Four stories. Real news. Zero fluff.
MARK ZUCKERBERG HAS HAD ENOUGH OF PAYING NVIDIA AND IS NOW MAKING HIS OWN CHIP
Here is a sentence you would not have predicted five years ago: Mark Zuckerberg is now in the semiconductor business.
Okay, technically he has been heading that direction for a while, but this week it became real. According to a report from Reuters confirmed by multiple sources, Meta’s custom AI chip, code-named Iris, will go into production in September 2026. TSMC is handling fabrication. Broadcom helped with the design. And the testing phase wrapped up in just six weeks with no major issues, which in the chip world is roughly the equivalent of passing your driver’s test on the first try after parallel parking a school bus.
Why does this matter? Because right now, Meta is spending enormous sums of money buying Nvidia chips. We are talking about a company that has publicly committed to spending up to $145 billion on AI infrastructure in 2026 alone. Jensen Huang is basically a silent partner in Meta’s AI ambitions, and the bill is astronomical. Iris is Meta’s attempt to build its own power source instead of renting one indefinitely.
The goal is 7 gigawatts of computing capacity by the end of this year and 14 gigawatts by 2027. For scale, that is roughly double what the company has today, achieved through a mix of Nvidia hardware and their own custom silicon. Iris is not designed to replace Nvidia GPUs entirely. It is more like Meta wants Nvidia for the heavy lifting and Iris for the specific AI workloads where a custom chip can run more efficiently and cheaply.
But read between the lines here. Every tech giant that starts building its own chips eventually starts reducing how much they buy from the outside. Apple did it with the M-series. Amazon did it with Graviton. Google did it with TPUs. The pattern is always the same: you start as a customer, you become more self-sufficient, and eventually you barely need the vendor at all.
Zuckerberg has also been publicly burned recently. He told his own staff on July 2 that Meta’s AI agent work has not accelerated as fast as he expected. The restructuring that cut 8,000 jobs and reassigned 7,000 others is still settling. The company is betting everything on AI being the future, and the last thing he wants is for that future to be bottlenecked by someone else’s hardware supply chain and someone else’s production schedule.
If Iris works as planned, Meta walks into 2027 with the ability to scale AI compute without being at the mercy of Nvidia’s pricing or shipping delays. That is worth a lot more than it sounds. The AI race is increasingly a hardware race, and right now Nvidia holds the keys to most of the kingdom. Meta just started building a back door. Jensen Huang will have noticed.
OPENAI LOSES ANOTHER EXECUTIVE AS ITS NO. 2 STEPS DOWN DUE TO ILLNESS AND GREG BROCKMAN TAKES THE WHEEL AGAIN
Here is the thing about OpenAI: they keep losing people, and they keep saying it is fine.
Fidji Simo, who served as OpenAI’s head of applications and effectively the company’s No. 2 on the product and business side, announced this week that she is stepping down to focus on her health. The reason is real and human and not corporate drama: she has been battling Postural Orthostatic Tachycardia Syndrome, known as POTS, a chronic condition that affects the nervous system and makes everyday functioning genuinely difficult. She stepped away in April for a medical leave and has now made it official. She is transitioning to a part-time advisor role.
Simo joined OpenAI in May 2025 from Instacart, where she was CEO. Before that she spent over a decade at Facebook, eventually running the app itself. OpenAI did not hire her to shuffle papers. She was brought in to turn the company’s products into a real, scalable business at a moment when the road to an IPO was starting to get real. OpenAI President Greg Brockman, who just returned from his own extended leave, has taken over her responsibilities in the interim.
The optics are not great, and the honest thing to say is that this is part of a pattern. OpenAI has seen significant leadership turnover. The list includes Ilya Sutskever leaving after the Sam Altman drama, Mira Murati stepping down, and a string of senior researchers heading for the exit. Every time it happens, the company says it is fine, focused, and building toward the future. And in terms of shipping products, that is actually true. GPT-5.6 launched this week. ChatGPT Work is live. The numbers keep growing.
But there is a difference between a company that can ship product and a company with deep institutional stability at the top. OpenAI is very much executing right now, but the leadership bench has been reshuffled so many times that it is hard to know who the permanent team actually is. Sam Altman is the constant. Everyone else seems to be on some kind of rotation.
Simo’s situation is clearly not her choice, and nobody with any decency is going to hold a serious health issue against her. But it does leave a gap at exactly the moment when OpenAI needs to convert its technical lead into a durable business. Sam Altman will be fine. He always is. The question is who he builds the next chapter with, and whether anyone stays long enough to build it.
NEWSOM SIGNS THE BIGGEST GOVERNMENT AI DEAL IN US HISTORY AND NOW EVERY CALIFORNIA BUREAUCRAT HAS CLAUDE ON THEIR DESK
On June 29, California Governor Gavin Newsom announced a deal that, depending on your politics, is either a visionary bet on the future of public services or a very large government contract for an AI company headquartered in San Francisco that has donated extensively to California politicians. Both things can be true at the same time.
The agreement gives every California state agency access to Anthropic’s Claude at a 50% discount, channeled through the state’s technology portal. Every city, county, and local government in California is also invited to join the deal at the same price. Free workforce training and technical assistance from Anthropic engineers is included. The California DMV is already using it to improve customer service and reduce wait times. The Department of Healthcare Services, which runs the largest Medicaid program in the United States, is using it for internal workflows. And the state has made Claude the first AI productivity tool officially available through its centralized Statewide Information Technology Shared Services portal.
In scale, this is genuinely historic. California has roughly 230,000 state employees across hundreds of agencies. If even a third of them use Claude in any meaningful way, you are looking at one of the largest enterprise AI deployments in history, run by a government rather than a corporation.
Anthropic’s angle is obvious and they are not hiding it. Government contracts are sticky, long-lasting, and extremely valuable as reference customers when you are selling to everyone else. When Anthropic walks into a meeting with a hospital system or a federal agency and says it runs AI for California’s Medicaid program, that is the kind of case study that closes deals. This is not charity. It is business development at the highest possible level with a built-in press release.
Newsom’s angle is also obvious. He has been building the narrative that California can lead on AI innovation while also being the state that takes AI risk seriously. He signed the responsible AI framework earlier this year, he is hosting tech summits, and he is positioning himself for whatever comes next politically. A deal this size generates attention and frames him as someone who gets the moment.
The cynics will note that Anthropic got a sweetheart deal in a state it already dominated without a competitive bidding process that would have taken three years. The optimists will say if AI can help the DMV answer questions faster and help Medicaid workers move more efficiently, the California taxpayer wins. Both camps have a point. The check is already signed and Claude is already answering questions inside state government. Now we get to find out whether any of it actually works.
THE AI REVOLUTION HAS A VELVET ROPE AND MOST PEOPLE ARE STILL STANDING OUTSIDE IN THE RAIN
Axios published a piece this week with the kind of sentence that sounds obvious once you read it but that nobody had quite said out loud this clearly before: AI has a class divide, and it is getting wider.
Here is the situation. The top of the AI pyramid right now is occupied by models like Claude Fable 5 and OpenAI’s Sol, which can run multi-hour autonomous research and coding sessions with minimal supervision. These systems are genuinely capable and transformative for the people using them. Lawyers, engineers, researchers, financial analysts, people who know exactly what to ask for and can afford to pay for access, are already running laps around their competitors using these tools every day.
But here is the problem. These frontier models are expensive. Sol is an enterprise product. Fable 5 costs real money that most individuals and small businesses cannot justify spending on a productivity tool. And even in cases where access is nominally free or discounted, the learning curve is steep enough that most people never get past the surface. They use ChatGPT to write a cover letter. They generate an image for a birthday card. They treat the whole thing like a slightly smarter search engine.
Meanwhile, the people who have figured it out are using AI to do the work of entire teams. Writing code, building products, analyzing markets, generating client reports, running customer service at a fraction of what it used to cost. The productivity gap between a serious AI user and a casual one is not a 10% improvement. It is beginning to look like an order of magnitude, and that gap compounds over time.
This is the uncomfortable thing that the AI industry does not spend much time discussing. All the talk about AI democratizing access to knowledge and tools comes with a catch attached: you have to actually know how to use the tools. And the people who know how to use them are, by and large, already highly educated, already working in high-paying industries, and already ahead of everyone else.
None of this means the technology cannot spread more widely. It does spread, just unevenly and slower than the press releases suggest. And the free tiers are not useless. But the honest version of the AI access story in July 2026 is that the frontier is currently reserved for people who can pay for it, work at a company that pays for it, or were lucky and curious enough to figure it out on their own.
Which brings it back to deals like the California and Anthropic partnership covered above. Giving 230,000 state workers subsidized access to a frontier model is not just a productivity play. It is, whether Newsom thought of it this way or not, a form of AI redistribution. Whether it actually moves the needle is still an open question. But at least someone is unlocking the door. Most people are still outside waiting for an invitation that has not arrived yet.