Explained

RPO vs Take-or-Pay: What's the Difference?

RPO is booked cloud not yet recognised as revenue. Take-or-pay bills reserved GPUs after delivery, even idle. Same stuck step: the hall, not the signature.

Sean

— 10 min read
Share
Photoreal GPU hall: an open ledger on a steel table beside a pulled-out server sled — the book is not the bill

EXPLAINED · Versus · Last reviewed: 1 Oct 2026 · Next review: after next hyperscaler RPO print or CoreWeave Q3 2026

Remaining performance obligation versus take-or-pay is the contrast between an accounting remainder of contracted cloud that has not yet been recognised as revenue, and a contract form that bills reserved GPU capacity after delivery whether the hours run or not.

In short. RPO is how the hyperscalers layer shows booked demand that is not yet a P&L line. Take-or-pay is how a neocloud — and sometimes a hyperscaler buyer — turns a reserved block into a cash clock after the hall is live. The stuck step on both clocks is still delivery.

What is the difference between RPO and take-or-pay?

RPO is an accounting disclosure. Under US GAAP (ASC 606), a public company reports the transaction price allocated to performance obligations that are unsatisfied, or only partly satisfied, at period-end — plus when it expects to recognise that amount as revenue (PwC on ASC 606-10-50-13). Oracle’s 10-K: remaining performance obligations “represent contracted revenues that had not yet been recognized,” including deferred revenues, uncollected invoices, and amounts that will be invoiced later (year ended 31 May 2026). Three buckets, one remainder.

Take-or-pay is a commercial form. CoreWeave’s Q2 2026 10-Q sells access “either through committed contracts, which are take-or-pay, or on-demand, which are pay-as-you-go.” Committed contracts “generally have a predetermined term and start either on a fixed date or when we deliver the capacity specified in the contract.” After that switch, the reserved fee still bills if the cluster sits idle.

The same operator can print both. CoreWeave’s Q2 2026 release called the commercial book a revenue backlog of about US$104 billion as of 30 June 2026. The 10-Q’s accounting cut is remaining performance obligations of US$103.7 billion. Take-or-pay is most of how that remainder is written. RPO is how it is counted.

A hyperscaler can sit on either side. It sells committed cloud that fills RPO. It also buys take-or-pay capacity. Microsoft’s FY2026 10-K put purchase commitments at US$194.060 billion as of 30 June 2026, “primarily relate to datacenters and include open purchase orders and take-or-pay contracts.” Stack: What Does Microsoft Do? · What Does Oracle Do?.

Axis Remaining performance obligation Take-or-pay GPU contract
What it is Unsatisfied contracted transaction price (ASC 606) Offtake form: reserved capacity bills after delivery, used or not
Who prints it Hyperscalers; some neoclouds as an accounting remainder Neoclouds sell it; hyperscalers also buy it
Cash Not cash. Deferred revenue is the prepaid slice After delivery, the reserved fee bills. Prepay is a deposit on the term
Idle hours May bill (committed) or not (usage-driven remainder) After the on-switch, idle still bills
On-switch Revenue when the operator performs Term starts on a date or when contracted capacity is delivered
Stuck step Energised halls, installed GPUs, accepted capacity The same hall — undelivered take-or-pay does not yet bill

Why does the mix-up bind?

It binds because the headlines travel together and answer different questions. RPO answers: how much contracted work is still on the books? Take-or-pay answers: who carries unused-hour risk after the block is live? Treating one as the other turns a longer clock into a cash print.

The hyperscaler remainders are not the same cut. Microsoft’s commercial remaining performance obligation was US$678 billion as of 30 June 2026, up 84% (FY26 Q4 release, 29 Jul 2026). The call added a 2.3-year weighted duration, roughly 30% recognised in the next 12 months, and +25% excluding OpenAI. Our arithmetic: 30% of US$678 billion is about US$203 billion in the next year. Amazon’s long-term AWS remainder was about US$496 billion (6.4-year remaining life; usage-driven) (10-Q, 30 Jun 2026). Alphabet’s RPO — it also says “revenue backlog” — was US$519.5 billion, US$513.9 billion of it Google Cloud, just over 50% over the next 24 months (10-Q).

Oracle’s year-end bands are the conversion people skip. As of 31 May 2026 remaining performance obligations were US$638 billion, with about 12% expected as revenue over the next twelve months, 34% over months 13 to 36, 34% over months 37 to 60, and the rest after (10-K). Our arithmetic: 12% of US$638 billion is about US$77 billion next year; 34% is about US$217 billion in the following two years. 12% + 34% = 46%, so about 54% — roughly US$345 billion — sat beyond month 36.

The Q1 FY27 print moved the headline without restating those bands. On 10 September 2026 Oracle said remaining performance obligations rose US$209 billion year-over-year to US$664 billion after booking more than US$30 billion of additional AI cloud contracts (investor release / PR Newswire). Our arithmetic: US$664 billion − US$638 billion = US$26 billion sequential. Bookings can exceed the net remainder when older obligations convert. Do not apply the 31 May bands to the later snapshot.

The neocloud remainder is a different universe. CoreWeave’s US$103.7 billion RPO as of 30 June 2026 expected 41% over the 24 months ending 30 June 2028, 39% between months 25 and 48, and the rest between months 49 and 78 (10-Q). Our arithmetic: 41% of 103.7 is about US$42.5 billion; 39% is about US$40.4 billion; the remaining 20% is about US$20.7 billion. Our check, not a ranking: 678 ÷ 103.7 ≈ 6.5×. Microsoft commercial RPO is not CoreWeave GPU hours.

Take-or-pay is why a neocloud can print 98% committed revenue while the hall lags. CoreWeave’s Q2 2026 revenue was US$2.575 billion; committed contracts were 98% of the quarter. Active power was 1.5 GW against contracted power of about 3.7 GW (11 Aug 2026). Our arithmetic: 3.7 ÷ 1.5 = 2.5×. The offtake clause does not bill a dark hall. See speed-to-power.

Who gets paid — and where is the stuck step?

Step RPO path Take-or-pay path
Booking Remainder goes up. Cash may or may not. Signature relocates unused-hour risk — after delivery
Stuck step Delivery — energised MW, installed GPUs, accepted capacity The same: the bill does not start until the contracted block is live
Cash-in slice Deferred revenue (prepaid before performance) CoreWeave deferred revenue US$9.7B; 15–25% weighted-average prepay (year-end 2025)
Who gets paid first Operator as it performs; upstream when trays and watts bind Operator as the delivered term runs; lenders against contracted cash flows
Idle risk after start Committed vs usage-driven remainder The customer — until the customer stops paying
Upstream feed CoWoS, HBM, watts Same trays and watts; no finished block, no term to start

Our check on CoreWeave’s cash-in slice: 9.7 ÷ 103.7 ≈ 9% of the RPO figure was already deferred revenue. Our check on the hyperscaler-as-buyer clock: 169.008 ÷ 194.060 = 87% of Microsoft’s take-or-pay-and-orders bucket sat in fiscal 2027. The buyer is on the hook for capacity it has promised to take. The seller is on the hook to deliver it.

Concentration cuts both clocks. Microsoft’s Q4 FY26 call said sequential commercial RPO growth came from customers outside frontier model companies, with RPO still up 25% excluding OpenAI (29 Jul 2026). CoreWeave’s 10-Q names Microsoft and OpenAI as significant customers and records that Jane Street committed approximately US$6.0 billion in April 2026. If a large offtaker does not perform, the operator “may have excess capacity and may remain responsible for expenditures for components, infrastructure, and data center leases and build-outs, as well as related financing.” Take-or-pay moves utilisation risk. It does not delete credit risk.

Oracle’s Q1 FY27 delivery print is the other side of the remainder: 850 MW of additional datacenter capacity and more than 300,000 GPUs since the end of Q4, with 97.9% utilisation (investor release and slides, 10 Sep 2026 — company-reported). Capacity delivered is not RPO. Wire: Oracle 850 MW — fleet still 97.9% full.

The US door is where most public remainders and offtake sit: US speed-to-power when halls wait on watts, 1.8 GW, $1 credit when interconnect is the story, and CoreWeave neocloud contracts when the specialist book is the offtake. China is the other door — China HBM access.

When does each clock convert?

RPO converts to revenue when the operator satisfies the performance obligation (ASC 606). Amazon’s 10-Q says timing is driven by customer usage and AWS’s performance. Alphabet’s backlog excludes cancellable contracts. Oracle recognises consumption infrastructure on utilisation, and subscription cloud ratably once the service is available (10-K, year ended 31 May 2026).

Take-or-pay converts to a bill when the term starts — on a date or when specified capacity is delivered — and to revenue on the same ASC 606 clock. That is why a neocloud can show a large remainder, a smaller deferred-revenue slice, and still live on 98% committed invoices once halls are accepted. Nebius is the heavier cash-in sibling: the Q2 2026 letter said roughly 70% of deals included prepayments covering 50–60% of related capex, with deferred revenue US$6.0 billion at 30 June 2026 (12 Aug 2026). Stack: What Does CoreWeave Do? · What Does Nebius Do?.

A useful test, not a ticker call: if conversion bands hold while delivered megawatts rise, the remainder is turning into hours. If CoreWeave’s committed mix stays near 98% while active power catches contracted power, the offtake is converting. If RPO jumps while halls stall, you are watching the book, not the bill.

Common misconceptions

  • People assume remaining performance obligation is the same as a take-or-pay GPU contract. Actually take-or-pay is the contract form. RPO is the unsatisfied contracted transaction price. CoreWeave prints both: ~US$104 billion revenue backlog and US$103.7 billion RPO (Q2 2026 release).
  • People assume a bigger RPO means the operator is already paid. Actually deferred revenue is the prepaid slice — about 9% of CoreWeave’s RPO. Oracle’s year-end bands put only about 12% of US$638 billion in the next twelve months.
  • People assume take-or-pay means the operator is already paid. Actually the 10-Q starts the committed term on a date or when capacity is delivered. A signature against a dark hall does not yet bill.
  • People assume take-or-pay is a neocloud-only trick. Actually Microsoft’s 10-K puts take-or-pay inside datacenter purchase commitments (US$194.060 billion as of 30 Jun 2026).
  • People assume hyperscaler RPO figures are a league table of AI demand. Actually the cuts differ: commercial RPO, long-term AWS, and Alphabet’s backlog including shorter contracts.

Not investment advice. Do your own research.

FAQ

What is the difference between RPO and take-or-pay?
RPO is the accounting remainder of contracted cloud not yet recognised as revenue. Take-or-pay is a contract form that bills reserved GPU capacity after delivery, even if utilisation is low. One counts unsatisfied work. The other relocates unused-hour risk once the hall is live.

Is remaining performance obligation the same as a take-or-pay GPU contract?
No. Take-or-pay is the offtake clause. Remaining performance obligation is the unsatisfied contracted transaction price under ASC 606. CoreWeave’s US$103.7 billion RPO as of 30 June 2026 is the accounting remainder of committed work still to perform.

Does a bigger RPO mean the operator is already paid?
No. Oracle’s 10-K (year ended 31 May 2026) expected to recognise about 12% of its US$638 billion RPO over the next twelve months. Deferred revenue is the slice where cash already arrived — about 9% of CoreWeave’s RPO at 30 June 2026.

When does a take-or-pay bill start?
On a fixed date or when the contracted capacity is delivered (CoreWeave 10-Q, period ended 30 June 2026). After that on-switch the reserved fee still bills if the cluster sits idle. Undelivered take-or-pay does not yet bill.

Who gets paid on each clock?
On both clocks the operator gets paid as it performs a delivered term, and upstream packaging and HBM get paid when the next tray cannot ship. The stuck step is delivery — energised megawatts and accepted capacity.

Why do hyperscalers show up on both sides?
They sell committed cloud that fills remaining performance obligation, and they buy take-or-pay capacity. Microsoft’s FY2026 10-K included take-or-pay contracts inside US$194.060 billion of purchase commitments, 87% due in fiscal 2027.

What would change the conversion clock?
A hyperscaler RPO print that restates conversion bands while delivered megawatts stall; a shift from CoreWeave’s 98% committed mix toward on-demand; or packaging and HBM delays that keep the next block from starting.

Markets and stack doors

Layer: hyperscalers. Sibling layer: neoclouds. Markets: CoreWeave neocloud contracts · US speed-to-power · US interconnect credit · China HBM access.

Get the Wire — short notes when backlog or hall clocks move. Stack: What Does Microsoft Do? · What Does Oracle Do? · What Does CoreWeave Do?. Analysis: where the stuck step moves.

How we check this

Figures come from company filings and company releases, each dated below. Press-report capacity is labelled reported. Arithmetic we work out is shown in full.

Last reviewed: 1 Oct 2026 · Next review: after next hyperscaler RPO print or CoreWeave Q3 2026

Spot an error? Tell us.

Our editorial standards →

Sources

More Explained