CoreWeave

Q1 2026 HAS PRINTED — the first neocloud primary source in this KB. CoreWeave's Q1 2026 IR release (2026-05-07) gives the entity its first financials: revenue $2.08B, $99.4B revenue backlog, >3.5 GW contracted power (>1 GW active), Q1 capex $7.70B (~3.7× revenue), operating loss $(144)M, net loss $(740)M. See CoreWeave Q1 2026 Results. Note: Q2 2026 is not yet released (CoreWeave reports Q2 in mid-August) — the discovery brief's "Q2 10-Q" does not exist yet; this Q1 actual is the latest primary print.

Type: Company (CRWV) — Neocloud / independent AI cloud

CoreWeave is the largest independent AI cloud ("neocloud") — a GPU-operator that buys Nvidia accelerators at scale and rents them as compute, sitting between the chip vendor and the model builders. In the NVIDIA / DOE supply-chain map it is named in the Clouds/model builders tier alongside Akamai, Google Cloud, Lambda, Microsoft, Oracle, Together AI, and xAI — i.e. it is one of the operators turning the GB300 NVL72 buildout into rentable capacity. It is also the topic's flagged first deep-dive target: per the frontier, the next research step is /deep-dive CoreWeave → /thesis new, and we already hold a position.

The neocloud model is the central tradable debate the entity embodies, and the Q1 2026 print now puts real numbers on one side of it. $2.08B of quarterly revenue against $7.70B of quarterly capex and a $(740)M net loss is the empirical shape of frontier debate #2 — "do neoclouds earn their cost of capital, or is it a depreciation mirage?" The gap between the $(144)M operating loss and the $(740)M net loss (~$596M) is largely financing cost — the tell of a debt-funded GPU build. Whether it earns its cost of capital hinges on the GPU depreciation schedule, which the release does not disclose; and the anchor-customer concentration (Microsoft/OpenAI share) is likewise not disclosed. Those two undisclosed items are exactly what the deep-dive must resolve — the Q1 release is primary but promotional; the 10-Q is the fuller, audited filing (flagged as the next fetch).

Key Facts

Q1 2026 — Actuals (printed 2026-05-07; three months ended Mar-31-2026)

• Total revenue $2,078M (~$2.1B). Evidence: strong (first-party IR release) (CoreWeave Q1 2026)
• Operating loss $(144)M; net loss $(740)M — the ~$596M gap is largely financing cost (debt-funded GPU build). Evidence: strong (first-party) (CoreWeave Q1 2026)
• Q1 capex $7,695M (~$7.70B) — property, equipment, capitalized software; ~3.7× quarterly revenue, the build far outrunning current revenue. Evidence: strong (first-party) (CoreWeave Q1 2026)
• Revenue backlog $99.4B — the neocloud-layer demand number beside Oracle's $523B RPO and Alphabet's ~$462B cloud backlog. Evidence: strong (first-party) (CoreWeave Q1 2026)
• Contracted power >3.5 GW (>1 GW active). Evidence: strong (first-party) (CoreWeave Q1 2026)
• Named customers this quarter: Meta ($21B commitment), plus Anthropic, Cohere, Jane Street, Mistral. Evidence: strong (first-party) (CoreWeave Q1 2026)
• BACKLOG-DOUBLE-COUNTING EDGE: Meta's $21B commitment sits inside CoreWeave's $99.4B backlog while Meta is itself a tracked hyperscaler — the KB now holds both ends of one such edge (overlap not yet quantified). See Contracted Backlog & RPO. Evidence: moderate (first-party; both figures in the release) (CoreWeave Q1 2026)
• NOT disclosed in the release: GPU depreciation schedule, customer concentration (Microsoft/OpenAI share), backlog duration/cancellability, FY2026 capex/revenue guidance. Call-circulated figures (Q2 guide $2.45–2.6B, FY $12–13B rev, $31–35B FY capex) are call-sourced, not filing-grade. Evidence: n/a (gap / call-sourced) (CoreWeave Q1 2026)

Borrowing cost — the three loan facilities (2026)

CoreWeave pays for its chips largely with loans secured on the chips themselves. Three of those loans, read side by side, are the one borrowing-cost curve this KB holds. All three are CoreWeave press releases, so every figure is the borrower's own, dated as shown.

Facility	Closed	Size	Interest margin	Rating	Customer behind it
DDTL 4.0	2026-03-31	$8.5bn (about $7.5bn at first)	SOFR + 2.25%, or about 5.9% fixed	A3 (Moody's) / A (low) (DBRS)	one "leading AI enterprise"
DDTL 5.0	2026-05-18	$3.1bn	SOFR + 4.50%	Ba2 (Moody's) / BB+ (Fitch)	two customers below investment grade
DDTL 5.5	2026-08-10	$2.6bn	SOFR + 5.50%	Ba2 (Moody's) / BB+ (Fitch)	contracts averaging about three years

• The margin went from 2.25 to 5.50 points over the benchmark rate in about four and a half months, and the Moody's rating fell five steps, from A3 to Ba2. Evidence: strong for each figure (first-party); the comparison is ours (DDTL 4.0; DDTL 5.0; DDTL 5.5)
• The first step lines up with a change of customer, not only a change of date. DDTL 4.0 finances a contract with a "leading AI enterprise" and is investment grade; DDTL 5.0 finances two customers the release calls below investment grade. So the curve mixes two things: the credit of whoever rents the chips, and the market over time. No release separates them. Evidence: moderate (inference from two first-party releases) (DDTL 4.0; DDTL 5.0)
• Every facility was "meaningfully oversubscribed", and DDTL 5.0 priced 0.50 points tighter than launch. Lenders did not walk away; they charged more for weaker customers and longer risk. Evidence: strong (first-party) (DDTL 5.0)
• DDTL 4.0 is non-recourse and ring-fenced in a single-purpose subsidiary, CoreWeave Compute Acquisition Co. VIII, LLC: lenders can claim that company's assets, not CoreWeave's. Evidence: strong (first-party) (DDTL 4.0)
• DDTL 5.5 runs about five years against contracts averaging about three, so lenders now carry some of the risk that the chips must be re-leased at a lower price. Evidence: strong (first-party) (DDTL 5.5)
• Fundraising totals: about $28bn of debt and equity commitments in the 12 months to March 2026, and more than $30bn raised in 2026 by August. Company framing, unaudited, and the two windows differ. Evidence: moderate (first-party framing)

Why it matters for the token price. The Q1 print shows a net loss about $596M larger than the operating loss, which is largely interest. These facilities show the margin over the benchmark rate on the newest slice of that debt more than doubling. That interest is part of what a CoreWeave GPU-hour must earn, and so part of the price of every token served on it. No source converts it into dollars per GPU-hour yet.

Positioning

• Named in the NVIDIA supply-chain map under "Clouds/model builders" (with Akamai, Google Cloud, Lambda, Microsoft, Oracle, Together AI, xAI). Evidence: moderate (technical-report — NVIDIA) (NVIDIA — Building America's AI Infrastructure)
• Largest independent AI cloud (neocloud) and the topic's flagged first deep-dive target; we hold a position. Evidence: moderate (KB frontier — internal) (see the open questions)
• Anchor-customer concentration is the key risk — and, per the Q1 print, still undisclosed. Evidence: moderate (KB frontier + confirmed gap on the primary)

Mentioned In

• Capital Structure index and frontier — the three facilities are the rung's borrowing-cost curve

• Contracted Backlog & RPO — CoreWeave's $99.4B backlog is the neocloud layer of the "who pays the rent" question, and the site of the Meta double-counting edge

• Rack Power Density — CoreWeave operates the GB300 NVL72-class capacity this concept describes

• Hyperscaler Capex Cycle — neoclouds are the independent demand layer alongside the hyperscaler buyers

• Hyperscaler Buyers — the hyperscalers are both CoreWeave's competitors and (in cases) its customers

Outstanding Questions

• Do neoclouds earn their cost of capital, or is it a depreciation mirage? Now with real numbers on one side — $2.08B revenue against $7.70B capex and a $(740)M net loss — but the answer still hinges on the undisclosed GPU depreciation schedule. (Frontier debate #2 — the core deep-dive question.)
• How concentrated is CoreWeave's revenue by anchor customer? The Microsoft/OpenAI share is not disclosed in the Q1 release; the contract duration backstopping GPU depreciation is likewise absent. Both must come from the 10-Q / 10-K.
• How much of the $99.4B backlog overlaps demand already counted at the hyperscaler layer? Meta's $21B is one confirmed edge; the total inter-layer overlap is unquantified.
• What is CoreWeave's build cost per MW vs the hyperscalers it competes with? (Cross-check: >3.5 GW contracted × ~$59M/MW all-in implies >$200B eventual capex against a $99.4B revenue backlog — capex leads revenue by years.)
• How much of the widening from SOFR + 2.25% to + 5.50% is customer credit, and how much is the market? A fourth CoreWeave facility for an investment-grade customer, or another borrower's GPU loan, would separate the two.
• Next action: ingest CoreWeave's Q1 2026 10-Q (depreciation life + concentration) now; the Q2 2026 print lands mid-August. Then /deep-dive CoreWeave → /thesis new for a tracked call.

Changelog

• 2026-06-24 — Created from 1 source (nvidia-doe-ai-infrastructure); financials flagged as a gap.
• 2026-07-23 — First neocloud primary folded in (CoreWeave Q1 2026 IR release): revenue $2.08B, $99.4B backlog, >3.5 GW contracted power, Q1 capex $7.70B, net loss $(740)M, Meta $21B commitment named. Confidence medium → high (primary filing). Recorded the Meta↔CoreWeave backlog-double-counting edge. GPU depreciation schedule + customer concentration remain undisclosed (kept open for the 10-Q / deep-dive). Q2 2026 not yet released (mid-Aug).
• 2026-09-24 — Folded in the three loan facilities (DDTL 4.0 newly ingested; 5.0 and 5.5 held since 2026-09-11 but never on this page). Recorded that DDTL 4.0 closed 2026-03-31, not in 2025, so the curve spans about four and a half months, and that its first step coincides with a change of customer credit. Fixed links to sources held on other rungs.