CoreWeave and Nebius — Two Divergent Paths for GPU Clouds

Treating 2026’s AI capex as a single business, the two most striking listed names are these two Neoclouds. CoreWeave (NASDAQ:CRWV) began as an Ethereum mining operation in New Jersey in 2017, IPO’d in March 2025, and posted $2.1B revenue in Q1 2026 with a $99.4B backlog. Nebius (NASDAQ:NBIS) is the parent that emerged in 2024 after Yandex N.V. divested its Russian operations — registered in Amsterdam, run by former Yandex founder Arkady Volozh — and had reached $1.2B ARR by year-end 2025. Its Q1 2026 results will be released on May 13. Same Neocloud model, almost diametrically opposed origins. This piece is organized along four axes: founding history, founding team, business status, and the essential difference from Hyperscalers. All data drawn from the most recent public disclosures.

Overview — different starting points · same destination as GPU factories

Same business, two paths — CoreWeave uses a commodities-trader mindset, packaging each GPU cluster into an SPV and issuing asset-backed debt against customer take-or-pay cash flows; Nebius inherits Yandex’s engineering team + ~$2.5B in spinoff cash + NVIDIA’s $2B strategic equity, running a “lighter” balance sheet.

— Don’t just look at Neocloud revenue growth; look at capital structure.

Founding history — mining farm vs search engine

CoreWeave: from Ethereum miner to AI training infrastructure

Nebius: from Yandex international business to independent AI cloud

The “original sin” and “dividend” of each path — CoreWeave’s dividend is the commodities trader’s instinct for capital structure, with mining-era GPU procurement, site selection, and power negotiation directly reusable in AI; Nebius’s dividend is the inheritance of a 1,300-person Yandex engineering team that has already built hyperscale search/ads/cloud, with Mäntsälä not a new build but an asset operating since 2014. The former uses money to solve problems, the latter uses people.

Founding team — commodity traders vs Russian internet veterans

CoreWeave — three commodities traders + one engineer

• Michael Intrator · CEO / Chairman · BA political science (Binghamton) + Columbia MPA; previously ran Hudson Ridge Asset Management (natural gas hedge fund)
• Brian Venturo · CSO · Hudson Ridge co-founder; Bloomberg estimates personal net worth ~$6.4B (2025)
• Brannin McBee · CDO · energy commodities trader background; personal net worth ~$4.7B
• Peter Salanki · CTO · engineering background, leads GPU cluster networking and Slurm/K8s scheduling stack

Viewing GPUs through a commodities trader’s lens — the three founders treat H100s as commodities like crude oil, natural gas, and electricity: lock in price with long-term contracts, arbitrage residual allocations at high frequency, collateralize inventory for financing. This is the cognitive substrate behind CoreWeave’s three consecutive $8B+ asset-backed loans in 2024-2026. A Hyperscaler CFO would not design a balance sheet this way.

Nebius — Yandex veterans

• Arkady Volozh · founder/CEO · born 1964 in Kazakhstan, founded CompTek in 1989 as a hardware distributor, co-founded Yandex with Ilya Segalovich in 2000; Russia’s “Bezos” of tech; sanctioned briefly by the EU after publicly opposing the war in 2022, sanctions later lifted
• Andrey Korolenko · CIO · Yandex infrastructure veteran, led Mäntsälä data center and RuNet uplink architecture
• Roman Chernin · CBDO · former Yandex business development executive, now Nebius commercial lead

Volozh’s generational advantage — he is one of the few founders who has built “consumer search + advertising + proprietary cloud + autonomous driving + vocational education” all to IPO scale. Nebius’s three non-cloud businesses (Toloka / Avride / TripleTen) look like noise to Hyperscalers, but in Volozh’s portfolio view, they are assets hedging the single GPU cycle.

— Nebius is not a pure GPU bet; this is part of why its valuation multiple diverges from CoreWeave’s.

Business status — scale / customers / capacity / capital

Table 1 · Financial comparison (most recent actuals)

Table 2 · Customer structure

Microsoft is the largest customer at both companies simultaneously — Hyperscalers are building their own capacity (2026 capex $190B) while outsourcing training peaks to Neoclouds. This “base load in-house, peaks outsourced” strategy protects Azure’s margins while preventing OpenAI / Anthropic GPU demand from hitting Azure’s capacity wall. Neoclouds are the externalization vehicle for Hyperscaler capital structure.

Table 3 · Capacity / data center expansion

Table 4 · Capital structure

The meaning of DDTL 4.0 — packaging GPUs + long-term customer contracts to a Moody’s A3 rating means insurance companies, pension funds, and mutual funds — that tier of capital — can compliantly buy “AI capex” debt for the first time. The pool expands from private credit to investment-grade public markets. Each new rating tier unlocks ~50-100 bps lower funding cost, flowing directly to Neocloud margins. Nebius has not yet reached this step, but Microsoft’s $7B upfront payment is essentially the same mechanism — customer prepayment substitutes for IG-rated debt.

Core difference from traditional cloud — Neocloud vs Hyperscaler

Ten-dimension comparison

Three essential differences (expanded)

① Vertical depth vs horizontal breadth

Neoclouds sell nine 9’s on a single GPU cluster — tuning InfiniBand fabric, NVLink, NCCL collectives, bare-metal Linux to the extreme; Hyperscalers sell a complete enterprise IT replacement solution, with GPUs as just a few SKUs. The former is like TSMC, the latter like Foxconn.

② Capital sources

Hyperscalers use 30 years of accumulated free cash flow as internal funding, with GPUs as one capex item across diversified businesses; Neoclouds package each GPU cluster individually into SPVs and issue debt secured by customer contracts — the asset itself is securitized first, then produces revenue. This structure makes Neoclouds relatively more upside-leveraged in low-rate environments versus Hyperscalers, but amplifies downside under rising rates + customer defaults.

③ Customer structure

Hyperscalers’ top 100 customers may not even account for 30% of revenue; Neocloud top 5 customers typically account for 80%+. CoreWeave’s pre-IPO disclosure showed Microsoft alone at 62% of revenue (2024). Even with OpenAI / Meta / Anthropic now in the mix, business still revolves around “top 5 customers.” This is the core reason Neoclouds must trade at a valuation discount.

Why do Hyperscalers themselves also rent from Neoclouds? Speed + capital structure externalization + training peak burst capacity. Microsoft Azure is pulling 2026 in-house capex to $190B while simultaneously signing $10B+ multi-year with CoreWeave and $17.4B 5-year with Nebius. On the surface, this looks like a “build + outsource” hybrid; in essence, it’s letting OpenAI / Anthropic / Azure’s own AI business training peaks be absorbed by the Neocloud layer, protecting Azure’s margin structure + balance sheet.

— This is the true demand function behind 2024-2026 Neocloud valuation expansion.

Investment view — valuation and three risks

Valuation multiples (May 2026)

Three common risks

1. Customer concentration / renewal risk — top 3 customers account for the majority; one adverse renewal is a revenue cliff
2. GPU residual value — Hopper (2023) → Blackwell (2024) → Blackwell Ultra (2025) → Rubin (mid-2026) depreciation curve steepens; ABS pool LTVs shift down
3. Interest rate sensitivity — Neoclouds are debt-driven; every 100 bps uptick in long-end rates compresses equity valuation by ~15-20%

Three differentiating factors

• CRWV = scale + Microsoft / OpenAI / Meta deep bindings + IG-rated ABS channel; story is “purer,” leverage higher
• NBIS = Volozh + Yandex engineering team + Europe / US / Israel multi-region footprint + Toloka / Avride / TripleTen non-cloud assets; story is more diversified, capital structure lighter
• Long term, CRWV is scale beta, NBIS is portfolio alpha; the two are complementary theses within the same industry

Signals worth tracking — 5 triggers

1. CoreWeave Q2 2026 earnings (expected August) — look at OpenAI / Meta / Anthropic order cadence; guidance upgrade headroom
2. Nebius Q1 2026 earnings (2026-05-13) — look at whether Q1 ARR reaches the ~$1.5B implied run rate; Microsoft Vineland DC energizing progress
3. GPU residual value secondary market — monitor H100 8-card hourly and retail prices on RunPod, Lambda, eBay; Hopper depreciation magnitude after Blackwell Ultra mass production
4. Regulatory stance on GPU-backed ABS — analogous to 2007-08 leveraged loans; if SEC / rating agencies start requiring thicker capital cushions, DDTL 4.0 replication costs rise
5. Microsoft in-house capacity ramp — once Azure 2026 capex $190B is fully online, will the Neocloud “peak outsourcing” window narrow? Watch OpenAI 2027 renewal negotiation signals

Essential judgment — the Neocloud story doesn’t hinge on “who do they sell GPUs to,” but on “can the capital structure keep rolling.” CoreWeave has elevated GPU-collateralized loans to IG rating; Nebius substitutes upfront payments for debt and fills equity with NVIDIA strategic stakes — both paths are still working, but this is a long-term bet measured in years, paid for in interest rates and customer renewals.

— The alpha at this layer is capital engineering, not GPU compute itself.

References — primary disclosures · media tracking · industry data

Primary disclosures

• CoreWeave Investor Relations — 2026-Q1 earnings release · DDTL 4.0 $8.5B announcement. investors.coreweave.com
• Nebius newsroom — Microsoft $17.4B agreement · Eigen AI acquisition. nebius.com/newsroom

Media tracking

• TechCrunch — Neocloud and Hyperscaler coverage. techcrunch.com
• CNBC — listed company earnings and AI infrastructure coverage. cnbc.com
• Bloomberg — GPU financing and capex coverage. bloomberg.com
• SemiAnalysis — deep analysis of GPU compute and data centers. semianalysis.com
• Sherwood News — capital markets and tech intersection coverage. sherwood.news

Industry data

• DatacenterDynamics (DCD) — data center and power capacity data. datacenterdynamics.com