Tokens by Michael Brenndoerfer, Week #3
OpenAI Offers the US Government a Five Percent Stake
Today's issue: OpenAI woos Washington with equity, AWS winds down MTurk, and Google gets deterministic with agents.
Welcome to #3 of your weekly AI newsletter.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
Main Story
OpenAI Tries to Put Uncle Sam on the Cap Table
OpenAI wants to make the federal government its newest shareholder. According to a report from Time, Sam Altman pitched a plan to grant the U.S. government a five percent equity stake via a public sovereign wealth fund during meetings with President Trump and senior officials.
The proposal comes as regulatory pressure mounts. Under a recent executive order, the government can subject any major model to a 30-day national security review before release. Bypassing that red tape is much easier if the regulator's own balance sheet depends on your next product launch.
While the equity transfer would require congressional approval, Washington has a history of taking cap table positions, including a 10 percent stake in Intel in 2025. With massive infrastructure demands growing (and SoftBank dropping another $10 billion tranche of its planned $30 billion follow-on), having a nation-state partner could make zoning permits for gigawatt-scale data centers remarkably easy to secure.
- the pitch: Altman wants to set up a sovereign wealth fund to distribute AI equity directly to Uncle Sam.
- the precedent: Washington already holds a 10 percent stake in Intel alongside positions in quantum startups.
- the cash: SoftBank added $10 billion to its planned $30 billion OpenAI follow-on. Gigawatt-scale computing is incredibly expensive.
Bottom Line: Putting the U.S. government on the cap table changes the dynamics of tech oversight. Instead of policing the models, Uncle Sam becomes a shareholder with a direct interest in the stock price.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
The Feed
-
Anthropic launched Claude Sonnet 5, featuring adaptive thinking for autonomous coding and web browsing. Cheap introductory pricing suggests Anthropic is bidding aggressively for developer mindshare.
-
Notion 3.6 introduced External Agents, letting teams launch Claude and Cursor bots directly from kanban boards. It also plugs into Miro, Box, and ClickHouse via the Model Context Protocol.
-
Fintech startup AIsa raised $6.5 million to build payment infrastructure for autonomous agents. Over 20,000 AI bots already use the platform to run transactions without human approval.
-
Together AI secured $800 million in a Series C round led by Aramco Ventures. Its valuation is now $8.3 billion as oil money floods into open-source infrastructure.
-
Abu Dhabi investment giant MGX closed its debut AI infrastructure fund at $49 billion. The capital is earmarked for physical hardware, specifically advanced data centers and semiconductors.
-
AWS is winding down Amazon Mechanical Turk and has closed signups for new customers. The quiet sunset of the original human-labeling engine shows how fast the industry is shifting toward synthetic data.
-
Anthropic's Claude Mythos model discovered Squidbleed, a critical memory leak in Squid proxy software that went unnoticed by human developers for 29 years.
-
Google Cloud quietly launched the Open Knowledge Format. The open standard uses git-tracked markdown files to build local agent memory and bypass complex external SDKs.
-
Google launched ADK 2.0 to replace raw LLM orchestration with deterministic directed workflow graphs. The structural change cut token consumption in half in a customer refund benchmark.
-
A new IETF draft outlines a standardized Agent Directory so autonomous bots can discover and transact with one another securely.
-
Cerebras will run OpenAI's upcoming GPT-5.6 Sol model at 750 tokens per second. This wafer-scale deployment targets complex agent loops, which drops multi-step runtimes from 30 seconds down to three.
-
Japan's Money Forward is building a back-office agent suite powered by Claude and the Model Context Protocol to automate HR onboarding amidst severe domestic labor shortages.
-
Researchers introduced Looped World Models. This design routes data through a single Transformer block multiple times, which allows a 1-billion-parameter model to match much larger LLMs on reasoning tasks.
-
At ICML 2026, Apple demonstrated local developer agents running on its native MLX framework. The demo showed that complex coding tasks can run on Apple silicon without a cloud budget.
-
Insilico Medicine launched ChemCensor to evaluate chemistry models on complex synthesis pathways instead of oversimplified single-answer test keys.
See you next week!
Michael
Get Tokens in your inbox 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
Join 5,000+ readers · No spam, unsubscribe anytime.