What’s the binding constraint in the $2.47 trillion backlog?
6 of the cloud and AI infrastructure sellers in PitchBook’s latest analysis are sitting on $2.47 trillion of contracted backlog.¹ PitchBook’s data shows backlogs have more than tripled since early 2025.
Every one of those dollars has been won. None of them has been earned yet.
AI demand is real, it’s accelerating, and the firms with the biggest backlogs are winning. Microsoft leads at $678 billion. Oracle follows at $638 billion. Whether anyone actually wants this compute is no longer a question.
But a backlog isn’t revenue. It’s a queue.
A contract becomes revenue only when the capacity behind it is switched on. And capacity is gated by 3 things: chips, power and land. Of the 3, chips are the one a cloud company can make for itself, and the one with the richest margin attached. Nvidia earned a 75% gross margin last quarter on $96.2 billion of revenue.² That margin is what scarcity looks like on an income statement.
So the question that matters isn’t how big the backlog is. It’s who controls the supply that clears it.
Sort the same $2.47 trillion that way and the picture changes.
Only 41% sits with firms that make their own chips at scale. Amazon’s chip business runs above $25 billion a year, with multi-gigawatt Trainium commitments from Anthropic and OpenAI.³ Google began recognizing revenue from TPU system sales in Q2.⁴ Another 27% sits with Microsoft, which deploys its own Maia 200 chip but hasn’t disclosed how much of its fleet runs on it.⁵ The remaining 32% sits with Oracle, CoreWeave and Nebius, which have no in-house accelerator and buy from merchant chipmakers such as Nvidia and AMD.⁶
Here’s the part the market underprices. Owning a chip doesn’t free you from Nvidia. Amazon says some of its customers will keep running on Nvidia for as long as it can see ahead.³ Google is renting third-party capacity as a bridge because supply is tight.⁴ What a chip program buys is a second supply line. In a long queue, the firm with 2 lines clears its orders faster, and keeps more of each dollar when it does.
That gives investors a better number to watch than backlog size. Call it Backlog Duration: backlog divided by annualized revenue, or roughly the years it would take to clear the queue at today’s speed. Google Cloud carries about 5.2 years.⁴ AWS carries about 2.9 years.³ A falling duration means a firm is converting. A rising one means it’s selling faster than it can build.
For anyone underwriting a neocloud or an AI infrastructure company in the private markets, this flips the diligence order. Ask for the chip allocation and the delivered megawatts before you ask for the backlog.
The model travels well beyond AI. In any market where one input is scarce, contracted demand is worth only what that input lets you deliver, and the value pools with whoever controls it.
Backlog tells you who won the customer. Silicon tells you who gets paid first.
Which leaves the question I am unable to answer yet. Amazon says some customers now want Trainium chips delivered outside its cloud, and it is exploring that.³ If it starts selling them, its rivals’ backlogs begin to depend on Amazon’s chip queue. At what point does the hyperscaler become the supplier everyone else is waiting on?
Footnotes
PitchBook, company filings and earnings calls, as of Sept. 24, 2026. Microsoft and Oracle figures are remaining performance obligations. Q2 2026 is the latest quarter reported by all 6 companies. PitchBook’s chart also includes SpaceX, whose $47.5 billion backlog covers launches, government work and Starlink rather than compute, so it sits outside the split.
Amazon Q2 2026 results, SEC Form 8-K, and Amazon Q2 2026 earnings call transcript. Backlog Duration uses $496 billion ÷ ($42.2 billion × 4).
Alphabet Q2 2026 earnings call transcript. Backlog Duration uses $514 billion ÷ ($24.8 billion × 4). Google Cloud revenue includes Workspace.
Comments