What Uber's AI budget actually tells you, and why the fix everyone applauded is the wrong one.
Uber handed its engineers agentic coding tools in December, put token consumption on internal leaderboards, and watched the entire 2026 budget disappear by April. The company's president and COO later said he could not connect that spending to a single better experience for a rider or a driver. The CTO has since declared the tokenmaxxing era over.
The industry read that as a cautionary tale about AI cost. It's a story about leadership, and the ending everyone applauded is the part that should worry you.
The fix was a cap. Roughly $1,500 per employee per tool per month. SemiAnalysis talked to more than fifty enterprises and found the same reflex everywhere, with caps landing anywhere from $250 to $2,000 a head and no two industries agreeing on the number. Everyone is confident there should be a limit. Nobody can tell you why theirs is the right one.
That's because the cap is the same mistake as the leaderboard, with the sign flipped.
Tokens Are a Cost, Not a Signal
Tokens have no meaning as a stand alone. They don't associate with quality. They don't associate with value, results, or outcomes. They associate with your bill.
That makes them a direct cost, and direct costs get governed. You watch them, you forecast them, you put controls around them, exactly the way you already do with cloud compute and storage and every other consumption line you've managed for a decade. Nobody has ever run a business by celebrating a rising AWS invoice.
What you don't do is measure by them. A token count with nothing beside it tells you a machine was busy. It tells you nothing about whether anything got better. Govern them, don't measure by them, and never publish the number without a value or quality variable sitting next to it.
I've made this argument before in a different costume. Enterprise AI ROI decks love the five minutes saved times a thousand employees times 250 workdays calculation, which produces a number with commas in it and no corresponding change anywhere in the business. Nobody walked out of work two weeks early. The leaderboard is that same fantasy running the other direction. Instead of counting imaginary minutes saved, we started counting real dollars spent and calling it engagement.
The Real Unit of Measure
Here's where most budget conversations go wrong before they start. They treat "tokens" as the unit. It isn't.
The minimum viable unit is tokens plus model plus reasoning level. Those three variables determine your cost, and a control that only touches the first one gets gamed on the other two immediately. Cap the dollars on someone doing genuinely hard work and they'll drop to a cheaper model to stay under the line, which costs you exactly the thing you were paying for. You didn't save money. You bought worse output at a discount and moved the loss somewhere it doesn't show up on a dashboard.
Now the harder question, which is how you allocate across a team.
The instinct is to sort by seniority. Seniors get the best return per token because they know what to ask for and can smell a wrong answer when it arrives. Juniors get the most raw lift, because the assistance is filling gaps that experience hasn't filled yet. Follow that instinct to its conclusion and you hand the seniors the big budget, throttle the juniors, optimize this quarter beautifully, and quietly defund your own apprenticeship pipeline.
But seniority is the wrong sort key, and I want to be precise about this because it's the part people get backwards.
The model should match the shape of the request, not the level of the person making it. If you know exactly what you want and what you want is complex, a frontier model matches that complexity and earns its price. If you're asking why something works the way it does, a frontier model is a waste. You're paying premium rates for a question a mid-tier model answers just as well, often more legibly.
Seniority only correlates with model choice because senior people show up with complex, well-specified requests more often. The variable that matters is the request. Which means the leverage isn't in the allocation spreadsheet at all. It's in teaching people to read their own question before they pick their tool.
That's a training problem wearing a budget problem's clothes. Almost every one of these is.
Education Is the Conservative Choice
So somebody has to fund the training, and this is where most technology leaders lose the argument.
They lose it because they walk in treating education as the soft line, the one they expect to defend on principle, the nice-to-have they'll trade away to protect something harder. The moment you frame it that way you've told Finance it's optional.
It isn't optional, and more to the point, it isn't soft. Hiring fully formed talent is the expensive path. It costs more money and takes more time, and you pay the full price on day one. Training someone costs less of both, and you collect the benefit while you're still paying the smaller cost. That's not a culture argument or a values argument. That's the fiscally conservative position, and it's the one you should be making out loud.
Don't split it into a separate line item either. The instinct to isolate education spend so it can be defended on its own terms feels responsible and is actually a surrender. You've just handed someone a discrete thing to cut, in a quarter where its return will read as zero. Education isn't adjacent to the AI program. It's the mechanism by which the AI program produces anything, and it belongs inside the plan, not beside it.
The Fight You Actually Lost
Which brings us to the conversation nobody wants to have.
A per-seat token cap arriving from Finance is not a budgeting event. It's a status report on a relationship. It says somebody with a technology title asked for money without a plan behind it, got a reasonable question, couldn't answer it, and lost.
Too many CIOs, CTOs, and CAIOs have accepted that position and started calling it pragmatism. I want them to take it back.
I've never had this problem, including in roles where IT reported through Finance, and it has nothing to do with speaking finance as a second language. I don't. What I know is what the CFO actually needs.
Your CFO doesn't care about your education spend. Your CFO doesn't care about your token spend either, whatever the internet tells you about finance people and their aversion to anything that doesn't show returns this quarter. Those lines get challenged because those are the lines where weak planning shows first. They're the soft tissue. A CFO probing there isn't attacking your priorities, they're testing whether you have any.
Show them the number is a piece of a strategy rather than a guess, and they move on. Mine has never pushed past a cursory question, because I've never shown up without a plan.
We're peers. If you think my ask is unreasonable, or that I'm flying by the seat of my pants, you absolutely get to ask. That's your job and it makes the company better. What you don't get to do is redline a plan you haven't been walked through.
And holding that line isn't about posture. It's about doing the job. When you hold P&L, the job is to build a strategy, build a plan to execute it, establish a budget, and defend all of it. Those aren't four separate skills. They're one skill, and the defense is the proof that the rest is real.
If you can't defend it, you can't execute it either. That's a much worse conversation, and it's the one waiting at the end of every cap you accept without arguing.
What To Do Monday
TO CIOs, CTOs, AND CAIOs: Walk into the next budget review with the plan, not the ask. If you can't explain in three sentences what outcome the AI spend is supposed to move, don't blame Finance for the cap. Write the plan first.
TO ENGINEERING LEADERS: Stop reporting consumption upward as though it means something. Govern tokens as a cost, pair every token number with a value or quality number, and teach your team to match model and reasoning level to the shape of the request rather than to their own title. That last one is free and it's the highest-return thing on this list.
TO CFOs: Ask whether the plan is fiscally sound. That's the right question and somebody should be asking it. But if the education line looks soft to you, ask for the strategy behind it before you take a pen to it. If your technology leader can't produce one, you've found a much bigger problem than a line item.
The tools are working. Engineers are the population most likely to turn this spend into real acceleration, and that deserves serious money rather than a nervous ceiling. What isn't working is the measurement, and underneath the measurement, the planning.
Govern the tokens. Fund the education. Defend the plan.