AI News: September 18, 2026

1. Crusoe Raised $3.9 Billion at a $30.9 Billion Valuation for Modular AI Factories

Crusoe closed a $3.9 billion round led by Atreides Management, Mubadala Capital, and Valor Equity Partners, with Founders Fund, GIC, Nvidia, Qatar Investment Authority, Radical Ventures, and TPG participating. The capital funds existing sites, including the large Abilene, Texas facility used by OpenAI, and a second bet on transportable modular data centers called Spark that deploy without assembling a large construction workforce on site. Source

2. PrismML Compressed a 27B Model to 5.9 GB With Ternary Weights

PrismML, an AI lab founded by Caltech researchers, released Bonsai 2 27B, a compression of Alibaba’s Qwen3.8 27B down to 5.9 GB, a 9x to 10x memory reduction. The method quantizes each weight from 16 bits to one of three values, +1, -1, or 0, and retains 98% of Qwen’s aggregate benchmark scores. The lab raised a $22.25 million seed from Khosla Ventures, Cerberus Capital, and Caltech. Source

3. The FAA Signed an $875 Million, 12-Year AI Contract for Air Traffic Flow

The FAA awarded Air Space Intelligence a $875 million, 12-year contract for SMART, a cloud platform that analyzes airline schedules, weather, airport capacity, airspace conditions, and operational constraints to predict traffic flows and flag conflicts before they materialize. Rollout starts in the Washington, D.C. metropolitan area before expanding to other regions. Source

4. King Charles Hosted a Private AI Summit and Asked for Control Mechanisms First

King Charles convened a private summit with Nvidia’s Jensen Huang, OpenAI CFO Sarah Friar, Anthropic’s Tino Cuéllar, UK AI Minister Kanishka Narayan, the head of UK foreign intelligence, and papal AI advisor Paolo Benanti. He called the technology both intriguing and deeply concerning, credited its potential in medicine and science, and asked whether sufficient means of control exist before it is too late. Source

5. Unredacted Filings Show a Microsoft Exec Calling AI Scraping the Largest Theft of Labor in History

Newly unsealed court filings show Microsoft privately characterizing OpenAI’s data practices as theft while both companies scraped paywalled New York Times content and built datasets from it. Internal messages warned that the practice would gut publishers, and one executive described AI scraping as the largest theft of labor in human history. Source

6. Base Labs, Hugging Face, and Goodfire Formed an Open-Weight Safety Partnership

Base Labs, the research arm Baseten spun up earlier this year, partnered with Hugging Face and Goodfire to develop and publish methods for training and monitoring open models. The stated target is the abliteration problem: Hugging Face now hosts more than 6,000 models with safeguards stripped out. The group is framing this as a safety standard built into model development rather than bolted on afterward, with an open call for developer contributions and no published timeline. Source

7. A Wave of Startups Is Betting That the Only Thing That Can Watch an Agent Is Another Agent

Agent oversight has become its own funded category, with Apollo Research’s Watcher, Goodfire’s Silico, and Embroidery competing on activation probes and reasoning analysis, while Braintrust, LangChain, and Judgment Labs have raised hundreds of millions between them and Y Combinator has funded 106 AI observability companies. The counterargument in the piece is that network-level logging and standard cybersecurity practice would catch more, sooner. The framing incident is the Hugging Face breach, where nearly 12,000 agents coordinated faster than humans could track. Source

8. OpenAI Is Reportedly Working on the Hodge Conjecture

OpenAI is reportedly working on the Hodge conjecture, a second Millennium Prize Problem, after its still-unconfirmed Navier-Stokes result. Employees expect a solution soon, though the report notes any announcement could be held back given the PR fallout from how the Navier-Stokes claim was handled. Source

9. GPT-6 Astra Finished Pokemon FireRed in 18 Hours, Then Farmed Potatoes in Minecraft for Hours

GPT-6 Astra showed a sharp jump on video game benchmarks, completing Pokemon FireRed in 18 hours against a prior 96, plus completions in Factorio, Fallout 3, and Portal. The reported mechanism is that the model distills experience into compact rules. The same trait produced the failure case: after one bad Creeper encounter in Minecraft, it spent hours farming potatoes instead of progressing. Source

10. A Bloomberg Developer Used Astra to Crack an 83-Year-Old Enigma Message in Ten Hours

A Bloomberg developer reported decrypting an 82-character 1941 Wehrmacht Enigma transmission using GPT-6 Astra, a message that had gone unsolved for 83 years. The plaintext turned out to be a soldier asking about his march route. The result still needs independent review. Source

11. An OpenAI Codex Developer Put a Number on the Coordination Tax of Agent Swarms

Eric Provencher, who worked on OpenAI Codex, argued that running more than two parallel sub-agents almost always burns tokens without improving quality, because agents do not trust each other and re-verify each other’s work. His example: a project that spent $20,000 in tokens refactoring a single Python file across 1,393 Fable agents, work he argues a single Astra agent could have done for far less. Source

12. OpenRouter Token Volume Is Up 25,000% Since January 2025, and That Says Less Than It Looks

Weekly token consumption on OpenRouter rose from 0.5 trillion in January 2025 to 126.2 trillion, a more than 25,000% increase. The analysis argues the curve tracks token-hungry reasoning models and a growing population of unoptimized agents burning through context rather than a proportional increase in actual AI usage, which makes it a poor proxy for demand in either direction of the bubble argument. Source

13. A NATO-Backed Startup Put Target Identification and Attack on Small On-Board Models

Scaleout, a NATO-backed startup, is deploying decentralized AI training across military bases and drones so that small models running on board can identify and attack battlefield targets autonomously. The decentralized learning setup is the point: models update from field data without centralizing it. Source

14. Watermarking Text Made Models More Likely to Follow Harmful Instructions

Researchers found that LLMs respond differently to harmful prompts when AI text watermarking is active, and that SynthID in particular can lead models to follow instructions they would otherwise refuse. The finding puts provenance tooling and refusal behavior in direct tension, since the watermark biases the token sampling that safety training relies on. Source

15. Pinterest Is Testing an AI Room Restyling Feature

Pinterest is testing Restyle, which lets users drop furniture, decor, and lighting into photos of their own rooms to visualize changes. The commercial logic is turning saved inspiration boards into purchases by closing the gap between a pin and a room the user actually has. Source