The figure came with the rest of the disclosure: $13.8 billion of contracted work that Anthropic had not yet performed as of September 30. About three-quarters of it is expected to be recognized as revenue within the next twelve months, the company said on November 3, with the remainder arriving after that.
The $13.8 billion is the company’s stock of unfulfilled contracts, and its composition matters more than its size. Anthropic said the measure consists mainly of deferred revenue, money customers have already paid or been billed for services not yet delivered, and of non-cancelable commitments written into contracts that have not started.
That definition is narrow on purpose. Remaining performance obligations, as the accounting standard calls them, count only the portion of a signed contract the company is obliged to deliver and the customer cannot walk away from without penalty. Trials, month-to-month usage and anything a customer can cancel on notice are excluded, which is why the number is usually smaller than the sales pipeline a company talks about on an earnings call.
The disclosure also arrives at an awkward moment for reading it. Anthropic filed a registration statement on October 30 for a public listing, a document that put numbers in front of investors that private backers had previously seen only in a data room. The same filing set out commitments to spend close to $900 billion on servers, buildings and electricity, and a net loss of $80.7 billion for the first half of the year.
Read together, the two sides of the ledger describe a company that has promised to buy a great deal of capacity and has sold a smaller, though growing, amount of it. Contracted revenue of $13.8 billion against contracted costs of a different order is the arithmetic public investors will spend the roadshow testing.
The twelve-month share is the detail that carries the most information. A backlog weighted toward the near term behaves like deferred revenue: it converts quickly, it is already priced, and it tells an investor little about demand two years out. A backlog weighted toward later years does the opposite, promising revenue that depends on the customer still wanting the service when the contract matures. Anthropic’s split, with roughly three-quarters landing inside a year, puts most of the weight on the first case.
For a company selling model access, that shape is expected. Enterprise buyers of AI services have been signing annual agreements rather than the multiyear deals common in traditional software, partly because the technology changes faster than a procurement cycle, and partly because nobody wants to commit three years of budget to a capability that may be repriced. Consumption-based contracts, in which the customer commits to a minimum and pays for what it uses above that floor, produce exactly this kind of backlog: a large number, spent quickly.
Software companies have used remaining performance obligations as evidence of durability since the standard requiring disclosure took effect. Investors learned to discount it after watching vendors pad the figure with long-dated renewals and then restate expectations when those renewals slipped. The useful comparisons are the rate at which the backlog converts, the share of it that is cancelable, and whether the company has the capacity to deliver what it has sold.
Capacity is the constraint that matters most here. Delivering contracted inference requires chips, data center space and power, all of which are bought years ahead and paid for on delivery. A company that has sold more work than it can serve does not lose the contract; it loses the timing, and the revenue arrives in a later quarter than the one the customer expected.
The composition of Anthropic’s customer base explains part of the concentration. The company sells through an API to developers and startups, through direct contracts to large enterprises, and through coding tools that individual engineers adopt before their employers buy anything. The last route produces usage that starts small and grows without a procurement process, which shows up as revenue rather than as backlog until someone signs a contract.
Competition sets the other boundary. OpenAI sells to many of the same buyers, Google offers models through its cloud, and every large enterprise now runs evaluations of at least two vendors. A buyer who can switch cheaply has little reason to sign a multiyear commitment, which keeps contract terms short across the industry and keeps the backlog figure modest relative to annual revenue.
What the number does not say is how profitable the work is. Revenue recognized against a contract signed when compute was cheaper carries a different margin from revenue signed last quarter, and the registration statement’s cost commitments suggest the company expects to spend heavily to serve what it has sold. Gross margin, disclosed separately, will be the figure that tells investors whether the backlog is worth having.
Anthropic has not said how much of the $13.8 billion comes from deferred revenue and how much from commitments not yet billed, a split that would show how much cash the contracts have already produced. Nor has it broken out the largest customers, another detail that matters when a handful of accounts can account for a large share of a backlog.
Those gaps will be filled during the listing process, under questioning from bankers and then from public shareholders. Until then the figure stands as a statement about delivery: $13.8 billion of work the company has promised to perform, most of it within a year, against a capacity plan that has to arrive on schedule for the promise to be kept.
