Question: Is there a real neocloud thesis, which names are actually defensible, and is the −25% to −47% cohort drawdown a buyable pullback or the start of something worse?

Verdict: The under-supplied AI-compute thesis is intact — but the neocloud equities are dangerous instruments: levered, hyperscaler-concentrated, circularly financed, and they break fast with no distribution warning. Own the thesis carefully. Prefer the structurally-defensible power-landlord layer (IREN) over the levered GPU-renters (CRWV); the drawdown is buyable on the thesis but must be sized for the fast-break risk — this is a scale-in-on-weakness special situation, not a clean trend-hold.

What we're asking

The whole neocloud cohort is down 25–47% from its highs, Meta just announced it will sell its own excess AI compute ("Meta Compute"), and the sector's financing is starting to look circular. So: is the thesis broken, or is this the buyable pullback the desk called on July 1? And within a cohort that trades as one block, what actually separates the names?

What we found

The cohort — one label, two very different businesses

They move as one block — but they aren't one business. Two models sit under the "neocloud" label:

• GPU-renters (CRWV, NBIS) rent out NVIDIA capacity to a few enormous customers. Capex-heavy, hyperscaler-concentrated, and financed with debt.
• Power-landlords / shell suppliers (IREN, and the miner-pivots APLD/CORZ/WULF/CIFR) supply the scarcer input — energized real estate and megawatts — and lease compute in on top. They're suppliers to the stack, so no single competitor's compute directly undercuts them.

The bull case is real: AI compute is demand-constrained, not supply-constrained

The load-bearing thesis (established July 1) is that new compute supply gets absorbed, not competed away — the sector is under-supplied. The cleanest proof was the SpaceX IPO: a huge block of compute was sold into the market and the neocloud cohort strengthened through it rather than breaking. And the contracts keep coming — Nebius ~$27B and CoreWeave ~$21B from Meta, IREN ~$9.7B from Microsoft — the booking behavior of a supply-short market, not a glut.

The three bear cases, ranked

1. Meta-as-competitor (the July trigger) — weakest. "Meta Compute" sent the cohort down ~15% in a day on the fear that Meta becomes a rival. But Meta is net-short compute for its own Prometheus/Hyperion roadmap; "selling excess" from a short buyer is a thin sliver plus an investor-narrative move to reframe capex as future revenue. It's still a report ("strategy could change"), not a priced product. Overblown.
2. Capex-digestion — the real cohort risk. The genuine threat isn't Meta; it's a broad hyperscaler capex pause that hits every name at once. This is the standing counter-thesis (ai-capex-digestion): same cohort, opposite call. The tell that decides it is whether neocloud bookings/utilization stay tight (absorption wins) or hyperscaler capex guidance rolls over (digestion wins).
3. Circular financing + concentration + leverage — the structural risk that makes them break fast. NVIDIA invests in the neoclouds → they buy NVIDIA GPUs → they pledge those GPUs as collateral for debt (CoreWeave's $8.5B facility is the first investment-grade GPU-backed financing) → the collateral is contracted cash flow from a handful of hyperscalers. Stack that on the concentration (a Meta or MSFT is often the majority of revenue) and the losses — CoreWeave's Q1 10-Q showed $2.08B revenue but a −$0.74B net loss — and you have equities that don't distribute before they break. If a single anchor contract renegotiates, both the debt collateral and the equity crack at once.

The caveat the bulls underweight: the biggest marginal dollar leaked to a competitor-landlord

The neocloud pitch is "frontier labs go multi-source; neoclouds capture the marginal compute dollar." But the largest single chunk of frontier-lab demand disclosed in 2026 — Anthropic's ~$45B — went to xAI as landlord, not to the neocloud cohort. Multi-sourcing is real; that the marginal dollar reliably lands with the listed neoclouds is not proven. TAM-capture is a claim, not yet a fact.

Verdict + reasoning

The sector thesis is intact; the instruments are dangerous — so differentiate within the cohort. "AI compute is under-supplied" survives every bear case except a broad capex pause, and the −25% to −47% drawdown is a buyable pullback on the thesis (the July 1 read holds). But these are not clean trend-holds: leverage + circular GPU-backed financing + hyperscaler concentration means they break fast, with no distribution warning, and they V-recover rather than base — so waiting for a base means never entering, and chasing the spike is what just cost buyers 25–47%.

The differentiated call:

• IREN — preferred. The power-landlord model is the defensible layer: it supplies the scarce input (energized megawatts) and isn't undercut by any single competitor's compute. Best risk-adjusted way to own the thesis.
• NBIS — the quality GPU-renter. Biggest cap, full-stack, sovereign-EU positioning, the largest anchor (~$27B Meta) — but also the largest absolute concentration. The blue-chip of the risky half.
• CRWV — highest beta, highest risk. Deepest drawdown (−47%), deepest losses, most leverage ($8.5B GPU debt), most concentration. The purest expression of both the bull thesis and the circular-financing risk. Trade it knowing that.
• APLD / CORZ / WULF / CIFR — the speculative shell/miner-pivots. Smaller, less proven; a basket sleeve, not core.

Watch-triggers (the tells that adjudicate this):

• Hyperscaler capex guidance — the single datum that decides absorption vs digestion. A roll-over hits the whole cohort.
• Any Meta / MSFT contract renegotiation or non-renewal — this is the one that breaks the debt collateral and the equity together. The fast-break trigger.
• "Meta Compute" ships as a priced product — only then is the competitor bear case real.
• Neocloud booking / utilization disclosures next quarter — tight = absorption confirmed; softening = digestion.

Sources: Meta Compute selloff and contract sizes — Yahoo Finance, 24/7 Wall St. Circular financing + the $8.5B GPU-backed facility — io-fund. Business-model split — The Diligence Stack. Builds on the desk's July 1 demand-absorption read and the July 6 power-operators investigation.