CoreWeave is the largest independent AI neocloud — it rents large-scale NVIDIA GPU clusters to the labs training frontier models. Founded by energy traders who started mining ethereum on a pool table in 2017, it pivoted to AI compute and rode the boom to a roughly $50B IPO. Today it runs on two vast numbers: a $60B-plus contracted backlog from Microsoft, OpenAI and Meta, and the debt it took on to buy the GPUs that serve them.
Research
The verdict
"SELLOFF IS MOSTLY OVERDONE (~60/40), MEDIUM confidence — but NOT a buy signal. On the filings, the Jul 1 ~15% drop has no disclosure event behind it (no 8-K on/after Jul 1; latest EDGAR filing is Jun 18, a debt raise). The Meta relationship is an EXPANSION ($21B order form under a pre-existing MSA, Customer B up to 20%) while Microsoft rotated DOWN 72%→45% — diversification-in-progress, not a whale defection. That argues the price reaction is out of proportion to any filed fact → overdone. BUT the structural fragilities are real and PREDATE the news: gross margin 74.3%→71.7%→65.5%, debt tripled to $25.1B, interest expense doubled to $536M/qtr, fresh Jun'26 notes at 9.625%/8.500%. So the correct read is: the SELLOFF (this specific 15% on this specific catalyst) is overdone; the STOCK is not cheap — it is a fairly-priced, structurally-levered bet whose fragility the market is (over-)pricing on a thin catalyst. Falsifier: if the Jul 1 catalyst turns out to be a genuine Meta capacity CUT / de-commitment (not internal optimization), flip to STRUCTURAL — that would be a real demand-side crack in the RPO, not sentiment."
Source documents — open to read in full
Refresh of
deep-dive-2026-06-29.md(7-day delta). This is a targeted refresh built around one position-relevant question, not a full 14-lens re-run: Connor HOLDS CRWV; it sold off ~15% on the Jul 1 "Meta-Compute" news — is that selloff overdone/mispriced, or structural? The Jun 29 dossier already re-grounded every moving lens in primary sources; the only genuinely new fact since is the Jul 1 selloff itself. So this pass re-judges Lenses 5 (what the numbers say about the Meta relationship), 8 (the catalyst / price reaction), and 12/13 (does the bull/bear hold specifically against the selloff thesis), carries the rest, and delivers a falsifiable verdict. Grounding note: this refresh leans on the freshly-staged 100%-filing-derived notedocs/context/crwv-primary-sources.md(10-K FY25, 10-Q Q1'26, three June 8-Ks, read directly from SEC EDGAR, CIK 0001769628, 2026-07-06). Every figure below was re-verified against the actual 10-Q text on the shelf this pass. Price context: sold off ~15% around Jul 1 2026 on "Meta-Compute" news [flagged — see selloff-trigger gap]. Reference pre-drop level ~$96–105 (Jun 27–29). Post-drop implied ~$82–90.
What changed (material moves, all sourced):
API Error: 529 Overloaded (the platform-wide external-fetch outage flagged in the brief). Per the no-retry-storm guardrail, the search was stopped after 2 attempts. The verdict below is built entirely on the filings and does not depend on identifying the catalyst — but the catalyst's nature is the hinge of the falsifier (internal-optimization by Meta = overdone; real Meta capacity cut = structural). See "Selloff-trigger gap" and "What would flip the verdict."What held: The entire structural thesis from Jun 29 stands unchanged — leveraged GPU-leasing arbitrage, $98.8B contracted RPO, six-year GPU depreciation flattering adj-EBITDA, deeply negative FCF, total dependence on continued (currently widening) credit access, live securities litigation over the Nov'25 data-center-delay disclosures. Nothing in the filings moved between Jun 29 and Jul 6.
Position-seed move: Direction unchanged — WATCHING (Connor is long; this refresh informs whether to hold/add/trim, not a fresh entry). The new judgment is specifically on the selloff: ~60/40 that this particular 15% drop is overdone rather than structural, MEDIUM confidence, with an explicit falsifier. No MarketCall tracks CRWV yet; a position SEED is handed to /thesis for Tuesday's verdict call (below). If a call is opened, the Jul 1 catalyst's nature is the fact that most moves it.
Framing. "Overdone" and "structural" are answering different questions, and the honest verdict splits them:
So: the selloff is mostly overdone; the stock is not cheap. The market took a fairly-priced, high-fragility name down 15% on a thin catalyst. That is an overreaction to the catalyst, not a mispricing that makes the equity a bargain. Net lean: ~60/40 overdone, MEDIUM confidence — capped below high conviction only because the catalyst is unidentified (web down) and CRWV's leverage means any real demand-side crack is disproportionately damaging.
Customer A45%72%Customer B20%*). The March 2026 Meta commitment is an order form under a pre-existing master services agreement — "initially committed to pay us up to approximately $21 billion... through December 20, 2032". The filing frames Meta as a growing significant customer, not a departing one. A "Meta-Compute" headline that spooks the tape is landing against a backlog where Meta is an expansion.Weighing it: the structural fragilities are real but are not new and were substantially priced (the stock had already faded from ~$117 to ~$96 on rising-data-center-cost worries per the Jun 29 dossier). The Jul 1 catalyst added a 15% leg on no filed information, against a Meta fact that is an expansion. Therefore the incremental move is more overreaction than repricing → overdone wins, ~60/40. The 40% weight on "structural" is the live possibility that the Jul 1 catalyst is a genuine demand-side signal the filings can't yet show (see falsifier).
Status: UNIDENTIFIED. The specific news item behind the ~15% Jul 1 move is not in any SEC filing (confirmed — latest filing Jun 18, debt-only). Attempting to identify it via the financial press requires WebSearch, which returned 529 Overloaded on both permitted attempts this pass (the platform-wide external-fetch outage). Per the brief's explicit no-retry-storm guardrail, the search was stopped after 2 attempts and the item is flagged deferred.
Working hypotheses (to test when web recovers) — labeled, not asserted:
On the filings alone, the base rate favors H1/H3 over H2 — because a de-commitment of a $21B MSA order form would itself very likely be an 8-K (material definitive-agreement change), and none was filed. That absence is soft evidence against H2, i.e., soft evidence for overdone. Not dispositive (an 8-K can lag, and non-material trims need no filing), but it is the tell the filings give us.
No new print since Q1'26 (reported May 7 2026); next is Q2 on Aug 13. Re-reading the Q1'26 10-Q specifically for what it says about the Meta relationship the market just punished:
Customer A 45% / 72% and Customer B 20% / * for the three months ended Mar 31 2026 vs 2025. Microsoft rotating down, Customer B (Meta-timing) rotating up to 20%. This is the filing's answer to the selloff: diversification-in-progress.The Jun 29 dossier's core finding was: CRWV reacts to balance-sheet / capex-intensity / accounting-credibility / customer-signal events — NOT to revenue (which keeps beating). The Jul 1 selloff fits that pattern exactly: a customer-signal event (a Meta headline) moved the stock hard, while the contracted revenue base was untouched. Historical precedents on the same axis:
Unchanged from deep-dive-2026-06-29.md — no new filing, transcript, or financial since. In particular: Lens 10 (forensic — six-year GPU depreciation, $25.1B debt itemized, live securities litigation), Lens 11 (forward projection — FCF-negative-through-FY27 forecast re-affirmed at p≈0.80), and Lens 14 (the 15 management questions, of which #4 "Microsoft 72%→45% — diversification or Microsoft deliberately slowing?" and a new implied #0 "what exactly is the Meta capacity signal?" are now the most load-bearing). See the prior dossier for the full text.
Research Trail
Covered in the Knowledge Base
Artificial Intelligence
China's #2 carrier-neutral data-center operator, re-rated into a leveraged AI-capacity call — wholesale/AI revenue compounding ~80%/yr with a fresh ~US$1B CATL strategic anchor, but a GAAP-loss-making, ~5x-levered VIE whose bottom line is hostage to interest, convert-fair-value swings and a punitive tax line; buy the EBITDA growth + CATL optionality at ~10x EV/EBITDA (a discount to GDS), underwrite the balance sheet and the China-ADR/governance tail.
>
The picks-and-shovels of Stargate — a DigitalBridge/Silver Lake-owned hyperscale developer whose credit quality is real (A-rated ABS on hyperscaler leases) but whose 2026-vintage growth is now underwritten by a single, unprofitable AI counterparty (OpenAI). Watch the tenant, not the towers.
The purest private-market proxy for the US colocation land-grab — a 1.5GW+ Tier-5 platform DigitalBridge is marking from $11B (2022) toward ~$50B (2025-26) on AI-factory demand; the alpha is in the re-IPO / secondary window, but a 10x-vote founder, a related-party real-estate history, and a parent (DigitalBridge) itself being swallowed by SoftBank are the three governance cracks that decide whether the markup is real or narrative.