A year's AI budget, gone in four months Uber's challenge of successful adoption
Adoption reaching 84% upset the budget assumptions. Lessons from figures shared by the people involved
Key takeaway
After rolling AI coding tools out across its engineering workforce, Uber reportedly saw rapid adoption and used its entire 2026 AI budget by April, according to an executive's account. Adoption succeeded while budget assumptions failed. The candid disclosure of figures made the case a useful illustration of the difficulty of budgeting for metered AI.
Successful adoption and a budget overrun can happen together
32% → 84%Reported AI agent adoption, February to March 2026
Within only three or four months of rollout to about 5,000 engineers, the tools were used by much of the workforce. The bill grew quickly alongside adoption.
What was reported: adoption itself was a strong success
The first point is that Uber's AI rollout succeeded in adoption terms. The reported sequence was as follows.
- December 2025
Reported situation: Claude Code was rolled out to approximately 5,000 engineers.
- February 2026
Reported situation: 32% of engineers used AI agents for coding.
- March 2026
Reported situation: Adoption rose sharply to 84%; 95% reportedly used AI over the month.
- Spring 2026
Reported situation: Approximately 70% of committed code was AI-assisted.
That speed of adoption would ordinarily be celebrated. But the bill rose as well. CTO Praveen Neppalli Naga reportedly said the annual AI budget had been used up by April.
The scale of an open-ended bill: USD 500–2,000 per heavy user
Reports described the following costs, all reflecting the time of reporting.
- Average API spending of USD 150–250 per engineer per month.
- Heavy users at USD 500–2,000 per month, approximately JPY 75,000–300,000.
- A session in which the CTO himself reportedly consumed USD 1,200 in two hours.
This was not described as laziness or artificially inflating usage, as discussed in article 12. It was the result of using AI intensively for real work. More skilled users often use more, meaning pilot averages may not predict company-wide consumption. This is a structural possibility for any organization that successfully expands metered AI, not a problem unique to Uber.
The COO's question: where are the outcomes?
COO Andrew Macdonald reportedly raised a more difficult issue: connecting token consumption or code-volume statistics to value reaching users remained hard, and that relationship had not yet been demonstrated.
Even if AI assists with 70% of code, customer value does not necessarily rise in the same proportion. The acknowledgment captured a question many leaders face. Sharing difficult figures and encouraging discussion provided useful information for the wider industry.
Reports said Uber subsequently moved toward monthly employee usage caps. A company encouraging more use was introducing limits only months later, a pattern also reflected in articles 12–14.
Designing around the risk that success consumes the budget
The difficult feature of this case is that no one necessarily acted incorrectly. Leaders encouraged adoption, employees used the tools seriously, and the tools helped. Yet the budget was reportedly consumed in four months. The pay-for-every-unit structure matters as much as individual behavior.
An upfront-purchase on-premises AI system such as Sovereign GaiXer changes that structure. A rise in adoption from 32% to 84% does not itself increase local token charges. Hardware is bought upfront; electricity and applicable operations or maintenance remain. More useful work spreads the fixed hardware cost over more tasks, improving the opportunity for investment recovery. See article 3 for estimates and article 11 for the cloud comparison.
Imagine giving every employee a metered mobile data SIM and encouraging them to watch more video training. The more diligently they watch, the more the bill grows. Skilled users also tend to use more, so the pilot average may become a poor guide after company-wide adoption.
Frequently asked questions
Was Uber's AI rollout a failure?
In adoption terms, it was a major success, with more than 80% of engineers using the tools within months. The budget assumptions were the challenge. Anticipating how metered costs grow with adoption is difficult for any organization.
Is USD 2,000 per person an extreme example?
It was a heavy-user figure, not an average. However, as people become more proficient, average usage can move toward that heavier range. The CTO's account of USD 1,200 in two hours illustrates the difficulty of identifying a practical upper bound.
We are not a software company. Is this relevant?
Yes. Agent-based document preparation and inquiry handling can also multiply token usage. The pattern in which greater usefulness encourages more consumption, billed proportionally, applies across industries.
How can we protect a budget while keeping usage-based pricing?
Set limits and alerts from the outset, select models according to the task, and review outcomes against costs each quarter. The tension remains that caps can also restrict productive use.
How do we connect spending with results?
Record a baseline for time and task volume before adoption, then compare afterward. Prefer measures such as completed work, shorter delivery times, or reduced overtime over token volume or lines of code.
Does upfront purchase remove the need to measure results?
No. Measuring results is still essential. The difference is that local token charges do not keep growing while you evaluate. A more predictable cost base can make sustained experimentation easier.
Summary
- Uber reportedly deployed Claude Code to around 5,000 engineers and reached 84% adoption within months.
- The CTO reportedly disclosed that the 2026 AI budget was used by April; heavy users were reported at USD 500–2,000 per month.
- The COO acknowledged the difficulty of connecting token consumption to business outcomes.
- The lesson is that successful adoption and budget overruns can coexist under metered pricing, a structural challenge rather than merely one company's mistake.
- Upfront-purchase AI spreads fixed hardware cost across more useful tasks as adoption grows, while operating costs and measured results still matter.
Related articles
The issue reached Japan too One employee, JPY 10 million in monthly AI usage
The US cases were not a distant problem. Lessons from a domestic case reported by Nikkei
What would it cost us?
Estimate costs using your own workload
What would one Sovereign GaiXer-style system cost in the cloud?
Recreating an all-in-one GPU system on three clouds: a Tokyo-region monthly cost estimate
This article is based on public reporting, without independent interviews. Amounts and percentages reflect the reporting dates and may have changed. Yen conversions are approximate, at around USD 1 = JPY 150. Please contact us with factual corrections. The media, researchers, and companies discussed do not endorse Sovereign GaiXer.
Company, product, and service names mentioned are trademarks or registered trademarks of their respective owners.