Why Cloud AI Bills Are So Hard to Budget

Most business software has a price you can write into a budget: a licence, a seat count, a contract. AI usually doesn't. It is billed by use, the use keeps changing, and the tools themselves are getting more autonomous. Here is why the bill is hard to predict, what happened at one large company that found out the hard way, and what changes when the price is fixed.

Why the bill moves

• It's metered in tokens. A token is a small chunk of text. How many a task uses depends on the prompt, the model and the job, so the same request can cost very different amounts.
• Agents multiply it. An agent takes many steps for one request, and each step consumes resources and adds to the cost (Finance Derivative).
• Cheaper units, more of them. Per-unit prices have been falling, but usage has grown faster, so total spending can still rise (Fortune).
• Tools pile up. Different teams buy different platforms, and AI features arrive inside software a company already pays for, so costs can grow without anyone deciding to spend more (Finance Derivative).
• The billing model is shifting. Some providers, GitHub among them, have been moving from flat rates to usage-based billing (Built In).

What happened at Uber

Uber is the example most coverage uses. Its CTO said in April that the company had used up its whole 2026 budget for AI coding tools in about four months, and it later capped spending at $1,500 a month per employee for each agentic coding tool, according to Bloomberg's reporting (TechCrunch). Reports say usage grew after the company encouraged staff to adopt the tools and ranked teams by how much they used them (Fortune).

Forbes reported that the share of Uber engineers using Claude Code rose from around a third in February to 84% in March, with typical monthly bills of $150 to $250 per engineer and the heaviest users reaching $500 to $2,000. Uber's COO said it was hard to connect all that token use to more useful features for customers, and observed that engineers who don't pay the invoices may treat the tools as nearly free (Yahoo Finance). Analyst Simon Willison called the cap a rational response, and noted that people spending that much suggests the tools deliver value (SmarterX).

How companies are responding

• Spending caps. Per-person, per-tool monthly limits, as at Uber.
• Visibility. Dashboards that let people see their own usage (Finance Derivative).
• An inventory. Tracking token spend across cloud models, subscriptions, browser plugins, desktop agents and gateways, instead of treating each as a separate expense (Forbes Technology Council).
• Model choice. Not making the most expensive model the default for every task (Finance Derivative).

A Forbes Technology Council piece also cites a Gartner prediction that by 2028 the cost of AI coding tokens alone will pass an average developer's salary. That's a forecast, not a measurement.

What a fixed price changes

The same dynamic exists at small scale: a hobbyist on a pay-as-you-go API key can overspend too. The alternatives each trade one kind of certainty for another.

• Flat subscriptions make the monthly cost predictable but usually come with usage limits.
• Pay-as-you-go matches cost to use and gives the least certainty.
• Running models on your own hardware removes the per-use charge, but moves the cost up front, and the models are smaller than the largest cloud ones. We looked at that route in our overview of local AI hardware.

For disclosure: we make iPhone apps, OffgridStem, OffgridScribe, OffgridVox and OffgridCam.

What to watch

• Whether more providers move to usage-based billing.
• Whether companies find a reliable way to connect AI spending to results.
• How fast agents' share of spending grows. Gartner forecasts AI agent software spending near $207 billion in 2026, up from $86.4 billion in 2025 (Fortune).

What we don't know yet

We have no access to any company's internal figures. The Uber numbers come from press reports that cite company statements and other outlets, and the forecasts are predictions that may change. Pricing in this market is also moving quickly, so the picture may look different in a few months.

More from Offgrid Studio

Local AI Hardware in Autumn 2026: What You Can Actually Run at Home
OpenAI's Dots: Always-On Agents That Work on a Computer of Their Own