In the data industry, it is quite common to purchase credits for future consumption.
You estimate how many reports or datasets you will need over the next 12 months—or even over a three-year period—and then negotiate a deal with data vendors such as CRIF.
It’s a win-win.
The client secures a better price and reduces friction when the need arises (no need to issue a purchase order every single time).
The vendor, on the other hand, can plan production—especially in markets like Asia where investigations may be required—or at least plan revenues and therefore investments.
The backlog, meaning the total amount of credits purchased by clients but not yet consumed, is therefore an important KPI to monitor.
It should grow if the company wants to sustain future top-line growth. But it should not grow too much compared with actual consumption. If it does, it may indicate that clients’ demand is declining, that they are consuming from competitors instead, or simply that they overestimated their needs.
Bloomberg reported in a February article that the AI cloud backlog has reached $1.6 trillion. This refers to contracted but not yet recognized revenue held by the major cloud hyperscalers. The four largest players—Microsoft, AWS, Google Cloud, and Oracle—have combined contracted business (backlog) not yet recognized as revenue totaling $1.63 trillion. Roughly $550 billion of that comes from commitments tied to OpenAI for AI infrastructure, training, and inference workloads, leaving well over $1 trillion from other enterprise customers.
In 2026, only a portion of this backlog will actually be delivered by cloud providers, despite massive capex investments of around $600 billion in infrastructure.
Backlog can exist for several reasons: companies may not yet have the capacity to fulfill all the orders; customers may have signed future commitments but do not yet need the capacity; or the reality may be some opaque combination of both. So far, nothing unusual.
The real question is whether this backlog—also known as RPO (Remaining Performance Obligations)—is based on excessive optimism.
In some ways it is not so different from the data business. Except that here we are seeing a large number of “circular deals.”
A circular deal is a structure in which money flows through a closed loop: one company finances another, which then uses those funds to purchase products or services from the original investor, effectively creating a guaranteed revenue stream. The cycle can repeat multiple times through equity exchanges, service contracts, or infrastructure commitments.
The most emblematic case is the triangle Microsoft → OpenAI → Microsoft Azure.
Microsoft invested $13 billion in OpenAI. OpenAI then committed roughly $550 billion in cloud contracts with Azure. That figure directly inflates Microsoft’s RPO backlog.
In practice, it would be like investing heavily in the equity of my own customers so that they will have the money to buy a massive amount of data credits from me in the future. In the data industry we would probably say that revenues are being artificially inflated.
One could argue that this is different, because OpenAI is both an enabler and an accelerator of Microsoft’s broader software ecosystem, not just a cloud customer. That is true.
But if the speed of real demand does not eventually catch up with the commitments that have been signed, there could be a painful adjustment—leaving behind a lot of idle infrastructure.
If, on the other hand, demand materializes as expected, then the backlog will continue to grow—perhaps this time without the need for circular deals.
-SL