For two years the argument about artificial intelligence spending has run in the future tense: capital expenditure guidance, power contracts, chip orders, all of it money leaving the building against revenue that does not exist yet. In the third quarter the other side of the ledger finally arrived with a number attached.
Google Cloud’s revenue grew 103% from a year earlier. Its contracted backlog rose to $561 billion, up 256%. Amazon Web Services grew 39%. Between them the two businesses now carry roughly $1.14 trillion of signed work that has not been recognized as revenue.
That total is the most consequential figure in the sector and the least examined. It is also why two teams inside Goldman Sachs can study the same filings and reach opposite conclusions, according to analysis reported this week. One camp reads the backlog as three years of revenue already locked in. The other reads it as a delivery schedule the power grid cannot honor. The arithmetic is identical. The disagreement is about physics.
A backlog, in plain terms, is contract value a company has signed but not yet billed. Cloud providers book it when a customer commits to a minimum spend over several years, and they recognize the money only as the work is delivered. It sits between an order book and a revenue line and carries the weaknesses of both. Multi-year contracts can include termination clauses, ramp schedules and credits that shrink the headline number once they are exercised.
Here is the comparison that should worry a careful reader. Google Cloud’s backlog grew 256% while its revenue grew 103%. The ratio of contracted future work to current revenue roughly doubled in twelve months. In earlier cloud cycles the two grew closer together, because customers signed commitments as they consumed capacity. When backlog outruns revenue by that margin, one of two things is happening. Either buyers are signing years earlier than they will use the capacity, or the contracts are longer than they used to be, or both.
The bull case writes itself. If you own a share of Alphabet or Amazon today, much of the next three years of cloud revenue has already been agreed. That removes demand risk from the model, which was the risk that mattered most when the market was still debating whether companies would pay for AI at all. It also explains why the paper is being signed early. Land, transformers, turbines and interconnection agreements are scarce, so a customer who needs compute in 2028 has to commit in 2026 to hold a place in the queue. A signature is the price of admission.
The bear case is narrower and harder to answer. Delivery capacity is not a balance sheet item. A data center that is contracted but not energized earns nothing while its depreciation clock runs. Both companies raised capital spending through the year, and the assets behind that spending begin losing value the day they are installed. If the backlog converts over five to seven years rather than three, the revenue total is the same and the return is not.
The comparability problem makes this worse. Microsoft discloses a commercial remaining performance obligation that bundles cloud, Office and other contracts, so it cannot be set against Google Cloud’s number directly. AWS, the larger and older business, grew 39%, a rate that would have been extraordinary in 2019 and now marks the mature end of the range. Google Cloud compounds faster off a smaller base, which flatters every ratio computed from its backlog. Investors comparing the two are not comparing like with like, and most published tables do it anyway.
What the backlog does not tell you is margin. Committed revenue at a fixed price is good news only if the cost of serving it holds. Power prices, chip depreciation and the interest cost of financing construction all sit outside the contract. A three-year commitment priced in 2025 is a different asset in 2028.
There is a second reading of the Goldman split that matters more than which side is right. The two camps are not disagreeing about demand, because demand is no longer in dispute. They are disagreeing about the conversion rate between a signed contract and a billed dollar, and nobody publishes that number. Every model of these two companies now depends on an assumption that is not in any filing.
The judgment, then, is that $1.14 trillion settles the demand question and leaves the supply question untouched. Anyone still arguing that enterprises will not pay for AI has been answered by the signatures. Anyone arguing that the money will arrive on schedule has not been answered at all, and the market is pricing the first claim while ignoring the second.
The test arrives with the fourth-quarter reports in late January. Watch the gap between backlog growth and revenue growth. If backlog again outgrows revenue by more than two to one while capital spending guidance rises for a fourth consecutive quarter, the delivery-constrained reading is correct, and the pressure lands on gross margin before it lands on revenue. If backlog growth slows toward revenue growth while capital spending guidance holds flat, the demand-side reading wins and today’s multiples are defensible. One of the two will be legible in the same table, in the same filing, on the same afternoon.
