Skip to content

Industry thesis · 2026Q3

Interpretation

Neoclouds are becoming the balance-sheet and power extension of the hyperscalers — and the market is splitting three ways

As of 25 Sept 2026 · predictions stored in the immutable ledger

As of 25 September 2026 the Neocloud industry is defined less by GPUs than by energized megawatts and cost of capital. The largest players (CoreWeave, Nebius, Crusoe, Nscale, IREN) are converging on multi-year, take-or-pay supply to a handful of buyers — above all Microsoft, Meta, OpenAI and Anthropic — financed with increasingly institutional, contract-backed debt. A second group (Lambda, Together AI, RunPod, Lightning AI) competes on developer experience and software. A third, fast-growing layer of former bitcoin miners (Core Scientific, TeraWulf, Cipher, Hut 8, Galaxy, Applied Digital) has become the powered-shell landlord for both. Announced pipelines vastly exceed deliverable capacity; the next 24 months will separate operators who convert contracts into operational MW from those who announced them.

Market structure

  • Contract-backed capacity suppliers: CoreWeave ($104B backlog, 1.5 GW active), Nebius (Microsoft $17.4B, Meta ~$3B), Nscale (Microsoft SOWs up to ~$43.8B), IREN (Microsoft $9.7B), Crusoe (4.9 GW contracted incl. Oracle/OpenAI and Microsoft campuses).
  • Developer/software-led clouds: Lambda, Together AI (inference), RunPod, Lightning AI (post Voltage Park merger). Smaller revenue, higher customer diversity, lower capital intensity per dollar of revenue.
  • Powered-shell landlords (category G): former miners leasing to neoclouds/hyperscalers under 10-15 year leases (Core Scientific ~590 MW to CoreWeave; Galaxy Helios 526 MW to CoreWeave; TeraWulf ~360 MW, Cipher 300 MW and Hut 8 245 MW to Fluidstack; Cipher 300 MW to AWS; Applied Digital 400 MW to CoreWeave).
  • Sovereign programs: Stargate UAE (G42), HUMAIN (Saudi), Yotta (India), Firmus (Australia), Nscale/Aker (Norway), Fluidstack (France MoU).

Power and GPUs

  • Power is the binding constraint: the scarce asset is grid-connected, energized MW delivered on schedule. Companies that control interconnects (IREN Sweetwater 2 GW, Crusoe Abilene with onsite generation) have structural advantage.
  • GPU flow is dominated by NVIDIA Blackwell Ultra (GB300 NVL72) in 2026 contracts; AMD has a single meaningful neocloud champion (TensorWave, 8,192 MI325X online) and Vultr; Rubin appears in 2027 plans (Volta Norway).
  • Headline GPU numbers are mostly contracted or announced, not operational — the observatory tracks the maturity ladder explicitly.

Demand

  • Microsoft is the single largest disclosed buyer across the universe (Nebius, Nscale, IREN, Lambda, CoreWeave, Crusoe Abilene). This is sector-wide concentration risk: a change in Microsoft's build-vs-buy stance would reprice the sector.
  • Anthropic's capacity is increasingly delivered via Fluidstack on Google-backstopped leases; OpenAI via Oracle/Crusoe (Stargate) and CoreWeave; Meta via CoreWeave and Nebius.

Financing

  • GPU-backed debt has institutionalised: CoreWeave's DDTL 4.0 ($8.5B) was rated investment grade; IREN financed $3.6B of GPUs at <6%; Nebius issued converts at 1.00%/2.75%.
  • Hyperscaler credit support is a new financing primitive: Google backstops (~$3.2B at TeraWulf, $1.4B at Cipher) make neocloud leases bankable without Google owning the capacity.
  • Private valuations re-rated sharply in 2026: Crusoe $10B → $30.9B, Fluidstack $7.5B → >$18B, Nscale $14.6B pre-IPO. Public comps are volatile (aggregator market caps for CoreWeave diverge by >20% in the same week).

Consolidation

  • Consolidation has begun at the small end: Lightning AI + Voltage Park (Jan 2026), Rumble + Northern Data (Jun 2026), Brookfield acquires Ori → Radiant (Feb 2026), Volta acquires Genesis Cloud technology.
  • The failed CoreWeave–Core Scientific deal (Oct 2025) shows landlords may prefer independence while lease values are rising.

Five-year scenarios

Probabilities are not exclusive — consolidation can happen alongside the supplier scenario.

  • Neoclouds become permanent hyperscaler competitors

    Requires software ecosystems and diversified demand; only CoreWeave/Nebius plausibly.

    20%

  • Neoclouds become long-term infrastructure suppliers to hyperscalers/labs

    Base case: contract-backed capacity with utility-like economics.

    45%

  • Significant consolidation (top 5 absorb or merge with many of the rest)

    Not exclusive with the supplier scenario.

    55%

  • Move up the stack into software/inference platforms

    Nebius Token Factory, CoreWeave W&B, Crusoe managed inference are early signs.

    30%

  • One or more become AI-era hyperscalers (>$50B revenue)

    CoreWeave is the only credible candidate on current trajectory.

    12%

  • Capital-structure stress causes at least one major failure

    Most likely among highly levered, single-customer operators if AI capex decelerates.

    30%

Closest attention

  1. 1. CoreWeave

    Largest contracted backlog ($104B) and operator benchmark; its delivery vs 1.85 GW YE target and margin trajectory set the sector's cost of capital.

  2. 2. Nebius

    Most vertically integrated listed neocloud with the strongest balance sheet ($8B cash) and now Platinum-rated; 5 GW contracted target makes 2027 an execution test.

  3. 3. Crusoe

    Controls power development at GW scale (Abilene ~2.1 GW footprint); $30.9B valuation makes it the bellwether for energy-first AI infrastructure.

  4. 4. Nscale

    S-1 filed 2026-09-18: first audited look at a two-year-old company with up to $43.8B of Microsoft SOWs against $140.6M H1 revenue — the purest test of contract-to-capacity conversion.

  5. 5. Fluidstack

    Asset-light operator of Anthropic capacity using Google backstops — a financing model that could be copied or could concentrate risk.

Career perspective

  1. 1. Nebius

    Builds its own servers, data centers and software (Soperator, Token Factory, Aether) — breadth of hard technical problems across the stack; strong cash position lowers survival risk.

  2. 2. Crusoe

    Rare combination of energy development, construction and cloud software; rapid capital inflow ($3.9B Series F) implies expanding scope and responsibility.

  3. 3. CoreWeave

    Largest-scale operations (≈50 data centers, 1.5 GW) and best-rated cluster operations; public-company stability with high growth, though leverage is high.

  4. 4. Nscale

    Very early (founded 2024) with enormous contracted scope in Europe and the US — outsized responsibility per employee, with correspondingly high execution and financial risk.

  5. 5. Together AI

    Software/inference-centric neocloud with >$1B reported revenue run-rate and diversified developer customers; exposure to model-serving systems research rather than construction.

What would change this view

  • Microsoft, Meta or Google announcing a reduction in third-party capacity purchases.
  • A sharp drop in GPU-hour prices for Blackwell-class capacity (would stress spot-oriented clouds and GPU collateral values).
  • Investment-grade ratings spreading to more neocloud debt (would accelerate the supplier model and favour scale).
  • Evidence of large-scale delivery slippage at landlord campuses (Q4 2026 - H1 2027 delivery wave).

See the 22 predictions behind this thesis, with probabilities and resolution criteria →