The AI bill arrives

Two stories from the last few months have stayed with me.

Microsoft gave thousands of engineers access to Claude Code in December. Six months later it’s winding those licenses down and moving everyone to its own Copilot, because individual engineers were running up $500 to $2,000 a month in usage, and the annual budget was gone months ahead of schedule. Uber’s CTO said the same thing more bluntly: their entire 2026 AI budget, spent by April.

Here’s the detail that makes this interesting rather than just embarrassing: the budgets didn’t collapse because the tools failed. They collapsed because the tools worked. At Uber, adoption hit around 85% of engineers and roughly 70% of committed code now starts with AI. People used the thing constantly because it was genuinely useful, and constant use is exactly what usage-based pricing punishes.

Quick explanation for anyone outside this world: most AI tools charge by the token, which is roughly a fragment of a word, so you pay for every word in and every word out. The cost doesn’t scale with how many people have access. It scales with how much they actually use it. A flat-rate software budget has no way to model that, which is why so many companies got surprised by their own success.

And the supply side isn’t coming to the rescue soon. A third to a half of the US data centers planned for this year are expected to be delayed or cancelled, mostly for unglamorous reasons like transformers and switchgear being out of stock.

I work in fintech, where the first question about any product is its unit economics, what one transaction costs, what it earns. What strikes me about this moment is that AI just handed every team lead that same question, and most of us never had to ask it about our tools before. Nobody asked what a Figma seat cost per file. Now “what did that workflow cost” is a real question with a real number attached.

So here’s what I’m taking from it as someone leading a design team through AI adoption, not just watching from the side:

• Plan for the high end, not the average:
  The bill comes from your best-case scenario, everyone using it, daily. Model that before rollout, not after.
• Put AI where it earns its cost:
  This changed how I think about my own workflow. Research synthesis across hundreds of interviews? Worth every token. Generating a fifth round of options nobody asked for? That’s the design version of leaving the tap running.
• Watch the workflow, not the seat count:
  The expensive habit isn’t access, it’s long, meandering sessions. Tight, well-scoped prompts aren’t just better craft anymore. They’re cheaper.
• Don’t read this as “AI was a mistake”
  Microsoft isn’t leaving AI; it’s moving to the tool it controls the cost of. The correction is about pricing models, not about whether the tools work.

For design teams specifically, I think the pressure lands somewhere unexpected. The parts of design work AI does cheaply, variations, resizing, first drafts, exploration, were never where a designer’s value sat. They were the volume work around the value. What a business actually pays a designer for is judgement: knowing which of the twenty options is right for this user, this market, this moment. And here’s the thing about judgement, it doesn’t consume tokens. When every generated option has a visible price on it, the person who can look at a problem and say “we don’t need twenty directions, we need these two” stops being a nice-to-have. They become the cost control.

That’s also why I don’t buy the replacement story, and I say this as someone using these tools daily, not defending my turf from a distance. AI can produce screens. It cannot sit in a room with a distributor in a tier-3 town and notice what he doesn’t say about why he won’t use an app for payments. It can’t weigh a trade-off it wasn’t told about, and most of the trade-offs that matter never make it into a prompt. The designers who should worry are the ones whose entire contribution was production volume, and honestly, that worry predates AI. For everyone else, the economics just started arguing on your side: when output is nearly free, output stops being the differentiator. Understanding people, and being trusted to decide, is what’s left. That was always the actual job.

The uncomfortable version of the takeaway: for a couple of years, AI adoption was a faith decision. Leadership said use it, everyone used it, nobody watched the meter. That era just ended. The teams that keep their AI budgets will be the ones who can say exactly what they’re getting for them. That’s not a finance skill. That’s knowing your own workflow well enough to defend it, which was always the job.