Why the open-weight vs closed AI fight just became a problem for Nvidia

Nvidia just spent $100 billion locking OpenAI into its hardware roadmap. And in the same month, OpenAI got pulled into a public fight over whether models like the ones Nvidia's chips run should even be closed off in the first place. That's not a contradiction — it's the actual state of the AI industry right now, and it's messier than most headlines make it sound.

If you've been half-following the open-weight versus closed-weight debate and wondering why it suddenly matters, here's the breakdown, plus a few genuinely useful product updates that shipped in the middle of all the drama.

What's really going on between OpenAI and Nvidia right now?

Start with the money. OpenAI and NVIDIA announced a letter of intent for a landmark strategic partnership to deploy at least 10 gigawatts of NVIDIA systems for OpenAI's next-generation AI infrastructure to train and run its next generation of models on the path to deploying superintelligence. To support this deployment including data center and power capacity, NVIDIA intends to invest up to $100 billion in OpenAI as the new NVIDIA systems are deployed. That's about as tight as two companies can get.

But zoom out to the policy side and the relationship looks a lot more complicated. In late July, open-weight AI models received public support when about 25 technology companies led by Nvidia and Microsoft issued an open letter urging U.S. policymakers to avoid premature or sweeping restrictions. The coalition's letter, known as Open Weights and American AI Leadership, included signatories such as Meta, Palantir, IBM, Andreessen Horowitz, Hugging Face, Mozilla, the Linux Foundation, and Dell.

Here's the twist: OpenAI and Anthropic did not join the letter when it first went out. That's notable because the group defined open-weight models as AI systems whose underlying parameters, or weights, can be downloaded, inspected, modified, and run on local infrastructure — which is exactly the opposite of how OpenAI's core business works. OpenAI sells access to models it keeps locked behind an API. Nvidia, on the other hand, benefits from every additional AI model that creates demand for its chips, regardless of whether that model is open or closed.

Then, less than 24 hours after the letter dropped, something shifted. Jensen Huang's "Open Weights" letter rapidly doubled its signatories to 50 in a single day, including OpenAI and Google. That's a genuinely strange move for a company whose entire revenue model depends on nobody being able to download and self-host its best models. Critics weren't shy about pointing out the awkwardness, either — observers noted a self-serving aspect, since many signatories, including organizer Nvidia, maintain proprietary layers elsewhere.

Why does this matter if you're not building your own AI lab?

Because the open-weight momentum is real and it's already changing what's possible on your own hardware. Chinese labs have been especially aggressive here — Chinese open-weight models such as Moonshot AI's Kimi K3, released in July 2026, have narrowed the performance gap with leading U.S. systems and intensified regulatory concern. Jensen Huang has been blunt about how big this shift already is: at Nvidia's CES 2026 press Q&A, he put a figure on the shift, saying one in every four tokens generated today comes from an open model.

That number matters because open weights don't just mean "free to download." They mean the compute doesn't have to run through a handful of API providers. Weights that anyone can download get served from enterprise clusters, regional clouds, and on-premises racks rather than a handful of hyperscaler API endpoints. For a company like OpenAI whose moat is closed access, that's a real threat. For Nvidia, whose moat is selling chips to literally everyone doing the serving, it's a feature.

Can you actually run good models at home now?

Yes, and the hardware options have gotten weirdly good this year. If you're shopping for a local AI box, the three real contenders right now are Apple's unified-memory Macs, AMD's Ryzen AI Max+ 395 (also called Strix Halo), and Nvidia's own DGX Spark.

Each one makes a different trade-off. The DGX Spark wins on tooling and raw AI compute — full CUDA and up to 200B-parameter models — but costs the most. Ryzen AI Max+ 395 boxes like the EVO-X2, Framework, and GTR9 hit 128GB unified memory for far less, and double as a full x86 desktop. And on the Apple side, price per gigabyte of fast unified RAM favors Strix Halo — $3,299 for 128GB works out to $25.77/GB versus the M3 Ultra at $3,999 for 96GB, or $41.66/GB.

The practical upshot: a $2,000–$3,300 mini PC can now run 70-billion-parameter class models at home without touching a cloud API. That's a direct challenge to the "you need a data center" assumption that's kept most AI compute flowing through Nvidia-powered clouds. If you want to size a machine to a specific model before buying one, Layer3Labs' local AI hardware guide is a solid starting point, and TerminalBytes' mini PC roundup breaks down the price-per-gigabyte math in more detail.

What did OpenAI just change about ChatGPT tasks?

Away from the chip politics, OpenAI shipped something genuinely useful for anyone using ChatGPT for actual work. You can now set up ChatGPT to react to events instead of just running on a schedule. Start at the official Scheduled tasks documentation to see the full feature set.

If you're a paid subscriber, you can now configure ChatGPT to trigger a prompt as things happen in Gmail, Slack and GitHub. OpenAI made this feature available to Plus, Pro, Business and Enterprise users. That's a real shift from "run this every day at 2pm" to "run this the moment something specific happens."

Here's what each integration actually watches for:

• Gmail tasks can respond to a new message or filter by sender or subject.
• Slack tasks can respond to new channel messages, and you need to add @ChatGPT to each monitored Slack channel.
• GitHub tasks can respond to supported pull request activity in an authorized github.com repository.

How do you actually set one of these up?

The setup flow is short. According to OpenAI's own help documentation, you: go to Settings > Apps and connect a supported Gmail, Slack, or GitHub account, open Work and describe the event and the action ChatGPT should take, then review the Trigger, Condition, and Prompt and complete any required authorization. Once it's live, you can go to Scheduled to review, edit, pause, or delete the task.

One nice touch: these aren't locked to your account anymore. You can share an eligible scheduled or event-triggered task so another person can review its instructions and create a separate copy — including active or paused event-triggered tasks. If you build a genuinely useful automation, like flagging urgent Slack threads or summarizing GitHub PR activity, you can hand it to a teammate instead of walking them through the setup from scratch.

Worth noting before you go wild connecting every app you own: security researchers have flagged that giving an agent standing access to Gmail, Slack, and GitHub simultaneously widens the attack surface if any single connected app gets compromised. Configure narrow, specific triggers rather than "watch everything and act on anything."

What is Codex Sites and why should you try it?

The other big product update ties back to the video's sponsor link, and it's worth using regardless. You can grab it directly from OpenAI's Codex download page.

OpenAI launched Sites for Codex on June 2, 2026, at a live event called "Intelligence at Work." The framing matters: the headline isn't that ChatGPT can now build websites — it's that ChatGPT can now build the internal tools knowledge workers actually need, like dashboards, project trackers, and lightweight workflow apps, and host them at a shareable URL with workplace authentication baked in, all from a prompt.

The adoption numbers back up why OpenAI is pushing this so hard. Codex now has more than 5 million weekly active users, and roughly 20% of those users are knowledge workers rather than developers — a segment growing three times faster than the developer segment. That's OpenAI quietly turning a coding tool into a general workplace platform, which fits the same pattern as the tasks update above: less "chat with an AI," more "let the AI run things in the background."

Is Time's 100 most influential AI list actually credible?

This one deserves a skeptical eye. TIME divided the 100 honorees into four categories: Leaders, Innovators, Shapers and Thinkers, pulling from executives, researchers, policymakers, and cultural figures.

But the list drew real criticism for who got left off. TIME's 2026 AI list left out Jensen Huang, Mark Zuckerberg, Sundar Pichai, Satya Nadella and Demis Hassabis — people who actually built the infrastructure and models defining this era — while including Paris Hilton and Joseph Gordon-Levitt. One critic summed it up bluntly, arguing the "list prioritizes engagement over accuracy."

TIME's own editor framed the year as messier than any list can capture. As Editor-in-Chief Sam Jacobs put it, the field has been "explosive and divisive." You can see the full 2026 TIME100 AI list and judge the omissions yourself.

What should you actually do with all of this?

None of these stories are isolated. Nvidia wants more models — open or closed — because every model needs chips. OpenAI wants to keep its closed models valuable while not looking hostile to the open-source movement gaining ground in Washington. And meanwhile, the actual hardware you can buy for your desk keeps getting good enough to make the "who controls the cloud" argument less relevant every quarter.

If you want a practical takeaway: try the ChatGPT event-triggered tasks this week if you're on a paid plan — connecting it to a single GitHub repo or one