AI infrastructure backlog and RPO tracker

Everyone tracks how much the hyperscalers are spending. What those companies have already sold gets far less attention. This page collects every disclosed backlog and remaining performance obligation figure across the AI buildout, from the cloud platforms down to the people making switchgear, and says what each number counts and where the company disclosed it. Figures move only when a company reports, so this updates each earnings season rather than live.

Updated Sep 25, 2026 · figures move only when a company reports

The gap between a figure in an audited filing and a figure on a slide can be large. Core Scientific states more than $24B of potential contracted revenue while its own quarterly report shows about $9.2B of contracted lease payments. Both are on this page, labelled differently.

Cloud platforms and AI specialists, ranked by what customers have contractually committed to pay them. This is the closest thing to a direct measure of booked AI demand.

Data center landlords, kept separate because their metric runs on a different clock. A REIT backlog is stated as annualized rent, meaning revenue per year once the lease switches on, while almost every other figure on this page is a lifetime-of-contract total. The two must never be ranked against each other.

Turbines, switchgear, grid equipment and thermal management. Read these with care: they are company-wide backlogs in which AI is one customer among many, and none of these firms breaks out a data-center figure.

Engineering and construction contractors. This group needs the most care of any on the page, because most of them publish two different backlog numbers and the headline one is usually the larger, looser figure.

Servers, optics and networking. By far the thinnest disclosure on the page: most of these companies publish no backlog at all, and among those that do, the headline figure has usually never appeared in a filing.

Semiconductor capital equipment sits one step upstream of every AI chip. The three numbers here are measured nothing alike: ASML discloses backlog once a year and stopped publishing quarterly bookings in 2026, KLA's figure was spoken on a call and folds in long service contracts, and Applied Materials excludes anything shorter than a year, which keeps its figure structurally tiny. Rank inside this group with extreme care.

Four of the most heavily discussed names in the AI trade publish no backlog figure at all, and one has quietly stopped. Meta discloses nothing in either its quarterly or annual report, which is a policy rather than an oversight, since its advertising revenue is short cycle. Super Micro says it holds record backlog without giving a number, and the $60B widely attached to it is a full year of new orders rather than work outstanding. Bloom Energy reports no order backlog, so the roughly $494M sitting in its deferred revenue note should not be dressed up as one. Vertiv is the one that changed: it disclosed $15.0B for the end of 2025 and has published no backlog figure in any quarterly filing since, meaning coverage that ties that number to its summer 2026 results is recycling a December figure.

The optics and networking names are the quietest group of all. Coherent, Lumentum and Corning publish no backlog whatsoever, despite demand that has roughly doubled revenue at some of them, and Lumentum's annual report contains a section headed Backlog whose entire purpose is to explain why it refuses to give one. Coherent's filings do contain the word, but only as a fully written-off intangible asset left over from a merger, which is the sort of false positive that ends up in a headline. Arista publishes no backlog either, though its remaining performance obligations are real enough.

Dollar totals are not comparable across rows and are not meant to be ranked against each other outside their group. Amazon counts only contracts with original terms beyond a year. IBM excludes anything a customer can walk away from without penalty. Alphabet widened its own definition in early 2026, which is why its year-over-year column is blank rather than showing an impressive and meaningless number. Oracle's total is company-wide and includes $75B of prepaid or customer-supplied hardware. Among the smaller names, several headline figures assume every renewal option gets exercised, so Applied Digital appears here at its $36.2B base-term value rather than the $86B that assumes full renewal, and Galaxy at $30B including extensions against roughly $18B on the initial term alone.

The single easiest mistake to make with this data is mixing up a yearly rate with a lifetime total. Digital Realty's backlog is annualized rent, what the signed leases will pay each year once they switch on, so it is marked accordingly and kept in its own group. Setting $1.9B a year beside CoreWeave's $104.2B of lifetime contract value would understate the first by roughly the length of a lease. The same care applies to Nebius, whose $3.0B run-rate revenue is a yearly pace while its $37.5B obligation is a lifetime figure.

The construction group carries a second figure wherever the company publishes one, because the headline and the accounts often disagree sharply. Quanta's record $53.4B backlog sits against $33.6B of remaining performance obligations, and Primoris reports $13.9B against $6.0B. The difference is mostly estimated work under master service agreements, which customers are under no obligation to actually order. EMCOR is the exception that proves the point: it publishes only the GAAP figure, which is why its total looks modest next to peers running comparable businesses.

Megawatt figures need the same care. Cipher reports gross HPC capacity while Applied Digital, TeraWulf and Galaxy report critical IT load, which is a different denominator, so the power column is there for context within a row and not for ranking across them.

IREN shows why the basis matters more than the total. It books remaining performance obligations only for tranches already delivered and accepted, so its RPO read $710M while the roughly $9.7B Microsoft agreement contributed nothing. The first Horizon deployment was accepted on August 13, 2026 and the FY2026 10-K duly moved RPO to $5.1B. Euro figures converted at the ECB euro reference rate of 1.1534 USD/EUR on 2026-08-13. Yen figures converted at 159.3 JPY/USD on 2026-08-15.

This tracker collects every backlog or remaining performance obligation figure the companies in the AI buildout have disclosed, and refuses to add them together. The reason is that the numbers are not the same kind of number. Each row is labelled with the basis it was reported on, the document it came from and the date the company stated it, so a GAAP obligation is never stacked on top of a self-defined pipeline figure.

• In the financial statements. Disclosed inside a 10-Q or 10-K, subject to audit and to a defined accounting standard.
• Non-GAAP backlog in a filing. Published in the filing but calculated on the company's own definition, typically including estimated work under master service agreements that carries no customer purchase commitment. Usually much larger than the same company's GAAP figure.
• Company release or deck. Published by the company in a press release or investor presentation, but absent from the financial statements. No accounting standard governs how it is calculated.
• Earnings call only. Said aloud by management on an earnings call and never written into a filing or a release.
• Not disclosed. The company publishes no comparable figure. Listed so the gap is visible rather than filled in with an estimate.

Figures move only when a company reports, so this page updates through earnings season rather than live, and the stamp under the headline is the newest data on the board rather than the moment you loaded it. Nothing is estimated or modelled: a number appears here only once the company has published it, and every row links to where. There is no login and no paywall.

For the valuation side of the same question, our AI Bubble Index dashboard scores how much the market is already pricing in against what these backlogs actually commit anyone to.

• Dell raised its full-year outlook by $25 billion as AI server backlog nearly doubled to $95 billion
• Broadcom targets $230 billion in AI revenue by 2028, and the call turned an after-hours drop into a gain
• IREN signed a frontier AI lab and is paying 9% to finance the GPUs
• Nvidia-backed Nscale pitches $103 billion in contracted revenue ahead of its IPO after a $33 million year
• Anthropic's $35 billion Lambda deal runs through Hut 8's Texas campus, WSJ reports
• Riot postponed its earnings call. Six days later it announced a $9.1 billion AI lease
• GE Vernova is booking turbines years out. Data centers are buying Bloom Energy fuel cells instead
• AI's storage ladder: HBM, flash and Seagate's race to the 100TB hard drive
• Quanta raised guidance and rose 17.3%, while Sterling reported nothing and rose 17.4%
• Super Micro's revenue landed at the low end of guidance and the stock rose 25% anyway
• Nebius reported Q2 into a Michael Burry short position

Backlog is one input. To see who is actually buying these names, the smart money confluence scanner lines up insider buying, hedge fund 13F positions and Congressional trades on one ticker, and the hedge fund tracker shows which of the funds we follow hold the neoclouds and the miners. The AI data center sector hub and the HBM memory hubcollect our coverage of the landlords, the power names and the memory suppliers whose books appear above, and each ticker link in the table opens that company's stock page with its filings, insider activity and stock score, which reads the audited rows on this page directly.

Every figure here is taken from the company's own filing, release or presentation, linked in the final column, and nothing is estimated or interpolated. Where a company publishes two versions of the same metric both are noted. This is a record of what has been disclosed, not a forecast, and none of it is investment advice.