Welcome back! The White House pulled the top AI labs into a room and told them it wants to see their newest models up to 90 days before anyone else. That is a long time in this business. Anthropic, meanwhile, signed a giant compute deal with Elon Musk's SpaceX and is even floating the idea of data centers in space. A Pizza Hut franchisee is suing for $100 million and blaming an AI delivery system for its collapsing sales. And we kept our promise from yesterday's Atlas, so today you get the guide on running Claude agents inside your own network.
In today's Generative AI Newsletter:
White House review: Why does Washington want a 90-day head start on every major AI model?
Anthropic x SpaceX: What is Anthropic paying for, and why is space part of the deal?
Pizza Hut lawsuit: How did one AI system turn a profit into a $100 million claim?
Claude agents guide: How do you run Claude agents inside your own network without the risk?
The White House's Office of the National Cyber Director pulled the top AI labs into a briefing this week about a coming executive order. The order would let intelligence and other government agencies review advanced AI models before they ship.
OpenAI, Anthropic and Reflection AI were all in the room on Tuesday, and Trump could sign it as soon as today.
The plan is voluntary for now. Labs would tell the government before a major release and could hand over their newest models up to 90 days ahead of launch.
Anthropic declined to comment.
OpenAI and Reflection AI did not respond.
Ninety days is a long time when new models land every few weeks. Hand them over that early and the government can scrutinize, slow or block them before the public ever logs in.
That puts the White House in control of the release calendar if Trump makes it a law.
Special highlight from our network
Anthropic ships something new almost every week. Cowork, Skills, Connectors, Design. Most people are still using Claude like it's just a chatbot.
The first Claude-a-thon is a 2-day live workshop on what Claude can actually do in 2026. You'll run real research, build artifacts and dashboards, design full presentations, and set up Connectors that automate things like job searches and inbox triage.
Saturday and Sunday, 10 AM to 7 PM EST. Free for the next 48 hours, 1,000 seats only.
Special highlight from our network
Every swipe and scroll feeds billion-dollar models, while most people struggle to make ends meet. Feels unfair, right?
Mode Mobile lets you earn from what you already do (music, games, shopping, browsing, etc).
They're Deloitte's #1 fastest-growing software company, with 490M+ users earning $1B+ combined.
Share price changes on May 29, might be worth a look before then.
Disclosures
Please read the offering circular and related risks at invest.modemobile.com. This is a paid advertisement for Mode Mobile’s Regulation A+ Offering.
Mode Mobile recently received their ticker reservation with Nasdaq ($MODE), indicating an intent to IPO in the next 24 months. An intent to IPO is no guarantee that an actual IPO will occur.
The Deloitte rankings are based on submitted applications and public company database research, with winners selected based on their fiscal-year revenue growth percentage over a three-year period.
Anthropic is paying nearly $45 billion over 3 years to rent computing power from Elon Musk's SpaceX. The company has been rationing Claude for months and this is how it plans to stop.
The deal hands Anthropic more than 300 megawatts of new capacity, over 220,000 Nvidia GPUs, inside a month, all of it from SpaceX's Colossus One data center.
Btw, SpaceX bought xAI back in January, so this is Musk selling compute to a direct rival. The two are even talking about putting data centers in space, what SpaceX calls multiple gigawatts of orbital compute.
Anthropic already turned a profit last quarter at a $44 billion run rate while starved for chips. Feed it enough compute and that number has plenty of room to climb.
A Pizza Hut franchisee called Chaac Pizza Northeast, which runs more than 100 stores, is suing the chain for $100 million over an AI system.
After Pizza Hut rolled out its Dragontail delivery tool, deliveries that used to take under 30 minutes started taking 45 or more, and their sales swung from double-digit growth to -9.78%.
The old way: Chaac's managers sat between Doordash and the kitchen, releasing an order only when it was truly ready and blocking badly rated drivers.
What changed: Dragontail handed Doordash everything up front, the tip, whether it was cash and what else was cooking nearby.
What drivers did: They sat on ready orders for up to 15 minutes to grab several at once and skipped the low-tip and non-cash ones.
Dragontail worked as it gave gig drivers more information and they used it to serve themselves. Automate a process you do not fully understand and you can lose the thing that made it work.
In yesterday's Atlas we promised to show you how to set up self-hosted sandboxes for Claude agents. Here you go.
Anthropic now lets a Claude agent run its actual work on a machine you own, so your code, files and network traffic never leave your environment.
Normally Claude's Managed Agents run their tools in Anthropic's cloud. A self-hosted sandbox keeps the orchestration with Anthropic and moves the execution to your infrastructure, which matters when the agent touches data that cannot leave your network.
The bridge is a small program called an environment worker.
Setting it up with the ant CLI takes five steps:
In the Claude Console, open Workspace > Environments > New > Self-hosted to create the environment.
Click Generate environment key, then set the key and environment ID as variables on your worker machine.
Install the ant CLI on that machine.
Run ant beta:worker poll. It claims any session for your environment, runs the tools in your /workspace folder and sends results back.
Create a session that points at your environment and the worker picks it up.
Pair it with an MCP tunnel and that same agent can reach the tools locked inside your private network (private APIs, internal databases, knowledge bases, ticketing systems, developer tools) over a single outbound connection, no inbound firewall rules and no exposed endpoints.
Self-hosted sandboxes are in public beta on the Claude Platform now, and MCP tunnels are in research preview by request. Start small. One environment, one worker, one agent pointed at a single internal tool.
Have you tested what breaks first: your voice AI or your customer's patience?
In noisy real-world environments (contact centers, live events, busy streets), most transcription AIs fall apart.
Google manages 16% accuracy in contact centers. Speechmatics hits 71% in the same messy conditions, 89% on live football commentary.
The difference is that Speechmatics was built for real human environments. Who cares about results from soundproof labs?
If you’re building anything voice-related, add Speechmatics to your stack.
Antigravity unlocked: Google tripled rate limits across every Antigravity tier so you can push Gemini 3.5 Flash harder.
Math solved: OpenAI says one of its internal models cracked a famous conjecture that stumped humans for nearly 80 years.
Tech layoffs: Tech companies have cut more than 138,980 jobs so far this year.
Built by agents: Google Antigravity used Gemini 3.5 Flash to build an operating system from one prompt, with 93 subagents, 15,314 model calls and billions of tokens.
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.
