Thesis. $104B backlog up 246%, adjusted operating income turned positive, and management reports 5–10% better margins on recent deals — the business is now growing into its debt rather than away from it, though the debt is still the risk.
Reviewed 19 Sep 2026, after Q2 2026 results on 11 Aug. Next review after Q3 2026 earnings on 11 Nov 2026.
What it does
CoreWeave is the largest specialist GPU cloud. It buys NVIDIA accelerators, installs them in powered datacentres, and rents the compute by the hour to AI labs and enterprises. It is the category benchmark — every other neocloud is measured against it.
The number that decides whether this business works is not EBITDA. Adjusted EBITDA excludes depreciation, and depreciating GPUs are the entire asset base. In Q2 the adjusted EBITDA margin was 56% while the adjusted operating margin — the one that charges for the chips wearing out — was only just positive.
A GPU cloud is a leveraged bet that a chip generates more rent over its life than it cost to buy and finance. Everything else is detail.
Bull case
- Q2 2026 revenue $2.575B, +112% YoY. Q3 guided $3.4–3.6B (~158% growth), above the $3.43B consensus. FY26 raised to $12.4–13.2B
- First positive adjusted operating income: +$128M. Adjusted EBITDA $1.51B from $752M. Adjusted EPS −$1.14, beating −$1.41 by 27 cents
- Backlog $104B, +246% YoY — and that figure excludes more than $25B of new commitments signed in early Q3. The deepest contracted book in the category by a wide margin
- Contracts: Meta $21B through 2032 on top of a prior $14B (~$35B total); a multi-year Anthropic agreement for Claude compute; a $6B Jane Street commitment alongside a $1B strategic investment
- Capacity: 1.5 GW active across 51 datacentres, adding ~500 MW in the quarter. Contracted power 4.2 GW, up from 3.5 GW. Targeting 1.85 GW live by year-end. First bring-up of NVIDIA's Vera Rubin NVL72
- Pricing power is visible. CEO Intrator: recent deals carry 5–10% better margins, with Vera Rubin expanding from the start. Managed inference ARR went from $1M to over $100M, targeting $250M+ by year-end
- Exit-2026 ARR raised to $18.5–19.5B. Added to the Nasdaq-100
Bear case
- The balance sheet is the story. Raised more than $10B of unsecured debt and convertibles in a single quarter, plus a $3.1B term loan. Net loss widened to −$626M specifically on financing costs, from −$290M
- Operating loss −$49M, from +$19M a year earlier, once depreciation is charged
- Capex $9.4B in the quarter; FY26 guided $35–39B, raised from $31–35B. Free cash flow deeply negative
- Only ~36% of backlog converts to revenue within two years. A $104B number against $35–39B of annual capex is not the cushion it looks like
- Assumed GPU useful life is the most contested undisclosed number in the category. Shortening it would hit reported profit across every name here at once
- Competition rising: SpaceX has begun selling excess compute and Meta is reportedly considering launching a cloud
- The most rate-sensitive name on the watchlist — it funds growth with debt against negative free cash flow
Major customers
- Meta — ~$35B of total commitments through 2032, the single largest disclosed contract
- Microsoft — historically the anchor customer and a very large share of early revenue
- Anthropic — multi-year agreement for Claude compute
- Jane Street — $6B commitment plus a $1B strategic investment
- NVIDIA — supplier, investor and, through referrals, effectively a channel partner
Read-through. CoreWeave sits directly downstream of NVDA and competes with NBIS, IREN and WYFI for the same customers. Its capacity announcements convert into orders for VRT, NVT and the Grid Equipment names. NVDA's own Q2 put its ACIE segment (AI clouds, industrial, enterprise) at $40B, +138% YoY — growing faster than hyperscale, and CoreWeave is the largest single piece of it. Note the circularity: NVIDIA is supplier, investor and demand driver at once, a pattern that repeats across this category and concentrates risk more than the ticker count suggests. CoreWeave is not comparable to CIFR or KEEL, which are landlords and buy no GPUs.
What would change the view
- Free cash flow and interest expense — the survival metrics
- Backlog conversion pace against the ~36%-in-two-years figure
- Debt maturity schedule and any refinancing, particularly if rates move
- Assumed GPU useful life, and any change to it
- Whether annual guidance is ever provided
- Capex guidance revisions against the $35–39B plan
- Whether the 5–10% margin improvement on recent deals holds as Vera Rubin volumes scale
Update log
19 Sep 2026 — Converted to the standard template. No new disclosures since the Q2 print; content consolidated and the superseded pre-Q2 figures removed.
7 Sep 2026 — Verdict: Watch → Constructive. The thesis had been "best contracts, worst balance sheet, bimodal." Q2 moved the operating line positive, grew backlog 246% and demonstrated pricing power. The debt is still the risk, but the business is visibly growing into it rather than away from it.
11 Aug 2026 — Q2 2026 reported. Revenue $2.575B (+112%), adjusted operating income +$128M (first positive), adjusted EBITDA $1.51B, adjusted EPS −$1.14 (beat by $0.27), net loss −$626M, operating loss −$49M. Backlog $104B (+246%). FY26 revenue raised to $12.4–13.2B, adjusted operating income to $960M–1.15B, capex to $35–39B. Stock +13% after hours to ~$102.
Research and education only — not investment advice.
Made with Bullet