Welcome to Eye on AI. Emily Forlini here. In today’s issue:
- Nvidia, Meta, OpenAI, Microsoft, and others publish an open letter about open source
- Trump administration nears release of AI framework
- OpenAI president and co-founder Greg Brockman sits down with Fortune
- Taiwan detains Nvidia employee allegedly smuggling chips to China
- Growing calls for OpenAI to disclose details of Hugging Face hack
First, let me introduce myself. My name is Emily Forlini and I am new to Fortune, covering AI alongside Jeremy Kahn and Beatrice Nolan. What an elite crew I’ve been lucky to join! My first few weeks have been a whirlwind, from breaking the story that OpenAI hacked Hugging Face, to meeting OpenAI president and co-founder Greg Brockman at the company’s New York City office.
This week, the main talk in the AI industry is the Trump administration potentially banning Chinese open source models, a move it’s reportedly mulling with a small group of advisors. The idea is to prevent a crippling cybersecurity incident, and to keep the U.S.’s leading edge. These longstanding concerns became heightened after Beijing-based Moonshot AI released its Kimi K3 model this month, rivaling the best American models while being widely available to download for free.
AI leaders are sparring with the government on this point, with multiple top figures speaking out against a potential ban. When we met with Brockman on July 23, he made it clear he did not support the ban, though he fell short of directly saying that. Instead, when journalists peppered him with questions about the topic, he reiterated phrases such as, “We [at OpenAI] really believe in access,” and that “having more models is a good thing.” Brockman became a top Trump donor in January 2026 with a $26 million gift to the MAGA Inc. super PAC.
The topic blew up on social media a day later on July 24 when Nvidia CEO Jensen Huang posted for the first time ever on X to strike down a potential ban. He shared an open letter titled “Open Weights and American AI Leadership,” signed by Meta, Perplexity, Microsoft, and others. “Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty,” Huang wrote.
Microsoft posted the same letter on its website, with an expanded list of signatories, including OpenAI.
The following day, I spoke with Michele Catasta, the president and head of AI at Replit, which also signed the letter. Catasta said such a ban would “create an awful precedent,” and thought Chinese models could “apply some price pressure, which is something we would all welcome in this space, given how much money every company is spending on AI.”
It seemed like everyone I spoke to was against the U.S. government limiting access to open source models, including the powerful ones coming out of China. But there was one big gap in the conversation: Where was Anthropic?
Anthropic did not sign the letter. After facing pressure to weigh in, late last night CEO Dario Amodei published a blog post on the matter. Amodei reiterated his long standing position that he does not support banning open source models, a move he called protectionist and ineffective.
However, he subtly undermined the pro open-source letter in part by saying he was more concerned about other areas: Limiting the export of chips to China, preventing Chinese companies from distilling U.S. models, and implementing stronger safety testing on all models, open and closed.
Amodei also directly disagreed with one key point in the letter related to cybersecurity.
“I don’t agree with the letter’s assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers,” Amodei said. “It seems at least as likely to me that the opposite will be true.”
Both the Trump and Biden administrations have wrestled with how to deal with the serious, legitimate cybersecurity concerns posed by advanced AI models in the hands of a capable adversary like China. The question is whether banning open source models makes any progress on that issue, or if it just makes it worse.
The real reasons Silicon Valley doesn’t want a ban
As much as I appreciate the discourse, I’m honestly not buying what any of these guys are saying. Moralizing talk about democracy, humanity, and American leadership, has circulated within AI for years. Sometimes, it comes off as performative.
What’s really beneath these comments? Some of it is likely genuine concern, but also everyone weighing in on the ban has a vested business interest in preventing it.
Nvidia stands to profit handsomely from open source AI in the long run. More capable and widely available models, means more customers for the Nvidia chips required to run the models. And as frontier labs such as OpenAI and Anthropic (both big Nvidia customers), look to wean off of their Nvidia dependence, including by building their own chips, open source models could become Huang’s main customers over time.
Companies like Replit, which offer services based on AI models but do not currently make their own models, benefit from open source models in a few ways. For one, they offer more control and customization. And secondly, they can be cheaper. Replit markets its AI platform as a hub for creatives, not just vibe coders, and from the company’s early testing, Kimi K3 is “pretty exceptional” at design, Catasta told me.
When it comes to Brockman and OpenAI, they may be concerned about the federal government intervening in the AI industry, dictating which models companies can offer and which ones they cannot. OpenAI had to work “very closely with the government” prior to rolling out GPT-5.6 Sol in late June, Brockman said during a roundtable conversation with the press. That was a few weeks after the Trump administration effectively forced Anthropic’s Fable 5 and Mythos 5 models off the market in June, citing security concerns.
Those kinds of incidents slow companies down and create an unpredictable bureaucratic layer, costing more time, effort, and money, per model release. The Trump administration is in the process of finalizing a framework for AI companies to submit their advanced models to the government before public release, The Information reported today.
So as usual, the public is left grappling with mixed messages on a topic that’s unlikely to be resolved in the short term. This debate will most likely continue on for years, if not decades. Buckle up.
With that, here’s more AI news.
Emily Forlini
[email protected]
@emilyforlini
Before we get to the news, just a reminder to check out our new vodcast, Fortune AI Weekly on YouTube. Also: Fortune‘s Global 500 list also debuted today, with Amazon topping it for the first time.
FORTUNE ON AI
China’s Moonshot, Z.AI, and DeepSeek are challenging U.S. AI labs—and beating them on cost—by Nicholas Gordon
AI IN THE NEWS
Trump Administration nears completion of AI framework. As questions about the future of open source AI loom, the U.S. government is set to release a process for AI labs to submit their most advanced models for review prior to releasing them to the public. Although the framework will technically be voluntary, it remains to be seen how the administration enforces it. More in The Information.
Nvidia invests $5 billion in former OpenAI chief scientist Ilya Sutskever’s company. The partnership gives the secretive company, Safe Superintelligence, access to ten-times more compute, according to its X post on the deal. For Nvidia, it’s a strategic move to shore up a long-term customer. More in Bloomberg.
Taiwan detains Nvidia employee allegedly smuggling chips to China. The employee is allegedly involved in illegal exports of Super Micro AI servers equipped with Nvidia chips subject to U.S. export controls. The suspect has not been fully named, aside from his surname, Chang. Authorities searched Chang’s home and detained him after prosecutors argued there was a risk he could flee, destroy evidence or collude with accomplices or witnesses. More in Reuters.
EYE ON AI RESEARCH
All AI models are cheaters. The UK-based AI Security Institute (AISI) says every model it has ever tested has attempted to cheat on cybersecurity evaluations. Coincidentally, UK AISI posted its findings on the same day OpenAI confirmed its models hacked Hugging Face in an attempt to cheat on internal evaluations. This adds to growing concerns that humans are unable to fully control AI models, and the impact of their cheating will get more concerning as their capabilities increase.
Read more from UK AISI here.
AI CALENDAR
Aug. 4-6: Ai4 2026, Las Vegas.
Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.
Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.
Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.
BRAIN FOOD
Is the AI industry moving at an unhealthily fast pace? This week, another C-suite AI executive stepped down citing health concerns, the second to do so in the past month. Lilian Weng, co-founder of Thinking Machines, said she cannot “continue at a pace a startup requires.” She’s been more sick in the past seven months than ever before, and she’s decided “the consistent stress and workload have pushed me beyond what my health can sustain physically.”
Weng is a former VP of research at OpenAI, who left the company, along with Mira Murati, OpenAI’s former CTO, to found Thinking Machines. Weng joins yet another OpenAI veteran, Fidji Simo, who left her role as the company’s head of product in early July. Simo cited her longstanding health concerns, specifically a fierce relapse with the chronic autoimmune condition POTS. “It’s now clear that I’ve pushed a little too far and I really need to try new interventions to stabilize my health,” she told employees in April.
While these are specific cases that are not indicative of the entire industry, we know that AI startups are working at a breakneck pace. The industry has embraced China’s “996” schedule, or working from 9am to 9pm, six days a week, over the past few years. Are employees starting to reach a breaking point?
