01OpenAI Traded the AGI Escape Hatch for an AWS Account
For five years, one paragraph in OpenAI's contract with Microsoft set the terms of the AI industry's future: if OpenAI's board declared the company had reached artificial general intelligence, Microsoft's commercial rights would expire. That paragraph is now gone.
OpenAI and Microsoft announced an amended agreement Monday that drops the AGI termination clause, ends Microsoft's exclusive cloud rights, and rewrites the revenue split between them. The Verge reported that Microsoft remains OpenAI's "primary cloud partner" and OpenAI products will ship first on Microsoft. But OpenAI can now sell to customers on Amazon Web Services, clearing a $50 billion compute deal that had been stuck behind Microsoft's veto.
The AGI clause had become the most cited contractual provision in AI. Researchers, regulators, and reporters treated it as a structural check on Microsoft's control over OpenAI's roadmap. The board that declared AGI achieved would have inherited the company's commercial rights alone. That definition fight is over before it ever started.
In exchange, Microsoft locks in a longer revenue-share. TechCrunch reported the new arrangement gives Microsoft more cash from OpenAI over a longer period than the original deal. Microsoft loses exclusivity but extends payments past the original term's expiration window.
The AWS opening matters now because OpenAI's compute appetite has outgrown what one cloud provider will commit to. TechCrunch reported the renegotiation cleared the legal risk Microsoft posed to OpenAI's pending $50 billion Amazon agreement. OpenAI's next compute order, and presumably the one after that, no longer needs Microsoft's sign-off.
What gets quieter is the AGI conversation itself. The clause forced both companies to keep a public definition of AGI on hand, because triggering it would unwind tens of billions in commercial rights. Without that trigger, neither side has a financial reason to declare the threshold met. The board-level decision that was supposed to reshape the industry has been removed from the contract entirely.
Microsoft kept the right of first refusal on OpenAI's products and the cloud relationship. For OpenAI, the prize was Amazon access and a path to fund the next training run. The guardrail the industry treated as a floor under safety governance is no longer enforceable through the partnership that wrote it.
0220+ Google VPs ask Pichai to drop classified Pentagon AI; OpenAI clears FedRAMP for federal sales
More than 600 Google employees signed a letter to Sundar Pichai this week. They asked the CEO to block the Pentagon from using Google's AI models for classified purposes, according to The Washington Post.
Organizers say many signatories work in DeepMind. The list includes more than 20 principals, directors, and vice presidents.
The signers represent a small fraction of Google's AI workforce, but the seniority changes the calculus. Public letters from 20 VPs are not silent attrition — they are management overhead. The letter does not demand Google exit defense work entirely; it asks Pichai to draw the line at classified use.
OpenAI's federal posture moved in the opposite direction. The company announced FedRAMP Moderate authorization for both ChatGPT Enterprise and the OpenAI API, opening secure procurement for U.S. federal agencies. FedRAMP Moderate covers controlled unclassified information used across most civilian and defense agencies, and is the threshold federal IT buyers check before signing.
The compliance certificate is not a contract, but it removes a procurement bottleneck. Agencies that wanted ChatGPT before now have a paper trail.
The two moves describe divergent commercial postures. Google is fielding internal opposition to a category of defense work it already bids for. The 2018 Project Maven walkout produced public pledges, then the AI Principles, then a steady drift back toward defense contracting.
OpenAI is removing friction in the procurement chain while Google's HR side absorbs cost. Federal AI buying decisions for fiscal 2027 are being scoped now. Procurement officers weigh model performance, but they also weigh whether a vendor will still be selling the product after employee pressure peaks.
OpenAI says the FedRAMP authorization positions ChatGPT Enterprise and the API for federal agency adoption; the company has not disclosed which agencies are in active deployment. Google has not publicly responded to the letter.
03The iPhone home screen, the Ubuntu kernel, the car design studio: AI is moving below the app layer
An iPhone app that replaces the home screen launcher itself raised funding this week before shipping. Skye, built by Signull Labs, doesn't add an AI button to the phone. It restructures the screen users see when they unlock it, surfacing apps, contacts, and actions before any tap. TechCrunch reported investors committed ahead of launch. iOS does not permit launcher replacement the way Android does, which makes a home-screen-replacement product technically aggressive on its own.
Canonical published a roadmap the same week for embedding AI into Ubuntu Linux. VP of engineering Jon Seager outlined twelve months of work to ship AI features through standard apt updates, touching the package manager, system configuration, and developer tooling. Ubuntu's user base spans tens of millions across desktop and server installations. None of them will install a separate AI client. The features will arrive when their machine pulls its next update.
In Detroit and Yokohama, the entry point shifted further upstream. The Verge reported that General Motors and Nissan now route early-stage car design through neural concept platforms. AI generates and iterates sketches before any human draws a 3D model. Designers used to spend weeks refining hand sketches from every angle before digitization. The new pipeline puts a generative model in the seat traditionally held by a junior designer with a pencil, compressing the front end of the studio process.
The three moves come from unrelated industries and overlap by accident, not coordination. What they share is the layer they target. None sits behind an icon the user chooses to tap. They run at the launcher, the package manager, and the CAD pipeline, where opting out means refusing the default product. The next iPhone unlock, Ubuntu apt update, or new vehicle on a 2027 lot will carry AI whether the user asked or not.

China blocks Meta's Manus acquisition Beijing forced the unwinding of Meta's deal to acquire Chinese AI startup Manus, citing US-China tech rivalry. The collapse shows how founders with China ties struggle to exit through US acquirers as both governments tighten review of cross-border AI deals. arstechnica.com
EU pushes Google to open Android to rival AI assistants The European Commission is preparing rules that would force Google to give third-party AI assistants the same Android access as Gemini. Google called the proposed intervention "unwarranted." Gemini currently receives default placement and system-level integration on Android phones. arstechnica.com
Ineffable Intelligence raises $1.1B for human-data-free AI David Silver, the DeepMind researcher behind AlphaGo, raised $1.1 billion at a $5.1 billion valuation for his months-old British lab. Ineffable Intelligence aims to train models through reinforcement learning without human-generated data, a direct bet against the prevailing pretraining-then-RLHF approach. techcrunch.com
GitHub Copilot moves to usage-based billing GitHub will charge Copilot customers based on consumption rather than flat per-seat pricing. The shift mirrors pricing changes by Anthropic and Cursor as agentic coding tools generate highly variable token costs per developer. github.blog
OpenAI publishes Symphony orchestration spec OpenAI released Symphony, an open-source specification that connects Codex agents to issue trackers including Jira, GitHub Issues, and Linear. Tickets can trigger agent runs without custom integrations, with the goal of turning ticket queues into always-on agent workflows. openai.com
Google DeepMind partners with South Korea on scientific research DeepMind signed a partnership with the South Korean government to apply frontier models to national scientific priorities. The agreement follows similar arrangements DeepMind has reached with the UK and other governments over the past year. deepmind.google
Meta signs space-based solar power deal Meta agreed to buy electricity from Overview Energy generated by orbiting solar arrays and beamed to Earth. The contract is Overview's first and an early commercial purchase of space-based solar by a hyperscaler aiming to power data centers overnight. techcrunch.com
Canva AI tool replaces "Palestine" in user designs Canva's Magic Layers feature was caught silently swapping the word "Palestine" in user-uploaded images. The feature separates flat designs into editable components and is not supposed to alter visible content. Canva apologized after a user posted screenshots on X. theverge.com
ASU turns faculty lectures into AI-generated micro-clips Arizona State University is piloting ASU Atomic, which slices professor recordings into seconds-long fragments and uses AI to assemble them into learning materials. Faculty say they were not informed their lectures would be processed this way. 404media.co
Anthropic opens Sydney office and names ANZ general manager Anthropic appointed Theo Hourmouzis as general manager for Australia and New Zealand and opened its Sydney office. The expansion follows Anthropic openings in Tokyo and several European capitals over the past year. anthropic.com
Researchers propose "levels x laws" taxonomy for agent world models A new paper argues the term "world model" carries inconsistent meanings across labs and proposes a framework with three capability levels plus a separate axis for environmental laws. The taxonomy targets agents that manipulate objects, operate software, or coordinate with other agents. huggingface.co