Kiseki Labs Request a briefing
Business

Tokenmaxxing: What Uber's Blown AI Budget Teaches CFOs

17.06.2026 Updated 06.10.2026 10 min read Eddie Forson

Uber ran out of AI budget by April. What tokenmaxxing is, why AI bills beat forecasts, and why cost per outcome, not tokens, should set your 2027 budget.

Neon visualisation of a torrent of glowing tokens pouring into a holographic budget gauge pinned past its maximum, flanked by leaderboard bar charts

By April, Uber had spent its AI budget for the whole of 2026. "I actually blew through our AI budget in a quarter for the whole year, essentially," chief executive Dara Khosrowshahi said on the Invest Like the Best podcast. "And it is forcing us to adjust."

Nothing had gone wrong with the tools. Anthropic's Claude Code had spread across roughly 5,000 Uber engineers, they liked it enough to use it all day, and some were reportedly being ranked on internal leaderboards for how much of it they used.

Six months on, the companies that have gone public about their AI bills have taken opposite routes. Uber reportedly capped every employee at $1,500 a month per AI coding tool. Coinbase says it cut its AI spend nearly in half while its token usage kept growing. If you're setting next year's AI budget this quarter, which route you take matters more than the number you pick.

What is tokenmaxxing?

Tokenmaxxing is treating AI usage, measured in tokens, as a goal in its own right. Tokens are the units AI models bill by, roughly three quarters of a word each. As they became the most visible measure of AI adoption, some companies started ranking people on how many they burned. Uber reportedly ran internal leaderboards, and Meta reportedly built dashboards that ranked employees by token consumption.

Measure usage and you get usage. On the a16z podcast, Martin Casado described speaking to someone at one of the large companies known for counting tokens, who told him: "me and my coworkers have agents do useless tasks". The Observer's John Naughton named the underlying flaw: companies were rolling out "stunts such as internal leaderboards that rank teams by AI tool usage without calculating what the cost of success might be."

We think the leaderboards made sense for a few months. Organisations genuinely needed people to experiment, and a league table is a crude but effective way to get them started. The trouble is that nobody attached a measure of value to the usage, so when the bill arrived, nobody could say whether it had been worth it.

Why AI bills outrun AI budgets

Seat licences are predictable. Token billing isn't. Agentic coding tools are increasingly priced by usage, so an engineer using simple autocomplete costs a fraction of one running several agents across a large codebase. As Janakiram Msv put it in Forbes, "The same tool, the same engineer, the same workday, can produce wildly different invoices depending on workflow choice." Uber's engineers reportedly ran up API costs of between $500 and $2,000 a month each.

Agents also burn far more tokens per task than the chatbots they replaced. An agent reads files, plans, calls tools, checks its own work and retries, and every step is billed. That's why a ten-person pilot on a tidy set of tasks says so little about what five hundred engineers will spend once agents are doing real work.

And cheaper tokens haven't made AI cheaper. Gartner expects inference on the most sophisticated models to cost AI providers 90% less by 2030 than in 2025, but it doesn't expect them to pass the full saving on to customers. Manish Kapur summed up the result: "The unit gets cheaper. The bill gets bigger. Both are true at the same time."

Vendors have been rewriting their pricing accordingly. GitHub moved Copilot to usage-based credits on 1 June. Anthropic announced a similar meter for agent usage on its paid Claude plans, then paused it on 15 June, the day it was due to start, while it reworks the plan. It was still paused in early October. A budget that assumes flat-rate, all-you-can-eat AI will last is betting against the direction of travel.

Three companies, three responses

Three companies' next moves have since been reported.

CompanyWhat happenedWhat it did next
Uber2026 AI budget gone by AprilReportedly capped token spend at $1,500 per employee per coding tool
MicrosoftGave engineers company-wide access to Claude CodeReportedly began cancelling most direct Claude Code licences, moving engineers to GitHub Copilot CLI
CoinbaseToken usage growing fastSays it cut AI spend nearly in half using defaults, routing and caching

Commentators blamed Microsoft's switch on cost, though Microsoft also owns GitHub, so it's hard to read much into it. Uber's cap is the more telling response. Caps stop the bleeding, but they also cap the upside, and they treat your heaviest users as your biggest problem when some of them are your biggest wins.

Coinbase's route is the more interesting one. Brian Armstrong's summary was: "Not with friction and spend alerts. With better defaults, routing, and caching." In its LLM gateway, the internal layer its model calls pass through, Coinbase is experimenting with open-weight models (which anyone can download and run) as the default. It uses "a frontier model for planning, but not for execution". And it raised the share of requests answered from its cache, rather than paid for again, from 5% to 60% in its internal chat tool.

Measure cost per outcome, not tokens consumed

As Kapur puts it, "Measure cost per successful outcome, not tokens consumed." Tokens consumed tells you nothing about whether the spend is working. Cost per merged pull request, per resolved ticket or per correct answer does.

WorkloadThe outcome to countThe metric
AI codingPull requests merged, bugs fixedCost per merged PR
Customer supportTickets resolved without escalationCost per resolved ticket
Analytics and reportingQuestions answered correctlyCost per correct answer
Classification and routingItems labelled correctlyCost per thousand correct decisions
Document processingDocuments verifiedCost per verified document

Reliability is the lever most teams forget. On one of our engagements, a natural-language analytics system was consuming 20,000 to 50,000 tokens on simple queries, and around 40% of all queries failed. Every failure is spend with nothing to show for it. The arithmetic is simple: at a fixed cost per query, lifting success from 60% to 98%, the rate that system reached on its 35-case evaluation suite, cuts the cost per correct answer by about 40%.

Getting to cost per outcome takes visibility first. Msv observed that most enterprises "do not yet apply to artificial intelligence tooling the spending controls that DevOps teams routinely apply to cloud compute." Sebastian Buzdugan puts it neatly: "For fifteen years, FinOps meant compute, storage, and network. Now there is a fourth line: tokens." Attribute spend to teams, workloads and models, and inventory every API key. On limits, Buzdugan advises soft limits and alerts first, with hard caps only where abuse is obvious. John Hwang would make basic hard limits per employee the first priority. We'd start softer, because a hard cap set before you know your cost per outcome is a guess.

How to set next year's AI budget

Uber's problem was a budget, so it's worth being concrete about how to set one. A fixed annual number built from seat counts will be wrong. Hwang argues that "you can't put a clean, fixed budget on a spend that can increase geometrically and unpredictably. Instead, things should be aligned with work."

In practice we'd budget per workload: the outcomes you expect from it, multiplied by the cost per outcome you've measured, with a variance band for the months when usage jumps. Where volumes are predictable, Msv's advice to "negotiate committed-spend agreements at fixed rates" is worth taking to procurement. A budget built this way can grow when a workload proves its value, and it tells you which line to cut when one doesn't.

Explore on the frontier, then scale on the cheapest model that works

Khosrowshahi described Uber's own correction: "we're using the more expensive models to explore… Once we scale some of these experiences and interactions, we'll look to bring in more efficient models that are more efficient on a token basis or open source, again, lower cost."

The question to put to every recurring workload is the one Cosmin Vladutu asks: is this problem worth this model? A nightly summarisation job doesn't need your most expensive reasoning model. A novel architecture decision might. At list prices, classifying a million articles costs about $2,250 on a mid-tier model such as Claude Sonnet 5 and about $46 on a purpose-built decision model. On 540 Anansi News articles, the decision model also agreed with our reference more often than gpt-4o-mini, the small model it would replace (92.6% against 79.8%).

There's no universal answer, even for the same kind of step. Coinbase keeps a frontier model for planning. On our analytics engagement, the frontier reasoning model we first chose for the planning step took over 60 seconds on simple queries and had intermittent reliability issues, so we switched to a less advanced, faster model. Planning dropped to under ten seconds, and the full pipeline still reached 98% on the evaluation suite. Both choices can be right, for different workloads. The only reliable way to tell which situation you're in is an evaluation suite, a fixed set of real cases the system has to get right, run against each candidate model.

That's also why we'd route by workload before routing by request. Coinbase routes per request, but inside its own tools, where it knows the workload. A generic router that sees only the prompt is a different thing. Menlo Ventures, an investor in the router marketplace OpenRouter, counts more than ten companies that have launched their own routers, yet concedes that "routing based solely on the prompt is actually a bit of a fool's errand." Decide the default for each workload with evidence, and re-run the evaluation when prices or models change.

Treat tokens like capital

Uber's chief executive hasn't given up on usage. "So my view right now is drive adoption, encourage usage across the company," Khosrowshahi said in the same interview. The return is real: SemiAnalysis reckons tasks that once took tens of person-hours and cost thousands of dollars can now be done in minutes for a few dollars' worth of tokens. What Uber's bill exposed is what happens when usage has no measure of value attached.

Hwang's line is the one to put in front of a CFO: "The winners won't be the companies that spent the least; they'll be the ones that treated tokens like capital…" Capital goes where it earns a return, gets measured, and moves when it stops earning one. Nobody allocates capital with a leaderboard.

A four-week plan to get AI spend under control

Four weeks won't fix AI spend at a 5,000-engineer company, but it's enough to see where the money goes and set the first defaults.

  1. Week one: make spend visible by team, workload and model, and inventory every API key.
  2. Week two: pick one outcome metric for each of your three biggest workloads and calculate cost per outcome.
  3. Week three: build a small evaluation set for the highest-volume workload and test cheaper models, better prompts and caching against it.
  4. Week four: set defaults and a routing policy per workload, with soft limits and alerts, and hard caps only where abuse is obvious.

It's the same evidence-first approach we bring to every engagement: establish what good looks like, measure against it, then change the architecture.

Frequently asked questions

What is tokenmaxxing?
Tokenmaxxing is treating AI usage, measured in tokens, as a goal in its own right, often encouraged by leaderboards or dashboards that rank employees by consumption. The problem is that token usage measures activity, not value.

Why did Uber blow its AI budget?
Claude Code spread across roughly 5,000 Uber engineers and was billed by usage, and usage was reportedly encouraged through internal leaderboards. Engineers reportedly spent $500 to $2,000 a month each on API costs, and Uber exhausted its 2026 AI budget by April. It reportedly then capped spending at $1,500 a month per employee for each AI coding tool.

What is AI FinOps?
AI FinOps applies cloud cost management disciplines to AI: attributing token spend to teams and workloads, measuring cost per outcome, setting soft limits before hard caps, and routing each workload to the cheapest model that meets its quality bar.

Should we cap AI spend per employee?
Use caps as a backstop for obvious abuse, not the main control. Better defaults, workload-based routing and caching reduce spend without penalising the people getting the most value, and cost per outcome tells you where the spend is working.

If you can't yet say what last month's tokens bought, request an executive briefing and bring the invoice.


This article first appeared on 17 June 2026 as an edition of AI for Decision Makers, Eddie Forson's LinkedIn newsletter, under the title The token economics of tokenmaxxing: why your AI bill outran your budget. We've since updated it with what happened next and corrected a few details.

Next

Facing something similar?

Thirty minutes on your situation, and a straight answer on whether the work is worth doing.

Not ready for a conversation? Take the AI readiness assessment: twenty questions, about six minutes, and you'll see where your organisation is stuck before you talk to anyone.