Capital expenditures tell you what an AI company has already spent on physical assets. Purchase obligations tell you what it has promised to spend. The two are different disclosures in different parts of a filing, and for a company financing an AI buildout, the obligations figure is often the larger and more forward-looking of the two — because it captures multi-year commitments for supply and capacity that have been contracted but not yet paid or delivered.

A purchase obligation, in SEC-filing terms, is an enforceable and legally binding agreement to purchase goods or services that specifies the significant terms — fixed or minimum quantities, fixed, minimum, or variable pricing, and the approximate timing of the transaction. These are disclosed in the commitments and contingencies note within a company's audited financial statements, typically accompanied by a table that lays out the total committed amount and how it falls due over future periods. The commitments note is the authoritative location; the figure is not in the cash flow statement, because the cash has not yet moved.

As of January 26, 2025, we had outstanding inventory purchase and long-term supply and capacity obligations totaling $30.8 billion, an increase from the prior year led by commitments, capacity and components for new product introductions, including our new Blackwell architecture.— NVIDIA Corporation, Form 10-K (fiscal year ended January 26, 2025), source

That single disclosure is larger than the company's annual capital-expenditure cash outflow by an order of magnitude, which is precisely why purchase obligations matter to anyone modeling AI capital intensity. NVIDIA's filing reports purchases related to property and equipment of $3,236 million for the same fiscal year; its outstanding inventory and long-term supply and capacity obligations stood at $30.8 billion. The commitment dwarfs the cash already spent, and it is the commitment — not the historical capex line — that signals the scale of capacity the company has contracted for.

The cloud commitment inside the obligations

The same note in NVIDIA's 10-K breaks out a category that has become central to the AI economy. Beyond the inventory and supply commitments, the filing discloses 'other non-inventory purchase obligations' of $14.3 billion, 'including $10.9 billion of multi-year cloud service agreements.' The filing states these cloud agreements are expected to support its research and development efforts and its DGX Cloud offerings. That breakout is a window into a recurring feature of AI finance: companies that build and sell compute also commit to buying compute capacity from others, and those commitments are contractual obligations disclosed here rather than in the capex line.

This is where purchase obligations connect to the broader question of how AI infrastructure is financed. A company can secure future capacity in several forms — outright purchases (which become capex when paid), finance leases, and purchase commitments — and only the first of those shows up cleanly as a capital expenditure. The purchase-obligations disclosure captures the committed-but-unpaid layer, which is why reading 'how much has this company committed to AI capacity' requires the commitments note, not just the cash flow statement.

How to read the obligations figure correctly

Three cautions keep the reading honest. First, a purchase obligation is a commitment, not a completed expense; it tells you what the company is on the hook for, not what it has spent. Second, the disclosure often notes that some agreements are cancellable, reschedulable, or adjustable — NVIDIA's filing states that certain contract-manufacturer agreements 'are cancellable, able to be rescheduled, or adjustable for our business needs prior to placing firm orders,' so the headline total is not always a hard, non-cancellable floor. Third, the obligations table allocates the total across time periods, so the question 'how much is due in the next year versus later' is answerable from the table itself rather than from the single aggregate.

It is worth distinguishing purchase obligations from the related disclosures they sit beside, because each captures a different slice of forward commitment. Operating and finance leases cover the right to use an asset — datacenter space, for instance — over time, and are reported under lease accounting on the balance sheet and in their own note. Construction commitments cover assets a company is having built. Purchase obligations cover contracts to buy goods and services outright. A company financing AI capacity may use all three at once, which is why some filers present a single contractual-obligations table that lines them up side by side with their due-by-period breakdown. Reading any one category in isolation understates the total forward commitment; reading them together, from the notes that define each, gives the fuller picture of what a company is contractually on the hook for.

For the analyst tracking the AI buildout, the method is to read the commitments and contingencies note alongside the cash flow statement: the cash flow line gives historical capex, the obligations note gives committed future spend, and the cloud-services breakout reveals how much of that future commitment is for compute capacity the company is buying rather than building. The $30.8 billion in supply obligations and the $10.9 billion in cloud agreements are not press-release figures — they are line items in an audited note, which makes them among the more disciplined numbers available for sizing the real forward commitment behind an AI company's growth. Where a capex headline captures only the cash already spent, the purchase-obligations disclosure captures the promise — and in the AI buildout, the promise is often the larger and more telling number.