Anissa Gardizy, a reporter at The Information, surfaced a quote on Monday from Uber CTO Praveen Neppalli Naga that every engineering leader planning 2026 AI budgets should stop and read:

I’m back to the drawing board, because the budget I thought I would need is blown away already.

Uber opened Claude Code access to its engineers in December 2024. Back then, 32% picked it up. By February of this year, 63% were on it. By April, Praveen told Laura Bratton at The Information that the annual AI budget was already torched.

That isn’t a rollout story. That’s a pull story.

The curve isn’t what finance planned for

When a company moves from 32% to 63% internal adoption in fourteen months, the platform team isn’t pushing. The engineering team is pulling. Aakash Gupta broke down the rest of the numbers: 92% of Uber’s developers now use AI agents monthly. Between 65% and 72% of the code that gets written inside the IDE is already AI-generated. 11% of pull requests are opened by agents, not humans. Uber’s internal AI review system, uReview, touches more than 90% of the 65,000 diffs the company ships every week. 75% of the comments uReview posts get marked as helpful.

These are production numbers, not pilot numbers.

Which means Uber can’t walk it back. The engineering flow has absorbed the tool. If somebody cut Claude Code tomorrow, developer velocity would drop in a way the board would see immediately. If nobody cuts it, margin drops quietly. Pick one of those two problems — you’re not actually getting to skip both.

The real unit isn’t a seat

Most enterprises still budget software the way they budget SaaS: cost per seat times headcount times twelve months. That mental model breaks in the first week of agentic workflows.

Claude Code for enterprise teams runs $100 to $200 per developer per month on the Premium plan, with consumption billed at API rates on top of the seat. The community-reported average daily burn lands around $6 per developer. Multiply that across 5,000 engineers and you get $30K in daily base spend before you open the rest of the invoice. That number assumes linear usage. AI isn’t linear.

Agents spawn sub-agents. Each sub-agent carries its own context window. A single pull request can trigger a chain that runs for thirty minutes, consuming tokens with no human watching. Drop that inside a CI/CD loop and the spend detonates on every commit.

Gartner measured it in March: an agentic workflow uses between 5x and 30x more tokens per task than a standard chatbot. AnalyticsWeek named the failure mode “Agentic Resource Exhaustion”: one agent in a semantic loop can incinerate thousands of dollars in a single afternoon. Their aggregated estimate across the Fortune 500 this year sits near $400 million in unplanned cloud spend attributable to agentic AI.

Jason Calacanis wrote it out in public a few weeks back: his team’s agents were running $300/day. Annualized, that’s over $100K per agent, operating at a fraction of capacity. This isn’t a Fortune 500 problem exclusively. The people paying attention earliest are narrating the alarm in real time.

Uber can’t get off the train

When you see a bill of this size, the reflex is obvious: pause, renegotiate, roll back. Uber doesn’t have that option.

11% of your PRs opened by agents and 90% of your diffs reviewed by an internal AI system isn’t a layer you can switch off without breaking the development cycle. And you can’t break the development cycle when you’re competing with Lyft across rider-driver matching, dynamic pricing, and real-time logistics across 70+ countries. Uber built its operational moat on software. AI now lives inside that software.

That leaves one lever that still works: govern consumption, don’t cut it.

The CFO and CTO show up to the meeting with different numbers

Annual budgeting was built for predictable SaaS and headcount. Two line items with linear, explainable unit costs. Agentic AI is neither.

CIO.com reported Gartner’s finance practice VP saying CFOs don’t really know what AI costs — estimates are running off by 500% to 1,000%. Glean audited enterprise AI budgets and found most underestimate true total cost of ownership by 40% to 60%. McKinsey had already quantified the other side: for every dollar a company puts into model development, they typically spend three on change management, training, and workflow redesign.

Stack the three numbers and you get a consistent picture. Companies aren’t budgeting AI badly out of carelessness. They’re budgeting AI badly because the financial model they inherited was never designed to measure exponential consumption.

The part that doesn’t get repeated enough: per the 2026 State of FinOps report, 98% of organizations now actively manage AI spend. Two years ago that number was 31%. And at the same time, 91% of C-suite executives plan to increase agentic AI budgets this year. Nobody wants to spend less. Everyone is realizing the governance work nobody did when the contract got signed is the actual project.

Praveen is budgeting from that spot right now. Most CTOs I talk to are one quarter behind him.

Taste debt’s financial sibling

I wrote yesterday about taste debt — the quality bill that accrues when you pull the human out of the loop too early. Today’s bill reads differently on paper. The underlying mechanism is identical.

Taste debt shows up in brand, decisions, and customer complaints. Uber’s budget overrun shows up in P&L, board decks, and quarterly conversations with the CFO. Two faces of the same liability: silent accrual that compounds while nobody is looking, visible only when someone arrives to collect all at once.

The amortization method is the same too. You don’t stop using agents. You install a governance layer that measures, limits, pauses, and demands human sign-off where the cost of an error — either quality or financial — outweighs the cost of the review.

Team OS is the name we use for that layer. Same layer for both liabilities.

Three moves every CTO should have on next week’s agenda

After the Uber news, any technology leader scaling agents has three items that belong on the next steering committee agenda. I’m listing them in order of effort.

Per-workflow visibility. Not per seat, not per team. Per agent, per execution. If your AI dashboard only tells you how many licenses got activated, it tells you nothing. Step one is tokenized accounting: every agent, every run, every context window, tagged. Without that, you can’t set limits because you can’t see where the money leaks.

Budgets with circuit breakers. Cap per agent, cap per team, cap per workflow. Automatic circuit breakers when an execution crosses a threshold without a result. An agent in a semantic loop burning $3,000 in an afternoon shouldn’t be possible without a human-in-the-loop alarm somewhere in the middle. FinOps-for-AI platforms already exist. Adoption is slow because companies assumed the cost was predictable.

Quality gates that double as cost gates. Where error cost is high, a human signs. Where error cost is low, the agent runs free. That boundary doesn’t get drawn in a quarterly workshop. It gets instrumented. If a workflow wants to spend more than X dollars or modify more than Y files, it needs approval before continuing. That isn’t friction. That’s insurance.

None of the three slows productivity. All three make productivity measurable, and therefore budgetable.

What won’t work

Three moves I expect to see in the next six months that aren’t going to solve this:

Freezing AI usage. This is the finance team’s reflex when the invoice lands. It’s structurally impossible once the flow has absorbed the tool. Freezing AI at Uber today is freezing a meaningful share of the development cycle. No CTO signs that.

Renegotiating the enterprise contract. Useful at the margin. But the problem isn’t the vendor’s price sheet. It’s the internal consumption pattern without governance. Even at zero cost per token, an agent in a loop still steals machine time and context.

Hoping the next model generation will be cheaper. It will be cheaper per token. It will be more expensive per task. When a model gets more capable, teams point it at longer, more complex, more multi-step work. Total cost per useful task goes up, not down. That’s Jevons’ paradox applied to tokens, and it’s already visible in the pricing data.

What IQ Source does with this

The Uber story is the most visible version of a problem any company scaling agents without consumption instruments is going to hit. Fortune 500 or a fifty-person startup: the token math doesn’t discriminate by headcount.

Our job is installing those instruments. Not an abstract framework. A concrete method for answering four questions every CTO should be able to walk into the next CFO conversation with: how much is being spent, across which workflows, what ROI each one is returning, and where the circuit breakers sit.

We implement it with Team OS — the same governance layer we use for the taste problem, applied to the consumption problem. The two liabilities live in the same flow. Treating them as separate projects is a diagnostic error.

The advantage of installing this before your company ends up in The Information is that early governance compounds into a real moat. The companies that figure out token cost optimization first are going to hold a structural edge over every competitor still running annual budget cycles against exponential adoption curves.

Uber won’t be the last company to have this conversation. The question is whether you have it before the board meeting, or during it.

Before the next board meeting, know the number. That’s where we start.