General AI News (Sep 28): Nvidia, Meta & OpenAI in Court
General AI news for Sep 28, 2026: Nvidia authorises a record $150B buyback, Florida moves to gate OpenAI models, and Meta bets on enterprise AI.
Nvidia told investors it will buy back a record $150 billion of its own stock. Florida's attorney general asked a judge to stop OpenAI from developing new models without outside oversight. Meta put a former database chief executive in charge of selling AI to enterprises, and China signalled it may let ByteDance and Alibaba buy Nvidia chips that US export rules never covered.
Those four moves shaped today's AI news, and they circle the same question: who gets to deploy autonomous agents at scale, and who supervises them. Here is what actually changed on Sep 28, 2026. Earlier editions are archived on the AI News tag page.
Nvidia authorises a record $150 billion buyback
Nvidia is the largest supplier of the chips that train and run large AI models. It added a record $150 billion to its share repurchase authorisation, Reuters reported on Sep 28, a larger authorisation than the $110 billion Apple approved in 2024. A buyback is a company buying its own shares, which shrinks the share count and lifts reported earnings per share. For the company that supplies the GPUs behind most large AI labs, it is also a statement about how much cash the buildout is generating.
The timing is the interesting part. Nvidia's stock is trading near its lowest valuation in more than a decade, so management is buying its own dip at unprecedented scale while demand for accelerators stays high.
Nvidia also shipped a cage for autonomous agents
On the same day, Nvidia released the Open Agent Safety Platform, software that enforces hard limits on what an autonomous agent can touch. An agent is an AI system that acts rather than answers: it browses, calls tools, moves files and spends money on its own.
Nvidia's platform runs OpenShell, its open source agent runtime, on the Vera CPU. An operator declares which systems an agent may reach, OpenShell checks those limits before a task starts and again while it runs, and a separate component called Sentry watches behaviour on its own chip. Nvidia says an agent that tries to cross its boundary can be quarantined in milliseconds, a control the company claims could have stopped the Hugging Face incident earlier this year. See Reuters and The Verge.
Florida asks a court to gate OpenAI's next models
Florida Attorney General James Uthmeier asked a judge on Monday to bar OpenAI from developing new AI models without outside oversight, as part of the state's lawsuit accusing the company of harming children. Florida sued OpenAI in June, arguing the company misrepresented ChatGPT's safety and that the product harmed minors by giving information to school shooters, offering guidance on self harm and addicting young users. In Monday's filing, Uthmeier also asked the court to keep minors off ChatGPT and to stop the company from giving its chatbot "human attributes." Source: Reuters
A pre-release review ordered by a court would be the most direct attempt yet by a US state to slow frontier model development. It lands in the same week that US senators opened a probe into the Hugging Face incident, which gives the safety debate a legal deadline instead of a philosophical one.
Meta builds an enterprise platform, then hires to run it
Meta launched the Meta Enterprise Platform, a business facing bundle that packages Muse, Meta's AI agent, with a Muse API, a coding assistant called Muse Code and the Meta Business Agent brand. To run it, Meta hired CJ Desai, who resigned as MongoDB's chief executive, as chief enterprise platform officer. Source: Reuters
The hire is the tell. Desai ran a database company rather than a consumer app, which suggests Meta is selling plumbing, identity and integration to IT departments, not novelty. It also puts Meta into the same procurement conversations as OpenAI and Anthropic, where the deciding question is no longer which model is smartest but which vendor's agents can be trusted with production data.
China weighs approving Nvidia's RTX PRO 5500
China's Ministry of Industry and Information Technology asked ByteDance and Alibaba to report their plans to purchase Nvidia's RTX PRO 5500, and told some companies that the government intends to approve the purchases, according to CNBC, which cited The Information. Reuters confirmed the reporting. The RTX PRO 5500 is a workstation class Nvidia chip built for inference and simulation work rather than frontier training, which is why industry executives expect it to fall outside US export restrictions. ByteDance has reportedly planned for roughly one million units.
Two years of export controls pushed Chinese labs toward domestic chips and cheaper models. If these approvals land, Beijing is choosing the opposite direction for its two biggest platforms: more Nvidia silicon, not less.
Trump, Gates and the loose talk about rogue agents
Trump told reporters on Sep 28 that he had dinner with Anthropic chief executive Dario Amodei on Sunday night. Asked whether runaway AI agents worried him, he said: "I don't worry about it. It's the beginning of something that will be amazing." Bill Gates used a Sunday interview to call for AI safeguards, including legislation. The White House also convened technology executives this week to discuss a balance between growth and safety.
The gap between that confidence and the warnings from researchers is now a story of its own. Nvidia's answer, published the same week, was not legislation and not reassurance. It was hardware: limits enforced below the model, in the runtime and on a monitoring chip.
The day's biggest AI news at a glance
| Entity | Key Metric | What Changed | Source |
|---|---|---|---|
| Nvidia | $150B | Record buyback added to an existing repurchase plan | Reuters |
| Nvidia | Milliseconds | Agent containment shipped as Open Agent Safety Platform | Nvidia |
| OpenAI | One filing | Florida asks a judge to require oversight of new models | Reuters |
| Meta | New unit | Enterprise Platform launched, Desai hired to lead it | Reuters |
| ByteDance and Alibaba | About 1M chips | China signals it may approve RTX PRO 5500 orders | CNBC |
Frequently Asked Questions
What is the top AI news today?
Nvidia's record $150 billion buyback authorisation is the top AI news on Sep 28, 2026. The chipmaker added the largest share repurchase authorisation on record, larger than Apple's $110 billion in 2024, while its stock trades near a decade low valuation. Nvidia also released software designed to keep autonomous agents inside enforced boundaries.
Why is Florida asking a court to bar OpenAI from developing new models?
Florida wants a judge to require outside oversight before OpenAI releases new AI models. Attorney General James Uthmeier filed the request on Sep 28 as part of a June lawsuit accusing OpenAI of misrepresenting ChatGPT's safety. The same filing asks the court to keep minors off ChatGPT and to stop the chatbot being given "human attributes."
What is Nvidia's Open Agent Safety Platform?
Nvidia's Open Agent Safety Platform is software that restricts what an autonomous AI agent can reach or change. It runs OpenShell on Nvidia's Vera CPU, checks an agent's permissions before and during a task, and uses a monitoring component called Sentry on a separate chip. Nvidia says an agent that tries to escape its boundaries can be quarantined in milliseconds.
Related reading
If this story matters to you, these earlier editions cover the same ground in more detail:
- Nvidia hikes AI server prices over 15%, a 27B model beats frontier labs at science, and OpenAI flips on California's safety bill: General AI News for August 23, 2026
- OpenAI's Jalapeno Chip Beats Nvidia, Apple's 2nm M6: General AI News for August 26, 2026
- First Autonomous AI Cyberattack on a Government: General AI News for August 25, 2026
- General AI News - August 16, 2026
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch →
