Cloud Backlog Before Q3 Earnings: Microsoft RPO, Google Cloud and AWS on a Page

Microsoft, Google Cloud and AWS ended June 2026 holding $1.69 trillion of contracted backlog between them, up from about $750 billion nine months earlier. That cloud backlog is the number every analyst will quote on results day. It is also three different numbers, measured three different ways, and a large slice of each traces to OpenAI or Anthropic.

The short answer. At 30 June 2026, Microsoft reported $678 billion of commercial RPO, Google Cloud $513.9 billion of backlog and Amazon about $496 billion, mostly AWS. Microsoft expects about 30% of its figure within 12 months. Alphabet expects just over half within 24 months. Amazon gives only a 6.4-year average contract life.

This post is written for a founder or director who buys cloud capacity on multi-year terms, or who sells into a market priced off these numbers. The output is a reading method you can reuse on 28 and 29 October, not a view on any share price.

Cloud backlog means three different things at three companies

Remaining performance obligations, or RPO, is a disclosure required under ASC 606, the US revenue recognition standard. It is the value of signed contracts for which the vendor has not yet delivered the service. It includes deferred revenue already billed and contracted amounts not yet billed.

All three companies report a version of it. None reports the same version. The differences are not footnote trivia. They change what a headline growth rate tells you, and they decide which comparisons are fair.

Sources: Microsoft FY26 earnings releases and calls; Alphabet 10-K for 2025 and 10-Qs for Q1 and Q2 2026; Amazon 10-Qs for Q3 2025, Q1 and Q2 2026. Amazon states its figure as "approximately" and rounds to the nearest billion.

Microsoft counts more than cloud

Microsoft's figure covers its whole commercial business. Microsoft 365 seats, Dynamics and other commercial contracts sit in it alongside Azure. Calling it a cloud backlog is a convenience, and it overstates the Azure portion by an amount Microsoft does not publish.

Amazon counts less than the others

Amazon's number covers only contracts with original terms over one year. Its filing also warns that recognition "will be driven by customer usage", which can run past the original contract term. Much of AWS revenue is pay-as-you-go and never enters the backlog at all.

Four quarters of cloud backlog, side by side

Here is the series as each company reported it, from the September 2025 quarter through June 2026. Microsoft's fiscal year ends in June, so its labels run one year ahead of the calendar.

Alphabet's Q1 2026 total includes about $7.3 billion from short contracts that earlier quarters excluded. Amazon's December 2025 figure comes from the earnings call rather than the 10-K, and its contract life for that quarter is not shown here.

Read the shape before the level. Each company's backlog grew in one or two jumps, not in a smooth climb. Steady enterprise demand does not produce a chart like this. A handful of very large contracts does.

Microsoft RPO is the only backlog that subtracts its biggest customer

Microsoft's commercial RPO jumped $233 billion in the December 2025 quarter. That followed OpenAI's restructuring in late October 2025, when OpenAI agreed to buy an incremental $250 billion of Azure services, as Axios reported at the time. Microsoft said about 45% of the $625 billion was attributable to OpenAI.

Since then Microsoft has kept reporting growth with OpenAI removed. That is unusual and it is useful. In the March 2026 quarter, CFO Amy Hood said RPO grew 26% excluding OpenAI, against 99% including it. Commercial bookings rose 7% without OpenAI and fell 46% with it, because the prior year's comparison held the giant OpenAI commitment.

The floor on OpenAI's share at June 2026

Microsoft did not repeat the 45% figure in July. It said backlog grew 25% year on year excluding OpenAI, according to the Motley Fool's account of the call. That is enough to set a lower bound.

A year earlier, commercial RPO was about $368 billion. Even if none of that was OpenAI, the non-OpenAI portion could be at most $368 billion. Growing 25% takes it to at most $460 billion. The remaining $218 billion or more of the $678 billion must be OpenAI.

My reading is that this floor is the most underused number in the Microsoft story. Coverage that says "almost a third" treats 32% as the answer. It is the minimum, and it assumes OpenAI held nothing at Microsoft in June 2025, which the record suggests was not the case.

Shorter duration, more due this year

The timing is improving. At March 2026, Microsoft expected about 25% of RPO within 12 months, with a weighted average duration of about 2.5 years. At June 2026, it was about 30% and 2.3 years, roughly $200 billion of revenue due within the year. Hood also said sequential growth came from "customers outside of frontier model companies", so the $51 billion added that quarter was not another lab deal.

Google Cloud backlog doubled in one quarter, with a definition change inside it

Alphabet's backlog went from $242.8 billion at December 2025 to $467.6 billion at March 2026. The same 10-Q changed the definition. Alphabet began counting contracts with an original term of one year or less, which added about $7.3 billion. Strip that out and the organic jump was roughly $217 billion in three months.

The definition change is small against that jump. It still matters on results day, because any year-on-year growth rate for Google Cloud backlog now compares two differently built numbers until mid-2027.

What the 24-month share implies

Alphabet expects just over 50% of its $519.5 billion to become revenue within 24 months. That is about $260 billion, or roughly $130 billion a year. Google Cloud revenue in the June quarter was $24.8 billion, up 82%, an annualised pace near $99 billion.

So contracted revenue due over the next two years already runs about 30% above Google Cloud's current revenue pace, before any new contracts or usage. The caveat is in the word "expect". Recognition depends on Google building and energising the capacity. Alphabet's capex and cash flow strain is covered in the quarter-by-quarter capex to operating cash flow ratio.

AWS backlog is growing, and its contract life is getting longer

Amazon's backlog rose from about $200 billion at September 2025 to about $496 billion at June 2026. The more telling disclosure is the weighted-average remaining life of those contracts. It went from 3.8 years to 5.5 years to 6.4 years in three filings.

Longer contracts are not a bad sign in themselves. They do mean the headline growth is being bought partly with time. A $496 billion backlog spread over 6.4 years is about $78 billion a year if it were recognised evenly, which it will not be.

The backlog is a minority of AWS revenue

AWS reported $42.2 billion of sales in the June quarter, up 37%, a $169 billion annualised run rate. The even-spread figure of $78 billion is less than half of that. Most AWS revenue still arrives from usage that never sat in the backlog.

That cuts both ways. AWS's backlog says less about next year's revenue than Microsoft's does. It also means a single long contract moves Amazon's backlog more than it moves Amazon's revenue, which is exactly what the last two quarters show.

Which deals sit behind the jump

Amazon's own filings name them. The Q1 2026 10-Q records that AWS and OpenAI expanded their "existing $38.0 billion multi-year commitment" by $100 billion over 8 years. The Q2 2026 10-Q records that AWS and Anthropic expanded their commitment by "more than $100.0 billion over 10.0 years". Neither filing says how much of either sits inside the $496 billion.

How much of each cloud backlog rests on one AI lab

This is the question the late-October prints should be read against. The answer comes with three different levels of evidence, and those levels should not be blended into one tidy chart.

Amazon: a ceiling, not a share

OpenAI's $138 billion and Anthropic's $100 billion or more total at least $238 billion. Against $496 billion that is about 48%. It is a ceiling for two reasons. Some of the older OpenAI commitment may already have been consumed, and Amazon has not confirmed that the April Anthropic expansion was booked in full by June.

Google: one reported contract

Anthropic announced on 6 April 2026 that it was expanding its agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, mostly from 2027. No dollar value was published. The Information reported a $200 billion, five-year Anthropic commitment to Google Cloud, as relayed by the Motley Fool in May 2026. Against $513.9 billion that would be about 39%. Alphabet's filings say nothing about it.

Microsoft's backlog also carries Anthropic. In November 2025 Anthropic committed to buy $30 billion of Azure compute. So the frontier-lab share at Microsoft is higher than the OpenAI figure alone. Who pays whom across these arrangements is mapped in the breakdown of the AI circular deals.

My position: the lab share matters less than the lab's ability to pay. A backlog is only as good as the counterparty's future cash. OpenAI's own numbers are examined in the OpenAI valuation and run-rate multiple analysis, and Anthropic's in the reconciliation of Anthropic's S-1 losses.

Where this comparison is weakest

Three honest problems, in order of how much they should change your reading.

First, a backlog is a contract number, not a cash number. None of the three companies publishes cancellation terms or what happens if a customer cannot pay. Oracle shows how far that gap can stretch. Its own backlog is read through a credit lens in the Oracle RPO credit analysis.

Second, the strongest bear point, that one customer dominates, is weakest where it matters most. Only Microsoft gives anything close to a customer split. The Amazon and Google shares in this post are a ceiling and a press report. They could be materially lower.

Third, the case against concentration as a worry is real. The labs are the fastest-growing software buyers on record, and the hyperscalers hold equity in them, which aligns incentives. Supply is also the binding constraint: if a lab walks away, the capacity has other buyers today. I think that argument is right for 2026 and untested for 2028, when much of this backlog actually comes due.

Smaller caveats still apply. Microsoft's ex-OpenAI growth rates in this post rest on call coverage rather than a filed document. Amazon's December 2025 figure comes from the earnings call. Amazon rounds every figure to "approximately".

The cloud backlog checklist for results day

Microsoft reports on 28 October 2026, after the close, per its 7 October announcement. Amazon reports on 29 October, per its 15 October release. Alphabet is widely listed for the same week; confirm the date on its investor page.

One discipline for the day itself. Compute the change in backlog in dollars before you look at the percentage. A $50 billion quarter at Microsoft and a $50 billion quarter at Alphabet are the same money and very different growth rates.

Frequently asked questions

Where to start

Before 28 October, copy the second table in this post into a sheet and add a fifth row for the September quarter. Fill each cell the evening each company reports, in dollars first, then percentages.

If you hold a multi-year cloud commitment, check one thing in your own contract this week: what happens to your committed spend if your provider cannot deliver capacity on schedule. The hyperscalers' backlogs depend on that same clause, written in their favour.

Related reading