Mark Zuckerberg built an empire on the promise that Facebook and Instagram would always be free. Now he's flipping the script.

Meta is reportedly launching Hatch, a consumer AI agent that can complete tasks on your behalf, within the next few weeks. A premium tier could run up to $200 a month, on par with OpenAI and Anthropic's priciest plans. Its next flagship model, arriving in October, has an equally serious code name: Watermelon.

The bird hasn't hatched. The fruit hasn't left the vine. Somehow both are about to run your calendar for $200 a month.

Here’s what happened in AI today:

😺 NVIDIA's newest AI chip is heading into orbit for Elon Musk's SpaceX.
📰 Anthropic's flagship AI model struggled to attract corporate spending despite its power.
📰 ICE banned staff from wearing Meta's AI smart glasses on duty.
📰 Chinese state-backed hackers doubled attack volume after adopting DeepSeek AI.
📰 Mistral struck a partnership with Saudi Arabia to build AI models.

😺 NVIDIA's Newest AI Chip Just Got a Ticket to Space, Courtesy of Elon Musk

AI agents (bots that can browse the web, write code, or book your flights on their own) need a traffic cop coordinating the work behind the scenes. That traffic cop is a boring, unglamorous chip called a CPU.

NVIDIA just landed its highest-profile customer yet for the CPU it built for that job. SpaceXAI (Elon Musk's AI arm, which powers the chatbot Grok) is deploying NVIDIA's new Vera CPU to run its next generation of AI agents, and taking the same chip architecture into orbit.

Here's what happened:

SpaceXAI will use Vera (a chip built to handle the busywork around AI models, like running code and coordinating tools) to speed up Grok's agentic systems.
SpaceXAI is also expanding its AI infrastructure with NVIDIA's Vera Rubin platform as it scales toward gigawatt-sized data centers.
Its first AI satellite, named Starmind, will run an optimized version of that same hardware, in actual orbit.

The specs, for the curious:

88 NVIDIA-designed "Olympus" cores
Up to 1.2 terabytes per second of memory bandwidth (how fast data moves in and out of the chip)
Up to 1.8x faster agentic task completion than traditional x86 chips, the standard Intel and AMD have sold for decades

Why this matters: Most of the AI spotlight goes to GPUs, the chips that do the heavy "thinking." But agents don't just think, they act: running code, checking databases, calling tools between every response. If the CPU handling that traffic can't keep up, an expensive GPU sits idle waiting for instructions.

That bottleneck is why NVIDIA built Vera, and why it's now selling standalone CPUs for the first time ever, chasing a $200 billion market it's never touched, per its own CFO. Intel and AMD have actually outperformed NVIDIA's stock in 2026 on bets that agentic AI keeps growing. If SpaceXAI's satellite plan pans out, "AI infrastructure" stops meaning just data centers on Earth.

Our take: A chip running an AI satellite sounds like a marketing flex more than a near-term product. But it's a real signal: Musk's companies want to own AI's whole stack, from the chatbot to the chip to the orbit it runs in. The open question is whether space-based AI compute becomes a genuine edge, or just an expensive science project.

Either way, NVIDIA just found a very literal way to say its chips are going to the moon.

FROM OUR PARTNERS

Most companies track the AI they approved. They're missing everything else: local LLMs on developer laptops, AI buried in SaaS tools, MCP servers quietly connecting models to internal systems. None of it shows up in a security review. All of it carries real risk.

Airia's free Shadow AI Risk Assessment scores your organization across 12 dimensions and delivers a personalized risk profile with a step-by-step action plan.

Discover what's running without your approval.
Understand your exposure across 12 risk dimensions.
Get a clear roadmap to close the gaps fast.

Take the free assessment now.

🎓 AI Skill of the Day: Make Claude Explain It With Big Pictures

Thariq says people at Anthropic have been using an /eli5 skill to understand a concept before diving into the details. It asks Claude to explain a topic as if you know nothing about it, using an HTML artifact with big pictures and very few words.

Try it on:

/eli5 how does this module work
/eli5 why did we make this tradeoff
/eli5 what caused this incident

Thariq says the point is not merely shorter output. It’s useful for explainers and building understanding before you attack the problem itself.

Install the community plugin:

claude plugin marketplace add anthropics/claude-plugins-community
claude plugin install eli5@claude-community

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

See how AI delivers fast, secure, personalized experiences

Watch Fin and Plaid now on demand to see how AI can resolve issues directly in customer conversations. This includes everything from reducing bank-linking friction, enabling proactive support, and helping customers complete real financial tasks.

Watch now

📰 Around the Horn

Anthropic's flagship model Fable 5 pulled in just 11% of corporate AI spending two months after launch, as businesses defaulted to cheaper tools like OpenAI's GPT-5.6, per data from 70,000 companies.
ICE banned staff from wearing Meta's Ray-Ban smart glasses on the job, after agents were repeatedly spotted wearing them during immigration raids.
Chinese state-backed hackers more than doubled their attack volume after folding DeepSeek into malware development and reconnaissance, researchers said.
Mistral struck a hundreds-of-millions-euro deal with Saudi Arabia's HUMAIN to build sovereign AI models for the Middle East.
OpenAI brought its GPT-5.6 models to AWS's Kiro coding tool, cutting task costs by roughly 82% in testing.

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

*Build production-ready AI agents with Google Cloud. Join the free 4-part Startup School technical series starting September 15.
Atlaso gives Claude Code, Cursor, and Codex one shared memory, so what you tell one tool, the others already know—free to try, then $10/month for Pro.
Viktor lives in Slack and Teams and finishes tasks itself, like auditing your ad spend and handing you the finished PDF instead of just a plan—free trial ($100 in credits), then $50/month.
AgentSky gives you one API key to run Claude Code, Codex, and other coding agents in the cloud, billed only for what you use—free to try ($3 credit), then pay-per-use.
Bolcho builds voice and chat agents that speak real Hindi, Tamil, and other Indian languages for your phone lines and website—free to try, then billed near provider cost per call.
port22 puts your Claude Code or Codex session on your phone, so you can read along and approve from the couch instead of your desk—free for one Mac, then $19.99/year for unlimited.
NudgeForMe scans your sent emails for proposals and pitches that never got a reply, then drafts the follow-up for you to send—free to try, then $39/month for Pro.

📖 Tuesday Tech Tip

Okay, so the “run AI locally” crowd just got some new ammunition. A developer testing Qwen locally says a sharpened Qwen3.8-27B running inside the Pi coding agent beat Claude Opus 5 High on the current slice of SWE-bench-Live, a benchmark built from recently published real software bugs.

The result is community-run and still in progress, so this is not “Qwen > Claude.” But the model card shows quantized builds around 18–23GB. Quantization shrinks a model so it needs less memory, usually with some quality tradeoff, which puts these versions within reach of a 24GB-class GPU.

Your gaming PC would like to discuss a promotion.

Speaking of: Berkeley PhD Shuo Yang says FreeToken can run official full model checkpoints without extreme quantization (paper). A checkpoint is the model’s actual learned weights, so the pitch is keeping more of the original model intact while still running it on consumer hardware.

His demo put Qwen3.6-35B at 39 tokens per second on an 8GB RTX 4060 laptop and DeepSeek-V4-Flash at 22–25 tokens per second on an RTX 5090 desktop. FreeToken uses bandwidth-aware CPU/GPU execution plus caching across agent turns, and its authors report 3–4× faster token generation than Ollama with coding-agent harnesses built in.

The practical shift: local AI is starting to look less like “run a weaker model for privacy” and more like a real cost, speed, and control option on hardware people already own.