CRWV — knowledge base

Overview

CoreWeave (CRWV) is a pure-play neocloud AI infrastructure provider whose Q2 2026 earnings, paired with Nebius results, were framed as a potential narrative inflection after 6–8 weeks of sector selling. Discussion centers on extended GPU useful lives validating longer depreciation schedules, a large multi-winner TAM across neocloud and colocation, and competitive edges in financing, prepayments, and GPU access amid capacity scarcity. Peers (Nebius, IREN/Iron, White Fiber, Riot, Cipher) supply comps on deal pricing, execution models, and bottlenecks (power, labor, ERCOT interconnection).

Key facts & figures

CoreWeave contracted Nvidia A100 GPUs through 2029 at attractive rates, implying ~9-year useful life from the 2020 launch and directly undercutting 2–3 year technical-obsolescence bear cases. [[s:122@00:05:47]]
Nebius (closest neocloud comp) reported payback periods compressed to ~1 year 10 months on recent deals (from prior 2–3 years) with pricing >$20M/MW; short-term/burst/auction deals reached $40–50M per MW as a portfolio slice. [[s:122@00:06:45]]
Colocation comps: Riot–Anthropic ~191 MW, 20-year, ~$9.1B lease at Rockdale equates to ~$2.4M average annual revenue per MW (add-on to prior AMD capacity, total ~241 MW). Fermi–TensorWave 15-year ~$6.5B/222 MW lease at Project Matador equates to ~$1.95M average annual per MW (speaker claim of ~$2.7M overstated/misleading). [[s:122@00:12:57]]
IREN (Iron) guided for 730 MW to come online/under construction to support 1.21 GW scale in 2027 (BC + Texas sites). [[s:122@00:10:04]]
White Fiber (WYFI) announced ~100 MW high-density GPU capacity deal (Krambu, exclusive ops, 2027) with stock response >20%. [[s:122@00:04:36]]
Texas ERCOT paused Batch Zero large-load reviews (early Aug 2026, per Gov. Abbott audit request), heavily impacting Cipher’s pipeline into 2027; NY imposed statewide hyperscale data-center moratorium (Jul 2026, up to 1 year) affecting names such as Wolf. [[s:122@00:16:02]]

Thesis & bull case

Extended A100 contracts through 2029 validate multi-year GPU cash-flow tails and refute short-obsolescence theses, supporting higher asset values and lower effective depreciation.
Large multi-winner TAM: hyperscaler anchors, frontier labs, and opportunistic burst demand can coexist; short hardware paybacks (<2 years at peers) leave multi-year residual cash flows.
Decisive edges are financing access, customer prepayments, and scarce GPU supply rather than pure power or real-estate ownership.
Rising colo lease comps (>$2M average annual/MW) and neocloud burst pricing ($40–50M/MW) signal sustained pricing power from AI capacity scarcity.
Constructive on specialized operators (tech-led vs pure power/real-estate) that can recycle capital via cloud contracts and financing flywheels.

Risks & bear case

Execution differentiation: Iron praised for power/vertical integration yet criticized as reactive and communication-light versus Nebius market agility and White Fiber’s Tier-3 retrofit speed + multi-revenue model; management credibility and per-share returns remain flashpoints.
Bottlenecks: skilled-trades labor (debated—oil-field analogy claims money clears mobile crews; counter-view flags remote Texas-style sites harder than coastal retrofits) and power/interconnection (ERCOT Batch Zero pause, NY moratorium).
Financing and GPU allocation remain zero-sum competitive variables; weaker balance sheets or slower prepay conversion lose share.
Residual obsolescence and utilization risk if next-gen GPUs or software efficiency compress older-fleet demand faster than contracted tails.
Tribal short-term stock focus and leverage amplify drawdowns even when underlying MW economics improve.

Timeline of developments

2026-07: NY statewide moratorium on new hyperscale/large data centers (thresholds ~50 MW+ or related 20 MW+) imposed for up to one year. [[s:122@00:14:55]]
Early 2026-08: Texas Gov. Abbott directs audit; ERCOT pauses Batch Zero large-load interconnection/transmission study, stalling Cipher-type pipelines into 2027. [[s:122@00:16:02]]
2026-08-12: CoreWeave Q2 earnings (A100 contracts through 2029) followed by Nebius Q2 (sub-2-year paybacks, $40–50M/MW burst) cited as sentiment turning point after 6–8 weeks of AI-infra selling; colo comps (Riot ~$2.4M, Fermi claimed ~$2.7M) and peer execution debate (Iron vs Nebius/White Fiber) dominate. [[s:122]]

Open questions

Will labor or power/interconnection prove the binding constraint on 2026–27 MW delivery, and does capital fully clear skilled-trades shortages at remote sites?
Which operating model wins on risk-adjusted returns: asset-heavy power-first (Iron), agile retrofit/cloud-contract (White Fiber), or hyperscaler-anchored multi-bucket (Nebius)?
How durable are >$2M/MW colo averages and $40–50M/MW burst rates once new supply clears ERCOT/NY bottlenecks?
Can CoreWeave convert GPU-access and prepay advantages into sustained share versus better-capitalized or more agile peers?

Notable predictions to track

Multi-year residual cash-flow tails materialize from sub-2-year hardware paybacks and A100 contracts running to 2029.
Narrative shift after CoreWeave/Nebius prints broadens beyond one-day relief into sustained multiple recovery for neocloud/colo complex.
730 MW IREN 2027 ramp and White Fiber 100 MW 2027 GPU deal deliver on time/cost, validating (or refuting) respective execution theses.
Money continues to attract mobile skilled crews at scale, preventing structural labor inflation from derailing Texas and other remote builds.