Welcome back! Anthropic told the world it's building its own chips, which is a polite way of saying it wants out from under Nvidia. Demis Hassabis handed off the day job at Google DeepMind to go chase AGI, and Alphabet dropped about 4% while the people who built Gemini headed for the door. Microsoft told its own engineers to stop using Microsoft's coding model and use OpenAI's instead. And Ilya Sutskever, after two silent years and billions of dollars, is reportedly about to launch a model.
In today's Generative AI Newsletter:
Anthropic: What is it building to get off Nvidia?
Google: Is AGI closer than expected?
Microsoft: Why pick ChatGPT over its own model?
Ilya: What did two silent years buy?
Anthropic is after Nvidia
Anthropic confirmed it's hiring a team to design its own chips, the first time it has said so out loud.
It wants to build the hardware and the models together, so Claude runs faster and cheaper on silicon designed for it. The Information reported last month that Anthropic has been scouting Samsung to build them.
The custom chips are meant to sit alongside what it already buys from AWS, Google, Nvidia and AMD rather than replace any of it. Demand for Claude keeps climbing while every lab in the business races to lock up infrastructure.
OpenAI showed off its Broadcom-built Jalapeño inference chip in June, Google has run on its own TPUs for years, and Meta builds MTIA accelerators.
Anthropic still has to buy Nvidia chips for years either way. This is about what it pays for them after that.
Special highlight from out network
If your team builds with Claude, Cursor, or Codex, your compliance workflow should too.
Join this live session from Vanta to see how you can manage security and compliance directly inside the AI tools you already use—saving time, reducing context switching, and keeping your team moving.
You’ll learn how to:
Run compliance tasks directly from your AI workspace
Spot risks and audit gaps with simple prompts
Assign and track fixes without switching tools
Keep Vanta as your single source of truth
Reserve your spot today. Can’t attend live? Register anyway, and you’ll receive the recording.
Demis Hassabis handed off Gemini to chase AGI
Demis Hassabis is stepping back from running Google DeepMind to become its chairman and Alphabet's chief scientist.
He told employees on Wednesday that AGI is close at hand and called this a pivotal moment in human history, which is why he wants out of the daily job.
He co-founded DeepMind, arrived at Google when it bought the lab in 2014, and keeps running Isomorphic Labs, Alphabet's drug discovery arm.
Koray Kavukcuoglu takes over as SVP reporting to Pichai.
He's worked alongside Hassabis for more than 13 years, was DeepMind's CTO and Google's chief AI architect, and Pichai credits him with building the deep learning team and driving breakthroughs like WaveNet and DQN. He owns Gemini 4 now.
Alphabet fell about 4%. It went cash flow negative last quarter for the first time on record and is guiding to up to $205 billion of capital spending this year, while cloud revenue grew 82% to $24.8 billion.
So the man who built Google's AI now spends his days on AGI and medicine, and the people who built Gemini are leaving.
Microsoft told its engineers to use OpenAI's model
Jay Parikh, who runs Microsoft's CoreAI group, told employees this week to default to OpenAI's GPT-5.6 Sol inside GitHub Copilot and use it most of the time.
Microsoft has built its own coding model and sells access to more than 11,000 models through its cloud, Anthropic's included.
He told staff that anyone with a big project needing a lot of tokens should go have a quick chat with their manager first.
Microsoft's cash generation fell 23% from a year ago. Capital spending across Microsoft, Amazon, Alphabet and Meta is expected to clear $700 billion this year.
The era of burning tokens without thinking about it is over at Microsoft. Yours is probably next.
Ilya Sutskever is about to show what he's been building
Safe Superintelligence is reportedly shipping its first model this month, which would be the first thing it has ever released.
If the name means nothing to you, Sutskever co-created AlexNet with Alex Krizhevsky and Geoffrey Hinton, the work that proved you could stack neural networks on GPUs and get something extraordinary.
Most of what you use today traces back to it. He then co-founded OpenAI, ran its Superalignment team, and left months after the failed attempt to remove Sam Altman.
He started SSI in 2024.
SSI raised $3 billion and hit a $32 billion valuation with nothing to sell. Nvidia added a stake last month that Bloomberg put at $5 billion, along with Vera Rubin chips expected to multiply its compute tenfold.
The August claim comes from investor Gavin Baker on a podcast. SSI itself has announced nothing.
Two years without shipping is a long time to ask people to wait. We find out this month whether it was worth it.
Tool of the Day: Google AI Studio
Google AI Studio is Google's free playground for its own models, and it hands you the dials the Gemini app hides. Text, image generation with Nano Banana and Imagen, video with Veo, speech and music, and real-time voice and video all sit on one page. You get system instructions, temperature and a thinking level control. The part worth the visit is the Get code button, which turns whatever you just built into working API code.
Try this yourself:
Open a new prompt and pick Gemini 3 Flash Preview from the model dropdown.
Write a system instruction describing exactly how you want it to answer, then run the same question with and without it.
Turn on Grounding with Google Search and ask about something that happened this week.
Hit Get code and paste the result into your editor to see what your prompt looks like as an API call.
Who it's for: anyone who has hit the ceiling of the Gemini app and wants the controls.
Everything else you shouldn't miss
OpenAI published its own account of two more escapes: Its models reached the live internet during third-party cyber evaluations by the UK's AI Security Institute and Irregular, and the post walks through how each one was contained.
Jeff Dean's new company wants to automate science itself: Discovery Loop, founded with Sanjay Ghemawat, Oriol Vinyals and Quoc Le, plans to automate machine learning experiments end to end, with Radical Ventures and Khosla leading the seed and Alphabet staying on as cloud partner.
Anthropic's reported $10 billion compute deal has Nvidia on both ends: Volta Infra, a startup founded this year that Nvidia and Michael Dell put money into, will fill a Norwegian data center with Nvidia chips for Anthropic through March 2027.
Nvidia opened its self-driving model for commercial use: Alpamayo 2 Super now ships under a license that lets carmakers deploy and modify it, and it ranks first among nearly 40 models on the LingoQA driving benchmark.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 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.
