Collaborate with us Welcome back! OpenAI built its own AI chip and it shows us they need to hire someone to name their products. It runs OpenAI's models for about half the cost, and OpenAI used its own AI to design it, the snake eating its own tail in silicon. Anthropic benched Dario Amodei for Trump . Google's flagship slipped to July, and ChatGPT got tuned to be more fun to talk to.
In today's Generative AI Newsletter: OpenAI: Why did OpenAI build its own chip, and what does it do to Nvidia?
OpenAI develops Jalapeño OpenAI built its own chip, called jalapeño, in collaboration with Broadcom. The new processor has one job, running OpenAI's models cheaply. Broadcom's chief says it does that for roughly half the cost of the Nvidia GPUs everyone rents, which for a company bleeding cash on compute is close to oxygen. OpenAI used its own models to help design the chip, so the AI built the hardware that runs the AI, and it did it in nine months.
It's also a shot at Nvidia, whose GPUs are both OpenAI's biggest cost and its biggest chokehold. Google and Amazon already make their own, so OpenAI joining means the companies that rent Nvidia's chips are turning into the companies that compete with it. Jalapeño is still in testing, with deployment slated for late this year, and it only handles inference, the running of models, so training still leans on Nvidia. The 50% number comes from Broadcom.
Owning the chip, the model and the product lets OpenAI squeeze cost out of every layer at once, and that's the whole game now.
This Tech Could Be a Star in AI’s Future Goldman Sachs projects data center power demand could eclipse supply by 2030. This demand is near impossible to meet with today’s data centers, since they take years to build. That’s why BluSky AI developed a faster way. Their SkyMod system deploys AI-ready data centers in months, not years. And it’s only the beginning.
BluSky AI has a portfolio of sites ready for deployment. Invest in BluSky AI as they scale their innovation across the U.S. Disclaimer: This is a paid advertisement for BluSky AI Regulation A offering. Please read the offering circular at https://invest.bluskyaidatacenters.com/ The AI differentiator other products can't rent OpenAI just spent nine months building a chip so it could stop renting the most important part of its stack. The companies that last in AI own the layer their rivals have to pay someone else for.
There’s no distribution platform built on actual winning data owned by the founders. A competitor can buy ads and chase reach for years and still not have a 14M-person AI audience or the record of which campaigns actually worked. GenAI Works owns the data behind how companies like Nvidia distribute their products to millions of people. We built that data into ToneUp, an enterprise platform that helps brands run AI go-to-market.
Now we’re welcoming more investors from $1,000 with up to 25% bonus shares, the kind of early seat that people feel proud claiming after a while. 6 days left at this price. Learn more and be one of the early investors In making an investment decision, investors must rely on their own examination of the issuer and the terms of the offering, including the merits and risks involved.
Works, Inc. has filed a Form C with the Securities and Exchange Commission in connection with its offering, a copy of which may be obtained here . Anthropic benched its own CEO A White House official reportedly called Dario Amodei a weirdo. So Anthropic stopped sending him to the technical talks, handing them to cofounder Tom Brown. Amodei, still the company's CEO, was off the negotiations that will write the rules. A week earlier the administration had branded Anthropic a national security threat and cut the rest of the world off from its models.
Now Trump says the threat was last week's problem. The ban situation is not resolved yet, but people are optimistic seeing Fable on Amazon Bedrock yesterday and expecting it to be back soon. Google's AI flagship delayed to July Google's flagship is late again. At I/O in May it promised Gemini 3.5 Pro, its next frontier model, by June, and the room groaned at even that wait.
June is now ending, the model has slipped to July, and Google still has not shown a single benchmark or price for it. July is the worst place to be late. OpenAI has GPT-5.6 aimed at either this week or mid-July, so Google's delayed flagship now could launch at the same time as its biggest rival's next model. A release that was supposed to reset the race has to walk into someone else's launch instead.
This is becoming Google's signature move. The research and the demos are there, and the headline model keeps sliding. Google invented the transformer that kicked all this off, and it still can't ship its own flagship on schedule. OpenAI's upgrade this week is a personality For a week the internet braced for GPT-5.6. What OpenAI shipped is a new version of GPT-5.5 Instant, the model most people actually use, tuned to be more fun to talk to.
It reads your intent better and handles fussy multi-step requests more reliably, with shopping and local tips that finally hang together. Paid users have it now, free users right behind. OpenAI is tuning its most-used model for how it feels to talk to, which tells you where the consumer race is heading, toward who you want to chat with as much as who tops a benchmark. A friendlier chatbot is a nicer experience and a stickier product, and engagement is the same metric that turned social media into what it is.
Worth knowing which one the tuning is really for. Tool of the Day: Obot MCP is the standard that lets AI assistants reach your outside tools. Obot is an open-source gateway that puts all of those connections behind one control panel. It comes with a catalog of ready servers like Slack, Notion, GitHub and Outlook, so your AI reaches them without credentials scattered across a dozen apps.
Try this yourself: Start free at obot.ai with the hosted version, no install or infrastructure required. Open the built-in catalog and pick a ready server like Slack, Notion or GitHub. Generate a connection URL and drop it into Claude Desktop, Cursor or VS Code, or chat with the tool right inside Obot. Set a policy so each teammate reaches only the tools they are cleared for, with every call logged.
Who it's for: teams plugging AI assistants into real company tools that want one gateway instead of a dozen loose wires. Everything else you shouldn't miss Japan's $2.3 trillion industrial plan puts chips and AI first : Prime Minister Sanae Takaichi's 14-year push spans 17 strategic sectors from defense to shipbuilding, with semiconductors and AI as the centerpiece and a goal to quintuple chip sales by 2040. Two more Gemini brains defect to Anthropic : Jonas Adler and Alexander Pritzel, both key contributors to Google's Gemini, are leaving for Anthropic, the latest in a steady exit of senior talent from the team. OpenAI says GPT-5 cracked a 3-year science mystery : Immunologist Derya Unutmaz fed GPT-5 Pro a T-cell puzzle his lab had chased since 2022, and the model predicted the experiment's result before he'd published it. Qualcomm wants its own way around Nvidia : It's in talks to buy AI-chip startup Tenstorrent for as much as $10 billion, a data-center bet aimed straight at Nvidia's lead.
Learn more about AI from the experts building it 📸 Follow us on Instagram for fast, visual AI updates in 30 seconds. 🐝 Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business. 📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators. 🐦 Follow us on X for breaking AI news and real-time industry updates. 🧠 Learn how to build your next AI application with practical resources and expert guidance. 🚀 Explore investment opportunities in the future of AI and join our community-backed growth journey.
Get in touch.