CoreWeave

The hyperscaler of GPU clouds — NVIDIA's preferred neocloud, backing the AI buildout via debt-financed capex against multi-year hyperscaler contracts.

1. Core Product / Service

GPU cloud platform built natively on Kubernetes, with InfiniBand-interconnected multi-thousand-GPU clusters for LLM training and inference [1][2]. Originally founded in 2017 as Atlantic Crypto (Ethereum miner); pivoted to AI infrastructure when GPU demand shifted [1].

Stack and capabilities:

• ~250,000 NVIDIA GPUs across ~43 data centers, ~850 MW active power as of FY2025 [2]
• Mix includes H100, H200, GB200/Blackwell — gets allocation priority on each new NVIDIA generation
• Bare-metal Kubernetes, persistent parallel storage, custom VM shapes
• Post-acquisition: Weights & Biases experiment-tracking + model lifecycle stack folded in (May 2025) [3]
• Pending Core Scientific deal adds ~1.3 GW gross power + ~1 GW expansion runway (announced July 2025, all-stock ~$9B) [4]

2. Target Users & Pain Points

Primary buyers are hyperscalers and frontier labs that cannot get enough GPU capacity from their own internal builds:

• Hyperscalers absorbing overflow — Microsoft alone was ~67% of FY2025 revenue [5]
• Frontier labs needing dedicated multi-thousand-GPU clusters — OpenAI signed up to ~$22.4B in cumulative commitments across deals in Mar / May / Sep 2025 [5]
• Enterprises and AI startups wanting H100/Blackwell capacity without waiting in AWS/Azure quota queues

Pain points solved: NVIDIA allocation (CoreWeave gets new SKUs early), faster cluster spin-up than legacy hyperscalers, GPU-hour pricing historically below AWS/Azure list [2].

3. Competitive Landscape

CoreWeave's differentiation: scale + NVIDIA proximity + capital-markets sophistication. Competitors don't have the same GPU-collateral debt machinery.

4. Unique Observations

• Business is as much capital-markets engineering as datacenter engineering. CoreWeave finances GPUs via debt collateralized by signed hyperscaler contracts — buy GPU, sign Microsoft/OpenAI multi-year, borrow against the contract, deploy more GPU. In March 2026 closed an $8.5B facility rated A3 (Moody's) — the first investment-grade rated GPU-backed financing [6]. ~$28B of equity+debt commitments in 12 months [2].
• Customer concentration is the central risk. Microsoft ~67% of FY2025 revenue [5]; Meta now ~40% of the $87.8B reported backlog [2]. If any single hyperscaler in-sources, the model breaks.
• NVIDIA backstop is structural. NVIDIA agreed to purchase any unused capacity under a $6.3B agreement [2] — converts CoreWeave's leasing risk into a put option on NVIDIA itself. Combined with NVIDIA's ~13% equity stake post-Jan-2026 $2B investment at $87.20 [2], the entity is functionally NVIDIA's external balance sheet for AI capacity.
• Customer commitments are structurally rigid. CoreWeave's backlog ($87.8B) is backed by multi-year take-or-pay contracts with hyperscalers and frontier labs. These are not soft reservations — they represent contracted, enforceable revenue commitments that underwrite the debt financing [local: 2026-05-24]. The rigidity of these contracts is what enables the GPU-collateralized debt model: lenders see signed hyperscaler commitments, not projected demand.
• NVIDIA's share in CoreWeave's debt structure is extraordinary. A significant portion of CoreWeave's debt financing flows directly to NVIDIA for GPU procurement — effectively, NVIDIA lends (via debt facilities) to CoreWeave to buy NVIDIA GPUs, which then serve as collateral for the same loans. The exact percentage of debt attributable to NVIDIA purchases is material but not publicly broken out [local: 2026-05-24].
• DDTL (Delayed Draw Term Loan) is the core debt instrument. CoreWeave uses DDTL structures — committed loan facilities where the borrower draws funds as needed rather than all at once — to match GPU procurement timing with capital deployment. This avoids paying interest on undrawn capital while maintaining committed purchasing power [local: 2026-05-24].
• The "neocloud" thesis vs reality. Industry narrative says CoreWeave proves neoclouds beat hyperscalers; financial reality says CoreWeave is a hyperscaler's overflow vendor, financed by NVIDIA, with hyperscaler-shaped concentration risk. Relevant context for ai-inference-engines and the broader gpu-kernel-optimization race.
• Miner partnerships provide power, not just capacity. CoreWeave's deal with Core Scientific adds ~1.3 GW gross power, but the miner partnership model goes deeper: bitcoin miners transitioning to AIDC bring locked-in power purchase agreements, existing grid interconnections, and permitted sites — the three scarcest resources in AIDC development. CoreWeave provides the GPU capital and hyperscaler contracts; miners provide the physical infrastructure and power access [local: 2026-05-24].

5. Financials / Funding

• IPO: March 28, 2025 on Nasdaq (CRWV). Priced at $40/share (below $47–$55 range), 37.5M shares, raised $1.5B at ~$23B fully diluted valuation — largest US tech IPO since 2021 [7][8]
• Stock: traded $33.52 low, $187 high; $81 in March 2026 ($43B market cap) [2]
• Revenue: $16M (2022) → $1.9B (2024) → $5.13B (FY2025, +168% YoY) [2][5]
• Backlog: $87.8B reported; Meta ~40.1%, Microsoft and OpenAI dominant [2]
• Capex: $30–$35B planned for full-year 2026; $6–$7B Q1 2026 alone [2]
• Financing: ~$28B equity + debt commitments raised in last 12 months; March 2026 $8.5B facility, first investment-grade GPU-backed [6][2]
• NVIDIA stake: ~13% post Jan 2026 $2B investment (vs ~7% at IPO, ~1.2% three years prior) [2]

6. People & Relationships

• Founders: Michael Intrator (CEO), Brian Venturo, Brannin McBee — original Atlantic Crypto co-founders
• Major investors: NVIDIA (~13% stake, strategic), Magnetar Capital, Coatue, Blackstone (debt), Fidelity, BlackRock
• Key customers (FY2025): Microsoft (67% revenue), OpenAI ($22.4B cumulative commitments), Meta (~40% of backlog), NVIDIA, IBM, Cohere, Mistral [5][2]
• Acquisitions:
  • Weights & Biases — closed May 2025, adds AI dev platform / experiment tracking [3]
  • Core Scientific — definitive agreement July 2025, all-stock ~$9B, expected Q4 2025 close, adds 1.3 GW power [4]
• Strategic partner: NVIDIA — supplier, investor, capacity-purchase backstop