CoreWeave

One-line summary: Leading US neocloud GPU infrastructure provider (NYSE: CRWV); 9 of top 10 AI labs as customers; 1.5 GW active power at Jun 30, 2026 (up ~500 MW in Q2); ~3.7 GW contracted; $21B+ financing YTD 2026; contract duration extending to 5-year take-or-pay; primary beneficiary of the powered-shell bottleneck thesis; inference >50% of platform utilization.

What it is

CoreWeave (NYSE: CRWV) is a hyperscale AI cloud infrastructure provider ("neocloud") that specializes in GPU compute for AI training and inference. Co-founded by Brannon McBee and others; original background in crypto mining (ETH). IPO in 2025. CoreWeave does not sell GPU time speculatively — it only commits capex when a client signs a specific build contract ("success-based capex approach"). All infrastructure built to Nvidia DGX reference spec.

Why it matters to stock-market

CoreWeave is a primary-source window into the AI infrastructure demand cycle. McBee's testimony on: (1) contract duration extending 3→5 years; (2) inference now >50% of utilization; (3) "powered shell" as the binding bottleneck (not GPUs); and (4) enterprise customer diversification are all load-bearing claims for the AI capex sustained-demand thesis. CoreWeave is also a listed neocloud directly exposed to the picks-and-shovels AI infrastructure theme.

Key facts

• Customer mix: 9 of top 10 global AI labs; 10+ clients with >$1B contracts each; financial services backlog ~$10B (direct enterprise, not via AI labs); Q4 2025 added 2× new logos vs. any prior quarter. Customer concentration risk (IPO criticism) is actively being addressed. From 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• Scale: 1+ gigawatt of active powered infrastructure deployed to clients (June 2026 Odd Lots). Q2 2026 IR (as-of Jun 30, 2026): +~500 MW in Q2 to 1.5 GW active; ~3.7 GW contracted. Aug 11, 2026 call (Fool transcript of issuer remarks): >300 MW in June alone; first self-build "expected to come online later this year"; Indonesia 360 MW "coming online in approximately 18 months"; YE 2026 active-power guide >1.85 GW; CEO ≥8 GW by 2030. Call-date contracted 4.2 GW (Aug 11). Fleet MW, not a single-campus clock. From 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right and 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
• Financing (YTD 2026 color, pre-Q2-IR): $21B in financing raised YTD 2026, including DDTL4 — investment-grade rated, non-recourse HPC infrastructure financing at SOFR+225 basis points (first of its class). From 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• DDTL 4.0 ($8.5B, closed 2026-03-31): borrowed by CoreWeave Compute Acquisition Co. VIII, LLC; rated A3 / A (low) — first IG HPC-infrastructure delayed-draw. Floating SOFR+2.25% / fixed ~5.9%. Anchored by Blackstone Credit & Insurance (not a named public BDC). Arrangers: MUFG, Morgan Stanley, Goldman Sachs, JPMorgan. From 2026-08-17-autoresearch-ai-capex-private-credit-contagion.
• Rating travels with the offtake, not the silicon. CCIR's filing read: DDTL 4.0 (Meta take-or-pay) is A3; DDTL 5.0 (two non-IG customers) is Ba2 at SOFR+4.50% with a full parent guarantee. Same GPU collateral class; the credit standing of the committed payments moved the rating. From 2026-08-17-autoresearch-ai-capex-private-credit-contagion.
• DDTL 5.5 ($2.6B, closed 2026-08-10): SOFR+5.50%, Ba2 / BB+, ~five-year maturity against ~three-year average customer contracts (lenders underwriting renewal risk). Official: >$30B debt-and-equity capital YTD. TechTimes (Bloomberg syndication): initial talk SOFR+425–450 at 99; deal blew out 100–125 bp to SOFR+550 at 97, YTM 10.44%, plus 1.35x DSCR and $112.5M minimum liquidity. Pricing stress, not a missed payment. From 2026-08-17-autoresearch-ai-capex-private-credit-contagion.
• Contract duration: Extending from 3-year (prior) → 4-year → now 5-year take-or-pay commitments. AI labs say: "we want explicit access for 5 years, can't cancel, exact same economics." "The deployments are getting larger and larger." From brannon-mcbee in 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• Inference >50% of utilization: CEO Mike confirmed inference workloads represent "well in excess of 50% of infrastructure utilization" on the platform. The same Nvidia infrastructure serves both training and inference — fully fungible across use cases. From brannon-mcbee in 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• Bottleneck shift: 3 years ago, the bottleneck was GPU access. Today: "Having a powered shell is the bottleneck." Powered shell = an empty, energized data center (power + cooling + all components) ready to accept racks. Supply chain constraints (transformers, electricians — 5-year apprenticeship to qualify) mean this bottleneck cannot be scaled quickly. From brannon-mcbee in 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• Nvidia exclusivity: Clients ask only for Nvidia infrastructure. "We really don't see demand on a material basis for anything but that Nvidia compute." Nvidia's CUDA ecosystem and 15-year investment are "most efficient, most scalable, most reliable." From brannon-mcbee in 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• GPU non-fungibility: Same GPU (H100) in CoreWeave vs. another cloud has different performance — measured by goodput and model flop utilization (MFU). CoreWeave differentiates via software stack that predicts GPU failures and maintains uptime. This is why GPU compute cannot be commoditized like natural gas. From brannon-mcbee in 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• Vera Rubin: CoreWeave has begun receiving testing racks for Vera Rubin. Expects scaling laws to hold through Vera Rubin. "It absolutely seems like tooling is important, but scaling laws are still holding. Your ability to advance your frontier model through accessing more infrastructure at scale holds." From brannon-mcbee in 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• Compute market structure: Not currently commoditizable — non-fungibility of GPU deployment quality prevents exchange-traded compute market in near term. "Until it starts becoming easier [to operate], you don't really have a path to commoditization." From brannon-mcbee in 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right.
• ClusterMAX 3.0 Platinum (2026-09-23, partial): SemiAnalysis tested 77 managed GPU-cloud providers. CoreWeave remains Platinum and "still sets the technical bar for others to follow." Nebius joins the same Platinum tier at seller's prices. Only 19 neoclouds globally achieve a Medallion rating. Attach, not a new chain and not an NBIS signal. From 2026-09-24-feed-semianalysis-clustermax-30-the-industry-standard.

Strengths (from a thesis-input perspective)

• Investment-grade financing at SOFR+225 confirms institutional confidence in revenue model
• 5-year take-or-pay contracts with largest AI labs provide revenue visibility
• Customer diversification accelerating (financial services direct relationships like Jane Street)
• Execution moat: the "gap between signed and delivered" is CoreWeave's secret sauce; most competitors can't execute at scale

Weaknesses (from a thesis-input perspective)

• Customer concentration still a risk (Microsoft + AI labs dominant)
• Massive leverage; rising powered-shell constraints could limit deployment
• Secondary to the pure picks-and-shovels plays (TSMC, AMAT, BESI) in a declining GPU-delivery world
• Not a pure-play long thesis for the project's current signal-feed instruments (CRWV is a neocloud, not a semis/energy/materials company)

Related

• ai-capex-to-power-and-materials-cascade
• datacenter-construction-electrical-picks-shovels
• hbm-supply-bottleneck
• brannon-mcbee
• ai-capex-derate-to-private-credit-contagion
• blue-owl
• energized-vs-announced-campus-mw
• colossus-cluster-clocks

Sources

• 2026-06-08-podcast-odd-lots-how-coreweave-sees-the-market-for-compute-right — Contract duration 5yr; inference >50%; powered shell bottleneck; GPU non-fungibility; $21B YTD financing; 1 GW active power; Vera Rubin testing racks received.
• 2026-08-17-autoresearch-ai-capex-private-credit-contagion — DDTL 4.0 Blackstone Credit & Insurance anchor; rating-follows-offtake; DDTL 5.5 SOFR+550 blowout; no public-BDC lender named.
• 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity — Q2 IR 1.5 GW active / ~3.7 GW contracted (6/30); call >300 MW in June, YE >1.85 GW, ≥8 GW by 2030, Indonesia 360 MW ~18 months. Fleet, not a campus clock.
• 2026-09-24-feed-semianalysis-clustermax-30-the-industry-standard (multi-context, partial) — ClusterMAX 3.0: CoreWeave still Platinum / technical bar; Nebius joins Platinum. Not a new chain. Do not emit NBIS.