 not pulled for CRWV — 3spread time, and the lock-up analysis in CBRS/dossier-2026-09-21.md is the template for when it is done.
• The out-year model is the routine's own, on stated assumptions, and is an EV/EBITDA frame because no earnings exist. On a balance sheet this levered, treat the three cases as a spread of outcomes rather than a valuation.