At Uber, AI is already exceeding budgets
According to The InformationUber Technologies CTO Praveen Neppalli Naga says heavy use of code generation tools has already consumed the entire AI budget planned for 2026, just a few months into the fiscal year. The rapid adoption of solutions like Claude Code, developed by Anthropic, has greatly accelerated this drift.
Beyond the budgetary anecdote, the episode is actually part of a logic already explained by the CTO himself. In a recent interview, Praveen Neppalli Naga describes an organization where AI is deployed at all levels: acceleration of software development, automation of customer service via agents, and direct integration into products on the driver and user side.
This positioning partly explains the observed drift. By internalizing a culture of massive use (“AI-first” in products, co-pilots for developers, automation of operations), Uber mechanically transforms AI into resource consumed continuouslyand not as a one-off investment. The bill then becomes proportional to the intensity of use.
In an environment where “fixing bugs” and continuously iterating are operational imperatives, code generation tools become essential. This functional dependence de facto reduces the ability to arbitrate usage based on costs.
Here we find a structural rebound effect: the more AI improves team productivity, the more it increases the volume of software production… and therefore the consumption of resources. The local gain becomes global inflation. In a large-scale organization like that of Uber, this phenomenon is all the more difficult to contain as it is diffuse and decentralized.
Finally, at Uber, economic optimization comes after competitive advantage. Exceeding the AI budget then appears less as a deviation than as the symptom of an assumed strategic choice, where cost control is temporarily relegated behind technological acceleration.
There remains a structuring question: as AI becomes a critical variable cost, will Uber be able to maintain this model of intensive use without fundamentally rethinking its stack, or even without internalizing part of these capabilities, to regain economic control?
An inspiring example in its ambition and speed of execution, but which reminds us that the industrialization of AI now requires an economic discipline as rigorous as its technical deployment.
Meta transforms employee interactions into training data for its AI
According to Reuters, Meta Platforms plans to collect mouse movements, clicks and keystrokes from its employees on certain internal tools in order to train its artificial intelligence models. The objective is to produce agents capable of reproducing real user behaviors (navigation, sequence of actions, decision-making in interfaces).
After having widely exploited public web data, AI players are entering a phase where the data becomes behavioral. It is no longer just content that feeds the models, but sequences of actions, in other words a fine capture of “how to do it”. We are thus moving from generative AI to execution AI.
Meta affirms that this data will not be used for individual evaluation purposes, without specifying whether it could, ultimately, fuel the development of agents capable of automating some of the tasks currently carried out by its employees.
At the same time, according to Reuters, Meta Platforms plans a first wave of 8,000 job cuts from May 20, or nearly 10% of its workforce, as part of a reorganization focused on artificial intelligence.
From Airbnb to OpenAI: Emmanuel Marill takes charge of the EMEA region
OpenAI appoints Emmanuel Marill as Managing Director for EMEA, a newly created position based in Paris. The former executive of Airbnb, where he led European operations, will be responsible for accelerating the adoption of ChatGPT in the region, which already has more than 120 million monthly users.
Having worked at LVMH, Roland Berger, Groupon and Meta Platforms, Emmanuel Marill brings experience in regulated markets and complex competitive environments. In particular, he will have to deal with the European framework, between Digital Services Act and AI Act, while strengthening the presence of OpenAI against competitors like Anthropic.
