If AI is now a line item in your budget, this is for you
Let’s start with the part that might surprise you coming from us: for most businesses, renting AI is the right call.
“Renting” means using an AI model through a vendor (OpenAI, Anthropic, Google, and others) and paying as you go, rather than downloading a model and running it on your own computers. (Our field guide, AI Isn’t One Thing, explains the difference in plain English.) Renting means someone else runs the servers, handles the updates, and carries a lot of the compliance paperwork. We build on rented AI ourselves.
But Mozilla’s new State of Open Source AI report is a useful reminder that renting comes with the usual renter’s risks. The rent can go up in ways you didn’t budget for. It’s harder to move out than you’d think. And the landlord holds the keys.
None of that is a reason to panic, or to buy a rack of servers. It’s a reason to rent like a smart tenant. Here are three things worth knowing.
1. The price list isn’t the bill
AI is priced in “tokens,” small chunks of text, both what you send in and what comes back. Price lists quote a cost per million tokens, and those prices have fallen dramatically: Mozilla reports about 60× cheaper in under four years.
So why can your AI bill still go up?
Partly because the price per token and the price per finished job aren’t the same number. Mozilla gives a clean example: one popular model “lists at $15/M output and bills like $31” on a standard set of test tasks, because of how many tokens it burns getting there. To be fair to that model, Mozilla also found it still came out cheaper per task than two leading rivals. That’s exactly why you have to measure the job and not the sticker.
And partly because usage grows in ways nobody plans for. Mozilla’s report cites the case of Uber, which, per Fortune, burned through its entire 2026 budget for AI coding tools in four months. According to Bloomberg’s reporting, some engineers had been running up $500 to $2,000 a month each, and Uber responded with a $1,500 monthly cap per tool, per employee.
If a company with Uber’s finance team can get surprised, a 30-person business certainly can.
2. The hardest thing to move is the part you can’t see
When you think about switching AI vendors, you probably picture swapping one for another, like changing phone carriers.
Mozilla points out a quieter kind of lock-in: “Thinking got cheap and memory got expensive. Lock-in now runs through state that accumulates on the provider’s side.”
Here’s what that means in practice. Vendors let AI “remember” the big pile of material it works with, like your instructions, your documents, and the conversation so far, so it doesn’t have to re-read everything from scratch. That memory (called a cache) is a big discount; Mozilla says a 90% hit rate cuts costs 80–90%. But it lives on the vendor’s side, and “no portability standard for cached context exists.” Switch vendors and you start paying full price again while everything gets rebuilt.
Add the tools, settings, and custom instructions that only work with one vendor, and “we’ll just switch if prices go up” gets a lot less simple than it sounds.
3. The off switch was never yours
This one isn’t hypothetical, and it isn’t about any vendor doing something wrong.
On June 9, Anthropic released its newest frontier models, Fable 5 and Mythos 5. Three days later, a US government export-control order required it to block access by foreign nationals. Anthropic had no reliable way to check nationality in real time, so it switched both models off for everyone. Fable 5 came back for all users on July 1: nineteen days later.
Anthropic complied with a government order; that’s what a responsible vendor does. But if your business had rebuilt a workflow around that model the week it launched, you spent nineteen days finding out what your backup plan was.
It’s not the first time something like this has happened. Mozilla recalls 2018, when Twitter acquired an anti-abuse service called Smyte and shut off its service “within the hour.” Customers on multi-year contracts got about 30 minutes’ notice.
Mozilla, which openly champions open models, puts the lesson in one line: “You can switch off a model. You cannot switch off a copy already running on a machine you hold.”
That’s true, and it’s why some organizations should run their own. But for most businesses, the practical takeaway isn’t “own everything.” It’s “don’t have a single point of failure you’ve never tested.”
How to rent like a smart tenant
Measure cost per finished job. Not per token. Pick your three most common AI tasks, and find out what one of each actually costs. That’s the number that belongs in your budget, and the only fair way to compare vendors.
Put a cap on it before you need one. Most AI vendors and business tools let you set spending limits. Set them per tool and per team, with an alert well before the ceiling. Uber’s cap came after the budget was gone.
Keep a tested backup model. Pick a second model, from a different vendor or one you can run yourself, and actually run your key workflows on it once. When Hugging Face was dealing with a serious security incident this summer, its published lesson was to have “a capable model you can run on your own hardware, vetted before an incident.” You don’t need to go that far. But “vetted before an incident” applies to everyone.
Build so the model is the swappable part. Keep your instructions, documents, and business rules in your systems, not locked inside one vendor’s settings. When the AI is one plug-in piece of a system you control, switching is a few days of testing, not a rebuild.
Know your exit. Write down, in a paragraph, what would break if your AI vendor disappeared tomorrow. If you can’t write that paragraph, that’s the first thing to fix.
What to do on Monday
Two things, one hour total:
- Find this month’s AI spend across every tool and every card, and check whether there’s a cap on each. If there isn’t, set one.
- Ask whoever built your AI workflows one question: “If this model disappeared for three weeks, what would we do?” If the answer is a plan, great. If it’s a pause, you’ve found your next project.
Watching this over time is ongoing work: prices change monthly, new models launch constantly, and the right vendor this year may not be the right one next year. That’s part of what our Steward clients hand off to us. If you’d just like to talk it through once, book a free half-hour with a senior engineer.