Top Stories
OpenAI Floats 5% Equity Stake for the US Government
OpenAI has floated giving the US government a 5% ownership stake as a way to ease tensions with the Trump administration and blunt mounting public backlash against AI, the Financial Times reported. CEO Sam Altman argued that giving the public a financial interest in the company would be the best way to share the upside of AI; he is said to have first pitched the idea to Trump early last year. At OpenAI’s latest $852 billion valuation, a 5% slice would be worth roughly $42.6 billion.
The proposal would extend to other US AI companies, with Anthropic, Google, and Meta reportedly in scope, though it is unclear whether any would agree. The political logic is direct: if the federal government becomes a direct equity holder in a frontier lab, the regulatory and procurement stories collapse into a single balance-sheet item. MIT Technology Review’s daily Download flagged the proposal alongside the Singapore seizure of a $42M mansion linked to Nvidia chip smuggling and Supermicro Taiwan-office raids.
Companies Throttle Internal AI Use as Costs Triple
Leaked Slack chats, dashboard screenshots, and emails from half a dozen enterprises - including Atlassian, Adobe, Amazon, and Citi - show employers cutting off access to specific AI models, urging workers to switch to less powerful alternatives, and watching monthly bills spike. One firm’s internal AI bill tripled to over $15 million per month. Adobe is ending unlimited access to Claude across its workforce. 404 Media’s companion Tokenpocalypse podcast (hosted by Joseph Cox) covers how some teams are rewriting prompts so models respond in compressed “caveman-style” text to cut token counts.
This is the first internal-data confirmation that inference cost has migrated from a forecast to a budget item. For workers, the practical move is to assume seats at the best models may disappear or get rate-limited mid-quarter, and to keep outputs portable.
Anthropic in Talks With Samsung on a Custom AI Chip
About a week after OpenAI’s Broadcom partnership for an in-house inference chip (called “Jalapeño”), Anthropic is in discussions with Samsung Foundry about manufacturing its own custom AI silicon. Anthropic declined specifics, telling TechCrunch that “a diversified hardware stack that includes chips from Google, Amazon, and Nvidia will continue to be pivotal to its compute strategy.” Reuters first reported Anthropic’s custom-chip thinking in April.
The custom-chip field is now four labs wide: OpenAI/Broadcom, Anthropic/Samsung, Google TPU, and AWS Trainium. Nvidia is no longer the only credible supplier the frontier labs will work with, which over time should narrow the pricing power of the dominant GPU vendor.
Zuckerberg Tells Staff That AI Agents Haven’t Moved Fast Enough
At an internal town hall, Meta CEO Mark Zuckerberg said his company’s work on AI agents had not “accelerated in the way” executives had hoped. Zuckerberg also acknowledged that the year’s earlier layoffs of roughly 8,000 people (about 10% of Meta’s corporate workforce) and reassignment of around 7,000 to AI groups were “not as clean” as they should have been, driven by worries that Meta “weren’t going to move fast enough.” Meta is on track to spend as much as $145 billion on AI infrastructure this year.
It is the most senior internal-meeting acknowledgment yet that the agentic-AI hype cycle is meeting production friction inside the most aggressive-spending lab of the bunch.
Microsoft Spins Up a $2.5B AI Deployment Company
Microsoft has launched Microsoft Frontier Company, a 6,000-person operating unit with a $2.5 billion Microsoft commitment, focused on delivering enterprise AI deployments using Microsoft’s own AI tools. Early named partners include London Stock Exchange Group, Unilever, Land O’Lakes, and Accenture. Microsoft’s commercial business CEO Judson Althoff framed the unit as “the largest, most capable, outcome-driven engineering organization in the industry” and “goes beyond what has been labeled as Forward-Deployed Engineering.”
The category of dedicated AI deployment companies is now mainstream rather than experimental: AWS, OpenAI, and Anthropic already operate similar groups, and Microsoft has just formalized one as its own line of business.
Anthropic Details Fable 5 Cyber Safeguards and the CJS Jailbreak Scale
Anthropic has published the cyber safeguards in place on the re-deployed Fable 5 and a new Cyber Jailbreak Severity (CJS) framework. Fable 5 sorts cyber actions into four categories - Prohibited, High-risk dual use, Low-risk dual use, and Benign - with a wider safety margin (more false positives, fewer harmful outputs) than earlier models. High-uplift vulnerability discovery is blocked; routine vulnerability finding is allowed. The CJS scale runs CJS-0 (Informational) through CJS-4 (Critical), scored on four axes (Capability gain, Breadth, Ease of weaponization, Discoverability) with a 10-point ceiling, built in partnership with “Glasswing partners,” and surfaced through a HackerOne program (anthropic-cyber-jailbreak).
The framework is the most concrete attempt yet by a frontier lab to put numbers on jailbreak severity, and the export-control episode that briefly suspended Fable 5 and Mythos 5 in mid-June before a deal with the Commerce Department lifted the restrictions on June 30 (Wired, Politico, Anthropic statement, Anthropic redeployment) is the most legible test yet of how the US government will handle frontier-model safety disclosures going forward.
Meta Puts Ray-Ban Smart-Glasses AI Features Behind a Paywall
Meta is moving the most advanced AI features on its Ray-Ban smart glasses to a paid subscription, Wired reports - the clearest signal yet that consumer AI features are being unbundled from hardware and resold as recurring services. The framing is consistent with the broader enterprise-cost story: when AI features cost real money to run, vendors start metering access regardless of whether the device was already bought.
Quick Hits
-
OpenClaw + Instagram dating: A user automated Instagram DMs into an OpenClaw + Claude code pipeline to “find potential international partners,” per TechCrunch; the package pairs a viral script with awkward public persona artifacts, and lands in the same week as we revisited OpenClaw’s supply-chain crisis from February.
-
Cursor inside SpaceX: With SpaceX’s acquisition of Cursor in view, Wired asks whether OpenAI and Anthropic will keep allowing the editor to serve as a neutral interface to their models - a question our head-to-head Cursor, Claude Code and Copilot test put on the table back in April.
-
Flare AI safety reporting portal: A new Wired-covered site lets consumers and researchers file structured AI-safety incidents; useful public infrastructure for non-developer users.
-
PamStealer macOS malware: Ars Technica reports a Rust-written macOS infostealer that uses the Pluggable Authentication Modules interface to validate stolen credentials before exfiltration - unusually careful tradecraft by Mac malware standards.
-
Meta’s vibe-coded Pocket app: Meta has quietly launched Pocket, an experimental AI app that turns text prompts into interactive mini games - a public bet that prompt-driven interactive content is a category.
-
Woodside Energy + 50 production AI agents: MIT Technology Review profiles Woodside’s deployment of a “Startup Advisor” LNG agent and around 50 production AI agents in its energy operations.
-
simonw/llm-coding-agent 0.1a0: Simon Willison released the first alpha of a Claude Code-style coding agent built on his LLM CLI, installable via
uvx --prerelease=allow. -
EFF LGBT data hygiene guide: EFF’s Paige Collings walks through self-search, broker deletion, password manager hygiene, Privacy Badger, and Google’s “Results about you” page.
-
Jersey Mike’s IPO AI mentions: A TechCrunch tally counts 22 mentions of “artificial intelligence” or “AI” in a sandwich chain’s S-1, with no specifics and no product - a textbook case of the AI-washing pattern we flagged in February.
Worth Watching
OpenAI’s 5% proposal. The pitch makes sense politically for Altman and ideologically for sovereign-wealth populism, but a direct US federal equity stake in private AI companies would need congressional approval and would collide with procurement, antitrust, and constitutional concerns. The first concrete test is whether Anthropic, Google, and Meta publicly accept or refuse inclusion.
Inference-cost as a political story. The 404 Media throttling story and the Tokenpocalypse podcast turn the cost story from a vendor problem into a worker-rights and labor-safety story. Watch for AI-vendor pricing tiers that look designed to extract rent rather than pass through compute.
The four-lab custom-silicon race. OpenAI/Broadcom, Anthropic/Samsung, Google TPU, and AWS Trainium are now the relevant shopping list for frontier workloads. The next data point is whether Nvidia responds by cutting H100/B100 list prices or by locking in multi-year supply guarantees with the hyperscalers.
Jailbreak severity measurement. Anthropic’s CJS scale is the first peer-reviewable jailbreak-severity rubric from a frontier lab. The open question is whether OpenAI, Google, and Meta adopt a comparable framework or stay opaque on severity scoring.