AI-Native Neoclouds

GPU-focused clouds built ground-up for AI workloads. The 2023-2026 capex story.

What this category is

Pure-play GPU clouds — companies that buy Nvidia (and increasingly AMD) accelerators, install them in data centers with the cooling and networking modern training requires, and rent them by the hour, by the cluster, or on multi-year reserved contracts. No model API. No enterprise data warehouse. No identity stack. The pitch is simple: raw, contiguous GPU capacity, faster than a hyperscaler can build it.

This is structurally different from Layer 4.1 (the hyperscaler bundle) and Layer 4.3 (inference specialists with custom silicon or a software stack). Neoclouds are not selling a platform — they are selling kilowatts of GPU time, wrapped in just enough scheduling, networking, and storage to make the cluster usable.

The structural thesis

From 2023 through 2026, the binding constraint on AI deployment has been physical: data center power, building shells, and Nvidia allocation. Hyperscalers can spend the capex but cannot conjure 300-megawatt campuses in eighteen months. Frontier labs need huge, low-latency, single-fabric clusters but don't want to be locked into one cloud's roadmap.

Neoclouds exist because that gap is real and persistent. They raise debt against signed customer contracts (or against the GPUs themselves), buy Nvidia silicon, energize a site, and deliver capacity in 6–12 months versus the 24–36 months a greenfield hyperscaler build takes. The hyperscalers buy from them — at one point Microsoft was the majority of CoreWeave's revenue — because it is faster and (sometimes) cheaper to rent than to build.

The implication for investment: a neocloud is a leveraged bet on three things. (1) Nvidia allocation, which determines whether they get any scarce silicon at all. (2) GPU economic life, which determines whether the depreciation schedule on the balance sheet matches reality. (3) Customer concentration, which determines whether a single renegotiation can break the equity. Every other line of the model flows from those three variables.

The players

CoreWeave (CRWV)

The category-defining pure-play. Started life as a crypto miner, pivoted to GPU rental in 2019, IPO'd in March 2025 after raising billions in debt collateralized by Nvidia hardware and signed contracts. Q1 2026 backlog reached $99.4B (up from a few billion at IPO), with 10 customers each committed to $1B+. Revenue doubled YoY to $2.08B in Q1 2026, and 2026 full-year revenue guidance is $12–13B. Contracted power is 3.5 GW, with over 1 GW already active.

The financial profile is unusual for a public company: very high revenue growth (triple-digit YoY in the early quarters), very high gross margin on a cash basis, very heavy GAAP depreciation, and a debt stack that needs to be refinanced as each tranche matures. Free cash flow is structurally negative because every dollar of contracted revenue requires more dollars of new GPU capex.

Two stories matter for CRWV. First, the Microsoft relationship: Microsoft was ~67% of FY2025 revenue and ~72% in 1H 2026, and any softening in MSFT's posture moves the stock more than fundamentals do. Second, the diversification trade. CoreWeave signed a series of direct agreements with OpenAI in 2025 — $11.9B (March), expanded +$4B (May), expanded again +$6.5B (September) — for total OpenAI commitments of up to ~$22.4B. The Meta relationship was expanded to up to ~$35.2B. Anthropic was added as a named anchor customer in Q1 2026. Each new anchor reduces relative Microsoft concentration but stacks new single-counterparty risks on top.

BullLargest pure-play; $99.4B backlog provides multi-year revenue visibility; Nvidia equity stake aligns allocation; OpenAI, Meta, and Anthropic deals diversify the customer book away from Microsoft.

BearConcentration (top counterparties still dominate); the 6-year GPU/server depreciation schedule is more aggressive than peers (Nebius uses 4–5 years) and Burry-style critiques flag it as the biggest single accounting variable; debt refi risk; equity is the residual claimant after a very thick capital stack.

Nebius (NBIS)

Spun out of the international assets of the former Yandex after the 2022 wind-down, re-listed on Nasdaq in October 2024, and re-launched as a full-stack AI cloud. Operates data centers in Finland, the U.K., Israel, and New Jersey, with additional U.S. and European sites under construction. Has secured 2+ GW of contracted power and is guiding toward 3+ GW. Nvidia is an equity holder and made a further $2B investment in March 2026. Also owns Toloka (data labeling) and Avride (autonomy).

Nebius is structurally less levered than CoreWeave — the Yandex spinout left it with cash on the balance sheet that funded the early GPU buildout without taking on equivalent debt. The customer mix re-rated dramatically in late 2025 / early 2026: a ~$19.4B multi-year contract with Microsoft and a ~$27B multi-year contract with Meta have moved Nebius from "diversified mid-market" into the same conversation as CoreWeave, while still keeping anchor concentration below CRWV's. 2026 revenue guidance is $3.0–3.4B, with a medium-term ARR target of $7–9B by end of 2026.

Nebius uses a shorter (and arguably more realistic) GPU/server useful-life assumption than CoreWeave — 5 years versus CRWV's 6 — which depresses reported EBIT today but should age better if Blackwell compresses Hopper economics.

BullCleaner balance sheet; MSFT and Meta anchors at scale; Nvidia partnership and direct equity check; more conservative depreciation; optionality from Toloka / Avride.

BearCapital needs will grow faster than internal cash generation as new sites energize; EBIT remains negative through the buildout; ex-Yandex provenance still carries political tail risk; new MSFT and Meta anchors introduce the same concentration risks CRWV is criticized for.

Lambda (Private)

The original GPU cloud for researchers and developers. Series D in February 2025 raised $480M at a ~$2.5B valuation; a subsequent $1.5B+ Series E led by TWG Global pushed total funding to ~$2.3B. Lambda hit ~$505M ARR by mid-2025 and signed a multi-billion-dollar Microsoft deal to deploy tens of thousands of Nvidia GPUs, plus a $1.5B Nvidia GPU sale-leaseback (18,000 GPUs over four years) that makes Nvidia Lambda's largest customer. Lambda has hired Morgan Stanley, JPM, and Citi for a public listing now targeted for H2 2026. The interesting question is whether Lambda can move upmarket into hyperscaler-anchor revenue (CoreWeave's playbook) without losing the long-tail developer base that built its brand.

Crusoe (Private)

Began as a stranded-gas-to-Bitcoin company, pivoted hard into AI data center development. Best known as the developer of the Abilene, Texas campus that became the first physical site of the Oracle / OpenAI / SoftBank "Stargate" buildout — two buildings live as of late 2025, with six more planned for ~1.2 GW total. The story changed in Q1 2026: OpenAI / Oracle scrapped the planned 600 MW expansion of Abilene after financing and demand-forecasting issues and a multi-day cooling outage strained the relationship. Microsoft has since stepped in to lease ~700 MW of the Crusoe-developed capacity OpenAI walked away from, building two additional "AI factory" buildings at the same Abilene site. Crusoe is increasingly positioned as a developer-landlord — power + shell + cooling — that flips capacity to whichever hyperscaler shows up, closer to a specialty data-center REIT than a traditional cloud.

Applied Digital (APLD)

Originally a Bitcoin-mining infrastructure host (Ellendale, North Dakota), now squarely an HPC landlord. Two ~15-year leases with CoreWeave totalling 400 MW at the Polaris Forge 1 campus — 250 MW announced May 2025, an additional 150 MW added August 2025 — contracted out to roughly $11B of total revenue. Buildings are coming online in waves through mid-2026 and 2027. In March 2026 APLD modified the lease structure to add credit enhancements after CoreWeave itself was re-rated. The thesis is the same: re-rate from miner-multiple to data-center-REIT-multiple as the campus fills with GPU racks. The risk is now single-tenant: APLD is, in practice, a leveraged way to own CoreWeave's North Dakota footprint.

The miner-pivot cohort

• IREN (IREN) — Texas-based, signed a $9.7B five-year agreement with Microsoft (Nov 2025) to deliver Nvidia GPU capacity at Childress across four "Horizon" buildings (200 MW critical IT load). Paired with a ~$5.8B Dell purchase agreement for GPUs and ancillary equipment. Targeting 480 MW of AI cloud capacity and ~150,000 GPUs deployed in 2026, with a $3.7B annualized revenue target by year-end. Among the most credible miner pivots.
• Cipher Mining (CIFR) — 10-year, ~$3B hosting deal with Fluidstack at Barber Lake (Colorado City, TX), covering 300 MW of critical IT load. Google is the financial backstop (~$1.4B of Fluidstack lease obligations) and took a ~5.4% equity warrant in CIFR, not a direct customer. First 168 MW targeted for delivery by September 2026.
• TeraWulf (WULF) — Lake Mariner (Western NY). Core42 (G42) anchor lease of 70 MW + expansion options. Separate 10-year Fluidstack agreement (Google-backstopped) for 200+ MW of critical IT load = ~$3.7B over the initial term, $8.7B if both 5-year extensions are exercised. Total ~522 critical-IT MW under long-term lease.
• Hut 8 (HUT) — Now explicit that bitcoin is no longer a long-term strategic focus. Anchored a $7B 15-year lease at the River Bend campus (Anthropic + Fluidstack as counterparties), then announced a $9.8B 15-year lease at the Beacon Point campus (Texas) in May 2026 with options that could take total value to ~$25.1B. Reported 8.5 GW development pipeline.
• Bitdeer (BTDR)— Singapore-listed, integrated mining + AI cloud + ASIC R&D. Sold its remaining Bitcoin treasury in Q1 2026 to fund AI data center expansion. AI cloud ARR ~$43M as of March 2026 (still small but growing rapidly month-over-month).

The thesis across this cohort is the same: existing land + power interconnect + building shell + ops team, retrofitted from air-cooled Bitcoin racks to liquid-cooled GPU racks, with a hyperscaler or AI-native customer providing credit enhancement. Notably, several of these 2025–2026 deals use a structure where a well-rated hyperscaler (Google, Microsoft) backstops a smaller neocloud counterparty (Fluidstack, CoreWeave) — a hybrid that gives miners investment-grade-quality cash flows while keeping the operator relationship with the neocloud. Still, mining-grade infrastructure (low-PUE air-cooled 5–10 kW racks) is genuinely different from HPC-grade (liquid-cooled 50–130+ kW racks, sub-millisecond network, redundant power feeds). Only some sites are geographically and electrically viable, and execution risk on the retrofit is real.

Private mid-tier and on-demand

Voltage Park (philanthropy-funded, GPU-as-a-service with a research bent) and Fluidstack (long-tail on-demand) round out the visible private operators. There is also a constellation of regional brokers — TensorWave, Hyperstack, Yotta in India — that are economically more "tier-2" than "AI-native." Tier-2 / regional clouds are covered in Layer 4.4.

Side-by-side: the publics

Figures directional. Pull the latest 10-Q / 20-F before sizing positions; RPO and concentration disclosures move every quarter.

Unit economics

A neocloud's income statement is two businesses stapled together: a revenue-recognition business (sell GPU hours under multi-year contracts) and a capex-and-depreciation business (buy GPUs, finance them, depreciate them). Cash margins look great. GAAP margins look terrible. Both are partial truths.

1. Cash gross margin — Revenue minus power, colo, bandwidth, ops. Typically 60–75% on contracted GPU rental at list price. Higher on reserved long-term commits; lower on spot.
2. GAAP depreciation — The accounting choice that determines reported profitability. CoreWeave depreciates GPUs over ~6 years. If the real economic life of an H100 is 3–4 years (because B200 / GB200 / Rubin make it cost-inefficient to operate on training jobs), reported earnings overstate true earnings by a large margin.
3. Interest expense — Debt is collateralized by either contracts (delayed-draw term loans backed by Microsoft RPO) or the GPUs themselves. Coupons in the 9–12% range on recent issuance. Every 100 bps of rate movement matters materially.
4. Maintenance capex — Often described as "we don't need much" but in practice each GPU generation requires new networking (NVLink switches, InfiniBand fabric upgrades), new cooling (air → liquid), and new power density. The cluster cannot be refreshed GPU-by-GPU; it is refreshed cluster-by-cluster.

The right way to look at any neocloud financial is on a per-cluster basis: total invested capital, contracted revenue, depreciation assumption, residual value at end of contract. The equity is the residual claimant after debt service and equipment refresh. In a bull case (4–5 year economic life, strong residual on Hopper into 2028+, contract renewals at flat prices) the IRR is excellent. In a bear case (3-year life, Hopper resale collapses by 2027, customers renegotiate down) the equity gets very thin.

Bull / Bear

What to watch (KPIs)

Quarterly:

• RPO growth and customer mix disclosures — top-customer concentration, weighted-average remaining contract life
• Contracted power capacity (MW under contract / energized / under construction)
• GPU depreciation schedule — any change in useful life assumption is a major signal
• Debt stack composition and weighted-average coupon — refi events matter
• Utilization (when disclosed) — GPU hours sold vs. installed
• Hyperscaler capex commentary — leading indicator for neocloud rental demand. A hyperscaler "moderating" capex growth is bad for neoclouds; "accelerating" is good.

Annual:

• Hopper resale prices in the secondary market (proxies the depreciation debate)
• Microsoft / OpenAI / Meta contract renewals and re-pricing events
• New silicon generation cadence (Blackwell Ultra → Rubin) and its impact on Hopper economics
• Power interconnect approvals at announced sites (a campus without grid sign-off is not a campus)

Players

Players: CoreWeave CRWV, Nebius NBIS, Lambda Private, Crusoe Private, Applied Digital APLD, Iris Energy (IREN) IREN, Cipher Mining CIFR, TeraWulf WULF, Hut 8 HUT, Bitdeer BTDR, Voltage Park Private, Fluidstack Private