Meta is building a cloud unit — and the neocloud selloff looks overdone.
- The ThesisOn Jul 1, Bloomberg reported Meta is building "Meta Compute" to sell excess AI capacity. Meta rose ~9%; CoreWeave (CRWV) fell ~14% and Nebius (NBIS) ~12–15%. Most of the viral claims check out — but the reaction reads as an overreaction to a modest, delayed supply addition, not a structural break.
- The CatalystIt started Jun 28 when the FT reported Google is rationing Meta's Gemini access over its own compute shortage — then Bloomberg's Jul 1 "Meta Compute" scoop lit the fuse, triggering notes from Morgan Stanley, JPMorgan, Bernstein, UBS, Rosenblatt, Cantor and BofA.
- The RiskThe real risk is long-term and structural: Meta is over a third of CoreWeave's backlog and could compete at renewal (2028+). Two viral claims are overstated — CRWV's "$131B backlog" (actual $99.4B; $131B is a Cantor projection) and "90% of 2027 pre-sold" (company says >75% contracted).
Think of AI compute like apartments in a red-hot city. The "neoclouds" — CoreWeave (CRWV) and Nebius (NBIS) — are landlords who built towers full of GPU "apartments" and rent them to AI tenants. Right now every unit is rented, the waitlist runs a year long, and rents are high.
Meta is a giant tenant that rents a huge block of these apartments for its own AI. On Jul 1 it said it might sublet its spare rooms — a new unit called "Meta Compute." Wall Street panicked that a massive new sub-landlord would flood the market and crash rents, so it dumped the landlords (CRWV −14%, NBIS −12%) and bought Meta (+9%).
Why that panic looks overdone: Meta only has a few spare rooms, it legally can't sublet the apartments it rents from CoreWeave or Nebius (the leases forbid it), and the city still has far more tenants than apartments. So near-term the landlords are fine. The real worry is years out — around 2028 — when Meta's big leases come up for renewal and it might build its own towers instead of renewing.
| Date (2026) | Event |
|---|---|
| Mar 16 | Nebius signs 5-yr Meta deal: $12B dedicated Vera Rubin capacity from early 2027, +up to $15B optional (up to $27B total) |
| May 7 | CoreWeave Q1: revenue $2.08B (+112% YoY), backlog $99.4B |
| May 13 | Nebius Q1: revenue $399M (+684% YoY), adj EBITDA +$129.5M, cash $9.30B |
| May 27 | Zuckerberg: selling excess compute is "definitely on the table" |
| Jun 28 | FT: Google limits Meta's Gemini usage over Google's own compute shortage |
| Jul 1 | Bloomberg: Meta building "Meta Compute." META +9%; CRWV −14% (→$85.69); NBIS −12% |
| Jul 2 | CRWV slides further (~$81, ~$38B cap); its junk bonds fall |
Why watts? The industry sizes AI data centers by IT power — the megawatts (MW) or gigawatts (GW) delivered to the racks — not by floor space or GPU count, because power is the one thing you can't fake or rush. Grid hookups, substations and cooling take years; that supply of electricity, not real estate or even chips, is the binding constraint. So everything gets normalized "per watt": build cost per watt, lease revenue per watt.
The economy of watts, in plain terms. One GW is a billion watts — roughly a large nuclear reactor, or the draw of ~750,000 homes. In AI it's shorthand for one hyperscale campus. Building that GW all-in — shell, power, cooling and the GPUs inside — runs about $35–50B, and the GPUs are ~60–70% of it and need replacing every 3–5 years. Rent the same GW out as full compute (GPUs included) and it can bring ~$20–40B per year. A hot GW can therefore pay back its build cost in roughly one to two years — which is exactly why everyone is racing to build them, and why "who controls the watts" is the whole game.
The catch — and where Meta fits. That payback math only holds while chips are scarce and utilization stays high. Bare-metal (raw servers, no software) — what Meta would resell — sits at the bottom of the $20–40B/GW/yr range; full-stack, high-touch enterprise inference sits at the top. At the retail level a single GPU-hour runs H100 ~$2–7/hr (median ~$3), B200 ~$2–6/hr — the same tiering, one chip at a time.
Every major supply datapoint as of early July says tight, not glut:
| Signal | Evidence |
|---|---|
| On-demand GPUs sold out | SemiAnalysis: on-demand rental capacity sold out across virtually all GPU types; reserved booked into Aug 2026 |
| Long lead times | ~36–52 weeks; Blackwell allocations slipped to Q1 2027. Constraint = CoWoS-L packaging + HBM, not fabs |
| Even Google is short | Rationing Gemini to customers AND paying SpaceX ~$920M/mo for ~110,000 GPUs as a bridge |
| Inference is the engine | Now ~60–70% of hyperscaler AI demand (up from ~40% in '24); agentic workloads compounding it |
| Capex still rising | 2026 guides: AMZN ~$200B · GOOG $175–185B · MSFT ~$150B · META $115–145B — all above prior |
| Buyers keep committing | Microsoft alone >$60B across Nscale, Nebius, CoreWeave, IREN & Lambda |
The honest caveat (from JPMorgan's own desk): whether demand is still steepening only gets confirmed in the July–August earnings season. Positioning was crowded going in, which amplified the drop.
Two names took the selloff head-on, and one small-cap sits alongside them as the higher-risk, higher-torque play. The tell is Meta concentration: the more of a company's backlog leans on Meta, the more the "Meta Compute" headline stings.
| Name | Read | Lean |
|---|---|---|
| NBIS Nebius | Least exposed of the two majors — full-stack, high-touch inference a bare-metal Meta offering doesn't touch, and the ~$17–19B Microsoft deal diversifies its book away from Meta. The selloff dinged it on sympathy, not substance. | Buy — the cleaner one |
| CRWV CoreWeave | Still a buy on demand and backlog ($99.4B), but Meta is >1/3 of that backlog and CRWV carries the most renewal + leverage risk (junk bonds sold off). The bull case is intact; the discount is deserved-larger. | Buy — but less so than NBIS |
| WYFI WhiteFiber | The outlier small-cap play — a neocloud carved out of a Bitcoin miner, ramping its NC-1 site toward ~40MW, up sharply since IPO. Barely any Meta exposure; the risk here is execution, liquidity and size, not the Meta headline. Torque, not safety. | Speculative — outlier upside |
- Demand is still strong. GPUs are sold out, waitlists run a year, and backlogs are at records. This is not a market with too much supply.
- Meta barely dents it right now. It has few spare GPUs to sell, and it legally can't resell the ones it rents from CoreWeave or Nebius. The impact through 2027 is tiny.
- The real risk is 2028. That's when Meta's big leases renew — and it may build its own instead. CoreWeave is most exposed; Nebius least.
Bottom line: the selloff looks overdone. Compute is still scarce, and Meta's threat is a 2028 story, not a 2026 one.
Want the full neocloud model — CRWV vs NBIS renewal risk?
Join the Discord to find out! →FT / CNBC / Bloomberg / Forbes / DCD (Google–Meta Gemini rationing, Jun 28) · Bloomberg / CNBC ("Meta Compute," Jul 1) · Morgan Stanley, UBS, Bernstein, JPMorgan, Rosenblatt, Cantor Fitzgerald, BofA notes (press/desk summaries) · SemiAnalysis GPU Shortage Report (Mar 2026) · CoreWeave Q1 2026 & Nebius Q1 2026 (6-K) filings · Giga Energy / Encor / Epoch AI / Data Center POST (power & lease economics). Reformatted from the "Meta Compute & the Neocloud Trade" research report, Jul 2, 2026.