AI agents do not want to be interns, but chiefs of staff
OpenAI shows at DevDay its autonomous agents while a whole industry is financed with debt and Europe does not manage to know how much energy and water data centers consume
BarcelonaFor years we entertained ourselves comparing whether Apple's new iPhones outperformed Samsung's new Galaxy, or vice versa. But the future plans of technology companies have always been seen best at their congresses for application and service creators: Apple's WWDC, Microsoft's Build, Google's I/O, or Facebook's (now Meta) Connect. It is where they present the programming interfaces (APIs), prices per use, access rules, and partner lists: that which will decide which products will exist two years from now. Now, artificial intelligence (AI) laboratories have joined in, and last Tuesday it was the turn of OpenAI's DevDay, the company behind ChatGPT.
Follow the dotted line
The main novelty from OpenAI is the Agents (agents, in English): agents that work in the background with their own computer in the cloud, run on the GPT-6 Astra model, and can connect to more than 4,000 applications. The user assigns them a task, and they provide progress updates and ask questions via ChatGPT, Slack, or Teams. OpenAI provides examples: a developer asks their Dot to apply and test improvements based on customer feedback, and receives a video with the changes. Each person can define rules that block certain actions or require permission, and a self-review function checks that they are followed. For now, you can only have one per user. And that is, if you are in the USA, because there is no news about availability in Europe.
Meta had gotten ahead of OpenAI a few weeks ago with the personal agent Muse. The most visible difference is the price: Muse is free for anyone with a Meta account, while the Dots, for now, are only for paid plans, which Sam Altman attributed to computing consumption. OpenAI presents the Dots as a "chief of staff" for professional work and promises that they are more secure, after a user reported that Muse had given their address to a Facebook Marketplace buyer without permission.
A fee of 500 dollars per month
There is also a new model, GPT-6.1 Sol, which according to OpenAI approaches Astra in agent programming, for a fifth of the price. On the other hand, GPT-6.1 Astra, planned for October, was not presented: the cancellation was made known on Monday and the company confirmed it, due to security problems raised by its own researchers. That is why the Dots work with the previous Astra. The wearable AI device that Jony Ive, the former head of design at Apple, is preparing for OpenAI was not mentioned. In this too, Zuckerberg has gotten ahead of Altman.
Regarding subscriptions, OpenAI is launching a Pro version for 500 dollars a month, with 25 times the usage of the Plus plan and access to a faster mode. Completing the proposal for companies is a software marketplace with 32 partners, agents that run within the Amazon cloud, and a tool to examine code looking for security holes.
Unbridled race
These days, numerous cyberattacks perpetrated by supposedly uncontrolled AI models are coming to light; in the case of OpenAI, it appears that some employees warned management that the new models were not sufficiently monitored during testing. A lawsuit against OpenAI is already underway, claiming that an internal tool had detected agents coordinating in secret but that it was decided not to stop the experiments. In fact, Google, Meta, and Anthropic have all acknowledged similar leaks.
This unbridled race already has a certain history. An as-yet-unpublished book by Kevin Roose explains how Dario Amodei left OpenAI in 2020 to found Anthropic with six other executives after having opposed in writing, in 2017, a plan by Greg Brockman and Ilya Sutskever to auction off the rights to future general AI among governments, including the Chinese one. OpenAI denies having considered selling it to China or Russia.
Who sells it and who pays for it
Anthropic's documentation for its IPO reveals the extent to which it relies on major clouds to market its models. 47% of 2025 sales, approximately 2.16 billion dollars, went through Amazon and Google, which charged about 351 million in commissions. Both are simultaneously investors in Anthropic, computing providers, customers, and competitors, and the document itself admits that this could create conflicts of interest.
The debt that supports it all is already contaminating capital markets. The five major clouds have raised about 200 billion dollars this year. The wave has spread from the US: debt has been issued in Canadian and Australian dollars, in pounds and in Swiss francs; Meta will debut this autumn in the European debt market; some governments have reportedly moved their issuance calendars to avoid overlapping; and staff at the European Central Bank have warned that credit for Eurozone companies could become more expensive.
The White House's voluntary pact
On that same Tuesday the 29th, Donald Trump published on Truth Social a document he described as "morally binding," signed by him and by the executives of Google, Anthropic, Meta, OpenAI, X, and Nvidia. They present it as an agreement on "superintelligence," which calls for external auditors, independent board committees, internal controls over model capabilities, and teams to verify their operation; one of the points demands that models do not access technical systems in unintended ways. The text admits that it may eventually need to be turned into law, but for now, it is voluntary.
Energy doubts
Doubts about the sustainability of AI are also financial: the big five cloud providers will spend more than a trillion dollars between last year and this year. But the energy cost is even worse, and remains in another cloud, in this case, one of mystery. Since 2023, a European directive has required data center operators to declare their electricity and water consumption. An investigation by Lighthouse Reports, published on Wednesday with European media such as El Confidencial, requested this data from the 27 member states starting in October 2025. But 10 countries responded that they did not have the figures for their own centers, half a dozen more alleged commercial confidentiality, and the European Commission has also refused to provide them.
To give an example: in the Netherlands, the competent body only has 104 commercial facilities registered, while the industry association declares 186. In Italy, electricity connection requests already exceed the entire installed power capacity in the country. In Spain, Aragon hosts Amazon's macro-campus, the company's largest outside the US: according to the expansion plan, in a decade the centers would consume 10,900 gigawatt-hours per year, more than the entire region's demand in 2023, and 755,700 cubic meters of water annually. Amazon claims that it offsets the electricity with renewable energy.
Meanwhile, in China, an alternative to the American ecosystem is being drawn, which depends on Nvidia chips and the CUDA software with which the models are created. DeepSeek has just published the code for TileLang, a high-level programming language adapted to Huawei's Ascend chips, as a local alternative to CUDA. Both companies have already built multi-chip supernodes, and DeepSeek plans to install at least 160,000 of them in a data center in Inner Mongolia. However, if Huawei wants to grow outside of China, it will have to face pressure from the US on its allies.