Contracted Backlog & RPO

Backlog — reported variously as cloud backlog or Remaining Performance Obligations (RPO) — is the demand-side counterpart to capex. Capex tells you what an operator is spending; backlog tells you what customers have contractually committed to pay. For a beat whose central risk is that hundreds of billions of dollars of capacity gets built ahead of the revenue to fill it, backlog is the closest thing the public data offers to an answer to "who pays the rent."

The 2026 backlog figures in this KB are large and growing fast. Oracle reports $523B of remaining performance obligations behind a $50B 2026 capex program. Alphabet's Google Cloud backlog was reported at $240B (+55%) as of the Futurum piece in February 2026, and at ~$462B in the July 2026 Alphabet preview — described there as having roughly doubled sequentially in Q1. Microsoft's figure is reported inversely, as a constraint: ~$80B of unfulfilled Azure backlog attributed to power limits, which makes backlog and the power bottleneck two views of the same shortage.

The honest caveats are large. First, backlog is a contractual commitment, not cash — duration, cancellability, counterparty quality, and concentration all matter and none of them are in this KB for any of these names. Second, the numbers are not comparable across companies: RPO, "backlog," and "unfulfilled backlog" are different constructs with different disclosure rules. Third, the Alphabet figure specifically comes from a paywalled preview supplemented by web-search snippets — it is weakly sourced. Backlog is the right question; the KB does not yet have the quality of data to answer it.

The backlog stack, one layer down — and the double-counting edge (added 2026-07-23)

The KB now holds a backlog figure at the neocloud layer, not just the hyperscaler layer. CoreWeave's Q1 2026 primary reports a $99.4B revenue backlog behind $2.08B of quarterly revenue and $7.70B of quarterly capex — a neocloud-layer demand number to set beside Oracle's $523B RPO and Alphabet's ~$462B cloud backlog. That third layer is what makes the double-counting hypothesis testable for the first time.

The backlog-double-counting edge (both ends held; overlap not yet quantified). CoreWeave's Q1 release names a $21B Meta commitment inside its $99.4B backlog — and Meta is itself a tracked hyperscaler in this KB. That is the first case where the KB holds both ends of a single edge: the same end-demand can book as a hyperscaler's own capex/commitment and as the neocloud's contracted backlog (and, one layer further, as the neocloud renting capacity that a hyperscaler also counts). Meta's $21B is direct evidence the same dollars appear at two layers. The overlap is not yet quantified — the release does not disclose how much of CoreWeave's $99.4B is contracted to names the KB already counts elsewhere (Microsoft/OpenAI/Meta), and the hyperscaler side does not net out capacity it rents from neoclouds. Until that netting exists, aggregate "industry backlog" figures should be treated as gross of inter-layer double counts, and this edge (Meta ↔ CoreWeave) is the one concrete, sourced instance of the mechanism.

Key Claims

• Oracle: $523B in remaining performance obligations, backing a $50B 2026 capex program (+136% YoY). Evidence: weak (single analysis source) (Futurum AI Capex 2026)
• Microsoft: ~$80B of unfulfilled Azure backlog attributed to power constraints — demand booked that cannot be served. Evidence: weak (single analysis source) (Futurum AI Capex 2026)
• Alphabet/Google Cloud backlog: $240B (+55%) as of Feb 2026 per Futurum. Evidence: weak (single analysis source) (Futurum AI Capex 2026)
• Alphabet/Google Cloud backlog: ~$462B as of the July 2026 preview, described as roughly doubling sequentially in Q1 2026. Evidence: weak (summary-derived from web search inside a paywalled analysis piece) (Alphabet Q2 2026 preview)
• CONFLICT — the two Alphabet backlog figures ($240B vs ~$462B) are not reconciled in-source. The gap is plausibly a timing effect (Feb 2026 vs Jul 2026 reporting vintages), but neither source states the reconciliation, and the KB does not have Alphabet's own disclosure. Evidence: n/a (unresolved conflict)
• Backlog is the watched signal, not headline EPS — the Alphabet preview frames capex guidance and backlog trajectory as the "key catalyst," above revenue/EPS beat-or-miss. Evidence: weak (single analysis source, market-expectation framing) (Alphabet Q2 2026 preview)
• Hyperscalers describe themselves as supply-constrained, not demand-constrained — the management framing that backlog figures are meant to evidence. Evidence: weak (single analysis source characterizing management commentary) (Futurum AI Capex 2026)

Added 2026-07-23

• CoreWeave (neocloud layer): $99.4B revenue backlog at Q1 2026, behind $2.08B quarterly revenue and $7.70B quarterly capex — the first backlog figure below the hyperscaler tier. Evidence: moderate (first-party company release; 10-Q is the fuller filing) (CoreWeave Q1 2026)
• BACKLOG-DOUBLE-COUNTING EDGE — Meta's $21B commitment sits inside CoreWeave's $99.4B backlog while Meta is itself a tracked hyperscaler — direct, sourced evidence that the same end-demand books at two layers. The KB now holds both ends of one such edge; the overlap across the full hyperscaler→neocloud→model-builder stack is not yet quantified by any source. Evidence: moderate (first-party; the $21B and the backlog are both in the release) (CoreWeave Q1 2026)
• CoreWeave contracted power: >3.5 GW (>1 GW active) — the power-denominated form of the same backlog. Evidence: moderate (first-party company release) (CoreWeave Q1 2026)
• CoreWeave backlog is duration-, cancellability- and concentration-blind, same as the hyperscaler names — the Microsoft/OpenAI share is not disclosed in the release. The defect the KB flagged for Oracle/Alphabet applies identically here. Evidence: n/a (gap, on a primary source) (CoreWeave Q1 2026)
• [UNRESOLVED BY THE PRINT] Alphabet's Q2 2026 8-K carries no RPO/cloud-backlog line — the watched ~$462B is a 10-Q item; the prior $240B-vs-$462B conflict stays open. Evidence: n/a (absent from the primary release) (GOOGL Q2 2026)

Benchmarks & Data

• Oracle RPO: $523B (against $50B 2026 capex, +136% YoY)
• Microsoft unfulfilled Azure backlog: ~$80B (attributed to power constraints)
• Google Cloud backlog: $240B (+55%, per Feb 2026 source) — conflicts with ~$462B (per Jul 2026 source)
• Backlog duration, cancellability, counterparty mix: GAP — none disclosed in any source, for any company (including the CoreWeave primary)
• Amazon/AWS backlog: GAP — no figure in any source
• Meta: no standalone Meta backlog figure — but Meta's $21B commitment appears inside CoreWeave's backlog (double-counting edge, below)
• Neocloud backlog (CoreWeave): $99.4B (Q1 2026) — filled 2026-07-23 (CoreWeave Q1 2026)
• Backlog-double-counting edge: Meta $21B ⊂ CoreWeave $99.4B backlog; Meta also a tracked hyperscaler — both ends held, overlap not yet quantified

Open Questions

• What is the average duration and cancellability of these backlogs? A $523B RPO over 10 years with weak counterparties is a very different asset from the same figure over 3 years with investment-grade ones. This is the single most decision-relevant gap in the topic.
• How concentrated is each backlog by customer? OpenAI-scale counterparties can make one name's backlog effectively a single-credit exposure.
• Are these figures mutually consistent, or does the industry double-count? A hyperscaler renting to a neocloud renting to a model builder can produce backlog at each layer against the same end demand. (Materially advanced 2026-07-23: the KB now holds one concrete edge — Meta's $21B commitment sits inside CoreWeave's $99.4B backlog, and Meta is itself a tracked hyperscaler. The mechanism is now sourced; the magnitude of overlap across the stack is still unquantified.)
• Which reconciliation explains Alphabet's $240B → ~$462B move: organic growth, a disclosure-definition change, or a source error?
• Does Microsoft's ~$80B power-constrained backlog convert to revenue once power arrives, or does it churn to a competitor first?
• What is the ratio of aggregate industry backlog to aggregate 2026 capex, and is that ratio improving or deteriorating?

Related Concepts

• Hyperscaler Capex Cycle — backlog is the demand-side test of whether that capex is justified or front-running
• Power: The Binding Constraint — Microsoft's unfulfilled backlog is the power bottleneck stated in revenue terms
• Data-Center Construction as a Macro Variable — if backlog is real, the macro exposure is durable; if it is not, the GDP contribution is a construction pulse

Backlinks

Pages that reference this concept:

• Alphabet
• Oracle
• Hyperscaler Buyers
• CoreWeave

Changelog

• 2026-07-22 — Created from 2 sources (futurum-hyperscaler-capex-2026-690b, alphabet-q2-2026-earnings-preview). Opens the "who pays the rent" demand-side thread; records the unresolved Alphabet backlog conflict ($240B vs ~$462B) and the total absence of duration/concentration data across all names.
• 2026-07-23 — Added the neocloud backlog layer from the CoreWeave Q1 2026 primary ($99.4B backlog, >3.5 GW contracted). Recorded the backlog-double-counting edge: Meta's $21B commitment sits inside CoreWeave's $99.4B backlog while Meta is itself a tracked hyperscaler — both ends now held, overlap not yet quantified. Confirmed Alphabet's Q2 8-K carries no RPO line (conflict stays open).