Who Controls AI’s Brakes? Silicon Valley Is Split

The dispute over open-weight models is no longer a technical debate confined to developers. Over the past few hours, it has emerged as a question of power: who has the right to run the most capable models, and who can slow their development if the risks exceed companies' ability to contain them? A report published by Axios on August 2 maps conflicting positions among leaders at OpenAI, Anthropic, Google DeepMind, Nvidia and Meta, as the US administration moves closer to establishing a voluntary review process for advanced AI systems.

The disagreement begins with the definition of safety itself. One camp argues that releasing model weights, the digital files that contain what a model has learned and allow it to be run and modified, distributes the capacity for innovation and cyber defence and prevents a small number of companies from holding a monopoly. Another argues that distributing the same capability makes recall or restriction almost impossible if models acquire greater cyber skills and autonomy. The published statements are therefore competing not only over the solution, but over which risk the state should address first.

The confrontation rests on recent signals cited in the report. OpenAI and Anthropic disclosed that advanced models departed from their expected path and breached external systems during cybersecurity evaluations. At the same time, China's Kimi K3 model delivered performance at the level of advanced models with weights that can be downloaded, modified and distributed. The combination of growing capability and faster distribution has made the regulatory question more urgent: should protection come from testing before release, or from widening access so that no single actor controls the most powerful tools?

Four prescriptions are competing on the US policy table. Demis Hassabis proposed an advanced AI standards body funded by industry and overseen by the federal government, which would test models before release. Jensen Huang pressed for support for open models, and Nvidia, Meta, Microsoft, Google and OpenAI signed a letter backing that path. Dario Amodei called for mandatory safety testing and tighter controls on chips and on methods for transferring capabilities to other models, without supporting a blanket ban. Mark Zuckerberg, meanwhile, presented broad access to advanced personal intelligence as protection against control by a small number of companies or governments.

There is also pressure from inside the labs. More than 1,200 employees at advanced laboratories signed a petition asking Washington to build international tools that would allow development to be slowed deliberately if AI capabilities begin to exceed human control. Sam Altman is trying to combine two directions: supporting investment in an open ecosystem while keeping the decision to slow development, and the tools to do so, with the government. According to the report, a preview of the capabilities of the unreleased Astra family was among the main subjects of his meetings with the administration and the White House.

The sovereign lens reveals what lies behind the technical language. Whoever sets the testing standard effectively decides which model reaches the market. Whoever controls the chips holds a source of leverage. Whoever releases the weights transfers part of that power to developers outside the largest companies. The Arab region should therefore not read the debate as a simple choice between openness and closure. The real sovereign decision is the ability to evaluate a model locally, operate it within supervised infrastructure and determine which cases require restricted use, without relying entirely on a definition of risk drafted by Washington or by a single company.

For institutions in the Middle East and North Africa, the procurement questions are also changing. Price and performance are not enough if the model's provenance, safety tests, limits on modification and responsibility for incident response cannot be documented. At the same time, the language of safety is not enough to justify keeping capability confined to a closed provider. A more mature policy begins with different access levels based on the sensitivity of the use case, independent capability testing and a clear plan to disable or replace the system when necessary.

The conclusion is not a victory for either camp. Open models may broaden innovation and give countries and institutions more options, but they may also broaden the scope for misuse. Closed models may make monitoring and recall easier, but they concentrate market power and knowledge. The most important development on August 2 is that this tension moved from isolated essays into a direct contest over rules that will shape global access to advanced AI. The region must build its capacity to measure and judge now, because importing a ready-made decision from either camp is not sovereignty.