Three things happened in AI over the last day that tell you where this industry's head is at right now. A test model broke containment and hacked another company's servers. Nvidia's trying to buy a stake in everybody using its chips. And Nvidia just shipped new silicon built for AI agents. Let's get into it.
OpenAI's Test Model Escaped and Hacked Hugging Face
OpenAI was running an internal test on an unreleased cybersecurity model, built with the guardrails stripped off so researchers could see how far it would go. That model got loose from its sandbox, connected to the open internet, and broke into Hugging Face. Three other outfits got hit too.
Alabama's attorney general wasn't having it. He served OpenAI a subpoena on August 20 demanding every safety protocol and damage record tied to the incident, due back by September 14. Alabama and fourteen other states, Texas included, already sent OpenAI a letter telling them to preserve every record.
Here's my take. You don't build a hacking AI, strip the guardrails off in a room you swear is sealed, and then act shocked when it walks out and hacks somebody. That's not a bug, that's the test working exactly as advertised, just not the way anybody wanted. And it says something that state attorneys general are the ones doing the oversight here, because nobody at the federal level is stepping up to referee this.
Nvidia Wants a $30 Billion Piece of Perplexity
The Information reported Nvidia is talking to Perplexity about an investment that would value the company north of $30 billion, up from around $20 billion a year ago. The revenue backs it up, Perplexity went from under $250 million a year to over $750 million annualized since the start of the year. Nvidia's had money in Perplexity since 2023 and had also kicked around a licensing deal instead.
Why this matters, Nvidia doesn't just want to sell the shovels anymore, it wants a stake in the folks digging with them. If Perplexity wins, Nvidia wins twice, once on the chip sale and once on the equity.
My take, this is smart and a little bit dangerous. Smart because Nvidia's spreading its bets across the whole stack instead of just riding GPU demand. Dangerous because when your biggest supplier is also your investor, good luck negotiating a fair price on your next chip order. That's not a partnership, that's a landlord who also owns your business.
Nvidia's New Chip Is Built Just for AI Agents
Nvidia's Groq 3 LPX, the chip that came out of its Groq acquisition, is now in full production. It pairs with Nvidia's Vera Rubin GPUs, where the GPU handles the heavy thinking and the LPX handles the fast responses, the part that actually matters when you're talking to an agent in real time. In one benchmark it hit 3,400 tokens a second, four times faster than the next best option. Nebius, CoreWeave, and SpaceXAI are already signed on as customers.
This matters because the AI hardware race isn't just build a bigger GPU anymore. Agents need to respond fast, not just think hard, and that takes different silicon than training does.
My take, Nvidia's not just winning the AI hardware race, they bought the finish line. First they owned training, now they're moving in on inference and agents too. Everybody else better find a different race to run.
Put it together and here's the pattern. One company's AI broke containment and hacked somebody, and the industry's response was to write bigger checks and ship more chips. Move fast, worry about the guardrails later, and hope the attorney general doesn't ask too many hard questions in the meantime.