$META

Summary

The analyst argues that Meta's plan to sell excess AI compute capacity is not bearish for neoclouds like Nebius and CoreWeave, expecting Meta to increase capex and build a cloud business in the future, supporting a long stance on these stocks.

Reasoning

Meta has excess older generation compute capacity it will resell while renting new capacity from Nebius under a multi-year deal, meaning Meta's excess supply is marginal and neoclouds have contracted revenue insulating them from impact; if selling compute proves profitable, Meta will increase capex to build a cloud business, benefiting neoclouds.

Paradis Research | $META the neocloud Meta may build a cloud business to sell excess AI compute capacity. TLDR: - $META selling excess compute because they literally have no roadmap to use old gen capacity - This is not bearish for neoclouds like $NBIS - $META would instead dump old Hopper/Blackwell it overbought while renting new Vera Rubin capacity from $NBIS for 2027. Just to frame Meta's own scale since it matters for what "excess" actually means: - Meta has guided 2026 capex to ~$135B, up ~75% YoY. - They've said they'll buy "millions" of $NVDA Blackwell + Rubin GPUs this year alongside Grace CPUs. - While simultaneously scaling MTIA, their in-house training/inference accelerator designed w/ $AVGO. Two implications: 1. At that spend, even a low % of "overbuilt" capacity is a multi $billion pool of resellable compute. 2. A cloud business is years from positively impacting Meta's rev mix which is still mainly from ads. But naturally, the market has sold off the neoclouds today out of fear: - $NBIS -14% - $CRWV -13% However, the threat to the incumbent neoclouds has been massively overstated ever since the news came out. I've been cringing reading bearish posts. Per $NBIS March release, the two signed a 5 yr deal worth up to ~$27B: - $12B of dedicated Vera Rubin capacity delivering early 2027 - Up to $15B of flexible capacity that Nebius intends to sell to third parties, with Meta buying whatever is left over as a backstop. So Meta is simultaneously renting frontier compute from Nebius and floating the resale of its own. Pretty funny lol. Meta wants guaranteed compute allocation speed which Nebius with their Nvidia partnership can deliver. While the capacity it would resell is older generation kit (Hopper, early Blackwell) it over ordered. That segmentation dissolves most of the paradox and probably tells you the resale pool is marginal + ageing supply. Naturally, the reflex reaction is to think that Meta will crush GPU pricing + breaks the compute shortage thesis: If Meta tip from net renters (Microsoft rents from $CRWV and Nebius exactly because they're short on capacity) to net sellers, the marginal price of rented compute softens, and the neocloud and neoclouds get impacted. But the bulk of the neocloud revenue is contracted and pre-paid over multi year deals. Nebius's ~$50B backlog and CoreWeave's own book anchored by Microsoft insulate them. And bitcoin miners like $IREN, $CIFR, $CORZ, $WULF, $GLXY etc sit further out on the risk curve, with more spot exposure. Across the basket, Meta's resellable excess is just a marginal addition to supply rather than a flood like some will have you believe. Then if Meta prove internally that selling excess compute is more profitable than ads... They'll almost certainly increase capex next year to Google/Amazon levels to go build a cloud biz in full.

Return by horizon