Uber burned through its entire 2026 AI budget in four months and has little to show for it

[Alfonso Maruccia]

In brief: Uber is one of the Silicon Valley companies that has invested the most heavily in AI. However, its top executives are now sounding the alarm that the expected productivity and efficiency gains have not yet materialized.

Uber has increasingly positioned AI as a core part of its technology stack, and AI models now represent a significant portion of its corporate spending. Costs are rising, but the company has yet to see meaningful returns in terms of ROI. According to President and Chief Operating Officer Andrew Macdonald, the ride-sharing company may soon begin to question its extensive use of "vibe coding" and other AI-driven development tools.

Macdonald was recently interviewed on the Rapid Response podcast, where he said Uber's internal engineering teams are increasingly using Anthropic's Claude Code to generate software. However, the impact of this AI-assisted coding on overall productivity is difficult to measure.

He noted that Uber cannot clearly connect usage metrics for tools like Claude Code with the delivery of "useful" features to customers.

Macdonald's comments come just weeks after the company revealed the financial impact of its AI spending, with its entire 2026 AI budget reportedly exhausted in the first four months of the year.

According to Uber CEO Dara Khosrowshahi, 10% of the changes made to the company's code now come from AI agents. Human employees still review the agents' outputs across areas such as marketing, legal, and software development – at least for now. However, Macdonald said that the lack of tangible results from vibe coding and AI agents is making the AI budget harder to justify.

Uber's AI over-commitment comes in light of a heated debate about costs and ROI across the technology industry. Many top executives are easily impressed by the apparent "magic" of vibe coding, but underlying organizations are struggling with the lack of a clear strategic approach to AI adoption. Even worse, end users are signaling growing concern over the widespread, unregulated adoption of LLMs and chatbot-like designs.

Credit: App Economy Insights

If an "AI-first" company like Uber cannot justify the cost of AI technology, many other Silicon Valley companies are likely facing even worse financial conditions. Earlier this month, Microsoft was allegedly forced to cancel its Claude Code licenses and migrate developers to its own GitHub Copilot CLI platform.

A recent Gartner study highlighted how AI model costs might decrease over the next few years, potentially falling to around 90% of their 2025 levels by 2030. However, agentic AI will require many more tokens to perform its tasks, which means enterprises are not likely to see meaningful savings on their AI expenditures anytime soon.

Permalink to story:

Uber burned through its entire 2026 AI budget in four months and has little to show for it

 

[bviktor]

Nowai, who woulda thought?!

 

[GeoffreyA]

On the universe's Wall of Infinite Drawers, pushing one drawer in causes another, or three, to spring out elsewhere.

 

[ScottSoapbox]

That cost change is likely supposed to be falling BY around 90% not falling to around 90%. A 10% increased efficiency over several years is nothing.

 

[daffy duck]

Given the CEO is an unabashed fElon Musk groupie, no surprise.

 

[Watzupken]

A lot of workers are forced to use LLM as part of their workflow and usage is tracked. To get around this, I won't be surprised that people are just using it for the sake of using it to demonstrate usage to management. So tokens are wasted just to check the "use LLM" box. The end result, higher cost to pay for LLM and minimal improvements to the companies using it. I don't believe this is an issue isolated to Uber of a handful of companies. There is a fundamental disconnect between senior leadership and workers on the ground around the results of using LLM to streamline work.

 

[RedGuard]

We are encouraged to use LLMs to generate code at work, but:

- my manager doesn't really know what we should build. It's been 3 times already that I constructed something and he told me he wants something else;
- we have to use Gemini which sucks big time. The code it generates is either too convoluted or goes to solve some probable issues from the future.

 

[noblody]

AI works well for generating greenfield projects, quick POCs, or small tools where code quality and long-term maintainability are less critical. The problems start when you introduce it into a large codebase with years of accumulated business logic and interconnected systems. In that environment, the model often loses track of context, starts making unrelated changes, rewrites code that was not part of the issue, or even breaks previously working functionality while trying to fix something else. It becomes very easy to burn through huge amounts of tokens with little real progress. The larger the project and context window become, the more output quality tends to degrade. Eventually you end up in loops of “sorry, I forgot” or “I accidentally removed that logic” while the model continuously overwrites its own work. If people were allowed to just use it where/when it works it would be great, but nooo, you need to prove you are using it intensely, fixing problems and shipping features has become secondary.

 

[user556]

And all this is still burning investor cash.

 

[stewi0001]

We are encouraged to use LLMs to generate code at work, but:

- my manager doesn't really know what we should build. It's been 3 times already that I constructed something and he told me he wants something else;
- we have to use Gemini which sucks big time. The code it generates is either too convoluted or goes to solve some probable issues from the future.

ouch! if you are not using claude for coding, then might as well not use AI

 

[wiyosaya]

Obviously, IMO, Uber's CEO is just following the lemmings over the cliff.