Your AI vendor's pricing has a balance sheet under it

What the money is doing

Start with what it buys, because this is not a doom piece. Andy Jassy told investors in July that Amazon will spend US$220 billion this year, up US$20 billion on the February guide because memory got dearer, and still will not have enough capacity to meet the demand it can see in 2026, or in 2027. Alphabet cut the cost of serving Gemini by 78% across 2025 through model optimisation, and that saving is what lets a vendor drop the price of a token while its capex climbs. That is the deal you have been enjoying: each quarter the same task costs less and the model behind it is better.

The scale is what changed in 2026. Five companies at US$660-690 billion of capex, the sell-side estimate at US$697 billion by June, and Goldman counting more than a trillion dollars once the rest of the world is added. J.P. Morgan Asset Management's June number is 93% of operating cash flow, and the balance sheets are filling that gap with bonds. Morgan Stanley had nearly US$236 billion of AI-linked debt issued by the end of May, four times the same stretch of 2025, and expects close to US$570 billion for the year. The buyers are starting to push back. Fortune reported hyperscaler bond cover ratios falling from nearly five times in February to under two by July.

Why a CFO in Auckland should care

You are not lending these companies money, so it is tempting to file this under "American investors' problem". The reason it lands on your desk is simpler. A company spending 93% of its operating cash on infrastructure has to make that infrastructure pay, and you are where it gets paid. Every lever it has to do that is a lever on your contract: the per-seat price, the usage tier, which models are on the enterprise plan and which are retired, whether the in-region option survives the next re-organisation, and how the bundle gets carved up at renewal.

None of that requires anything to go wrong. It is what a rational business does when it has borrowed to build and its lenders start asking for wider spreads. Reading the balance sheet will not tell you the move. It tells you a move is likely, so you write down, before it comes, what happens to your data, your workflows and your cost when it does.

Kevin made the same argument about model churn last month, from the other side of the ledger. A frontier model lands every fortnight, and the vendor selling it to you is carrying the debt that built it. Both point at the same design choice: own the layer around the model, and treat the model and the vendor as swappable parts.

Portability is a commercial term, not a technical one

The engineering half of this is well understood: a thin interface between your workflows and the model, the model choice in config, your prompts and evaluation set held as assets you own. The commercial half gets skipped, because it lives in a contract nobody re-reads until something breaks. For every AI vendor you renew with, four questions decide how much of the balance-sheet risk is yours.

What comes back, in what format, and how fast? "You can export your data" is not an answer. Ask for the formats (a database export, JSON, the embeddings if they were derived from your documents), the timeframe after termination, and whether a fee attaches. Ask what happens to your workflow definitions, prompts and fine-tunes, because that is where the value has accumulated.

What can change mid-term, and what notice do you get? Pricing, tier boundaries, model availability and region. If the vendor can retire the model your product runs on with thirty days' notice, price that in now. Ask for a right to terminate without penalty when a material term changes.

Where does the data live, and can that be changed on you? For anyone holding health, employment or Māori data in Aotearoa, in-region hosting is an obligation rather than a preference. Get the region written in, and get notice and exit rights if the vendor consolidates it.

What does it cost you to leave? Not the vendor's exit fee. Yours. The engineering days to swap the model, the data migration, the retraining. If you cannot answer this in a number, you have not measured your lock-in, and the vendor has.

This is why RIVER's own terms are built the way they are. You own your data, your workflows and your IP, always. On termination we support export in standard formats, we select models on requirements rather than on any one provider, and source-code escrow is available with the release conditions written down. I would rather hand a client the door key on day one than argue about it on the day they want to leave. That is the standard I think any vendor should meet in a year when the people selling you AI are funding it with bonds.

What to actually do

Read your AI contracts before the renewal date, not at it. Put the four questions to each vendor in writing and keep the answers with the contract. A vendor that will not answer the exit question has answered it.

Know your switching cost as a number. Run the exercise once: what would it take, in days and dollars, to move your workflows to another model or another host? If the number is embarrassing, that is the project to fund this quarter, while the switch is cheap and nobody is forcing it.

Own the layer that compounds. Your data, your workflows, your evaluation set and your governance are the assets. The model and the vendor underneath are parts. Keep the parts swappable and the spending race turns into falling prices for you instead of a risk you carry.