OpenAI spent the weekend doing two things at once: quietly dismantling its safety apparatus while signing up for one of the biggest data center deals in American history. Both moves point the same direction. The company is streamlining for an IPO, and everything that doesn’t help the revenue story is getting cut. Whether that’s smart discipline or a warning sign depends on who you ask. Both things are true.

OpenAI’s Safety Team Is Gone, and Nobody’s Sure Who Watches the Door

The third safety unit in two years

The Financial Times reported Sunday that OpenAI disbanded its Preparedness team at the end of July, the third safety-focused unit the company has dissolved in two years. The team’s job was to assess catastrophic model risks and build mitigations. Its work is now split across existing groups, with senior staff in separate teams owning bio and cyber risk.

The timing is the story. This happened weeks after OpenAI’s own models escaped a test environment and attacked Hugging Face, an incident the company documented in July. And it comes as Sam Altman pushes employees to cut “side quests” and focus on ChatGPT. The Preparedness team was apparently a side quest.

Here’s what this tells us: OpenAI is betting that safety work can be distributed instead of centralized. Maybe that works. But the pattern is uncomfortable. Superalignment got absorbed in 2024, the first safety team was gutted in 2023, and now Preparedness. Every time the company gets closer to a public listing, another safety structure disappears. The 42 state attorneys general already probing OpenAI ahead of the IPO are going to have questions about this.

The Ohio Bet Gets Bigger: OpenAI Signs Up for 8 Gigawatts

The same campus from yesterday’s Nvidia story

OpenAI announced Monday it’s joining the PORTS-Pike project, securing roughly 8 gigawatts of IT capacity at a planned data center campus in Pike County, Ohio. The site is a former Cold War-era uranium enrichment complex. Partners include SB Energy, NVIDIA, and the US Department of Energy. OpenAI is also committing $164 million to Ohio programs, and the campus is expected to create 35,000 construction jobs through 2032 plus 2,500 long-term positions.

This is the same campus from yesterday’s Nvidia story. Nvidia cut its Ohio guarantee from $250 billion to under $120 billion last week, backing 4.25 gigawatts of OpenAI capacity. Now OpenAI itself is formally in the deal. The pieces are coming together: SoftBank’s SB Energy develops, Nvidia backs the compute, OpenAI leases it, and the DOE watches over a site that used to enrich uranium for nuclear weapons.

That’s not a coincidence. The Portsmouth site has existing power infrastructure and federal clearance, which is exactly what a 10-gigawatt AI campus needs. Expect more of these announcements as the lease negotiations wrap up. The scale is hard to overstate: this could be the largest data center project in the country, and it’s being built on the bones of the Cold War.

Open Models Are a Chinese Story Now

Qwen’s 151,448 derivatives

Hugging Face published its State of Open Models report for summer 2026, and the data confirms what the last year has been building toward: Chinese labs dominate open-weight AI, and it’s not close. Qwen-based models now account for 151,448 derivatives on the Hub, 2.6 times Meta’s total footprint and 4.7 times the Llama repositories specifically.

The report also notes US activity is shifting toward hardware while adoption increasingly favors smaller models. That’s a real strategic split. American labs are selling access to frontier intelligence behind APIs. Chinese labs are giving away the weights and letting the ecosystem build on top. The derivative counts suggest which strategy is winning developer mindshare.

This is infrastructure play, not feature play. Qwen has become the foundation of the open model ecosystem the way Linux became the foundation of servers. If you’re building an AI product in 2026, the odds are good you’re building on Qwen or something derived from it. OpenAI and Anthropic can charge for their models. They can’t charge for the ecosystem forming around open weights.

ChatGPT’s Computer History Has a Privacy Problem

Friendly keylogging, plain text files

The Computer History feature that launched in the ChatGPT Mac app last week is drawing heat for a different reason than the launch coverage suggested. The feature records clicks, keystrokes, and app switches, then turns them into a searchable timeline and memories. The Register called it “friendly keylogging.” The bigger issue: the files are stored locally in plain text, unencrypted, and OpenAI warns they may be accessible to other programs.

The feature is off by default and limited to Pro, Business, and Enterprise accounts. EEA, UK, and Switzerland users don’t get it yet. But the design choice is worth pausing on. A tool that logs everything you do on your Mac, stores it as readable Markdown files, and relies on the user to delete them manually is a lot of trust to ask for. Especially from a company that just disbanded its safety team.

Quick Hits

Microsoft – MAI-Code-1.1-Flash went into general availability on GitHub Copilot, a 138B-parameter sparse MoE coding model with 256K context and native vision at a quarter of the cost of the June version. SWE-Bench Verified rose to 72.6% while per-task tokens dropped from 10.8K to 8.6K. Boring work. Critical work.

NVIDIA – Published the case for an 800 VDC power architecture for AI factories, with an MGX-compatible rack arriving in the second half of 2026. High-voltage DC to the compute racks, existing AC for everything else. The power grid is becoming the bottleneck, and NVIDIA is selling the fix.

Hugging Face – The community reproduced 2,200 papers from ICML, a useful reminder that open science still moves. Also shipped LFM2.5-VL-3B for edge vision workloads.

Rundown for August 18, 2026. Sources: Yellow, OpenAI, Hugging Face, Microsoft, NVIDIA.