Oracle and OpenAI Cancel Texas Data Center While Regulators Circle Mythos

The deal that wasn't

Oracle and OpenAI have cancelled plans to expand a flagship AI data center in Texas. The two companies couldn't agree on financing, and OpenAI's needs apparently shifted mid-negotiation. That's the polite way of saying someone's spreadsheet changed and the partnership math stopped working.

Data center deals are expensive and slow. You need land, power contracts, cooling infrastructure, and a timeline measured in quarters not sprints. When one party is scaling a foundation model business that reinvents its compute needs every six months, the other party—whose business is selling stable, long-term infrastructure—gets nervous. Oracle walked, or OpenAI did, or both decided the juice wasn't worth the squeeze.

The timing is awkward. OpenAI is raising capital, shipping products, and presumably still burning through GPU clusters. Oracle wants to be the enterprise cloud for AI workloads. A high-profile cancellation in Texas doesn't help either narrative, but it does clarify that even the biggest players are still figuring out how to build this stuff at scale.

Anthropic launches a marketplace and gets regulatory attention

Anthropic rolled out a new marketplace where enterprise customers can buy third-party software that integrates with Claude. Think AWS Marketplace but for AI tooling—SaaS vendors can list integrations, enterprises can browse and purchase without leaving Anthropic's ecosystem. It's a revenue play and a stickiness play. If your procurement team already has an Anthropic contract, adding a compliance monitoring tool or a customer support overlay becomes a line item instead of a new vendor conversation.

The marketplace launch is unremarkable by itself. Every platform company with enterprise traction eventually builds one. What is remarkable is that the Financial Stability Board—the global body that coordinates financial regulators and central banks—is now gathering information about Anthropic's Mythos model specifically.

Mythos is the cyber-offense research model Anthropic released earlier this year. It's designed to find exploits autonomously, which makes it both useful for red teams and terrifying for anyone imagining what happens when that capability leaks or gets sold. The FSB doesn't usually pay attention to individual AI models. They care about systemic risk: the kind of thing that could cascade across financial institutions or critical infrastructure if it goes wrong.

The fact that they're circulating a memo about Mythos suggests someone in a regulatory chair read the technical paper and decided this isn't just a research curiosity. Autonomous cyber-offense tools present a pretty obvious problem if they're commoditized. The FSB sharing insights across its member network is the first step toward either coordinated oversight or at least coordinated concern. Anthropic gets to be the test case for what happens when your safety research produces something regulators think is too sharp to leave on the counter.

A week of fine-tunes nobody asked for

The rest of the week was quiet unless you care deeply about uncensored Gemma variants or someone's guide to fine-tuning Qwen models. Hugging Face saw a flurry of uploads with names like "HERETIC-UNCENSORED-Thinking" and "The-DECKARD-V2-Strong," which tells you everything you need to know about the vibe. These are hobbyist projects: people stripping safety layers off foundation models, retraining them on private datasets, and publishing the results with benchmarks that claim the new version "exceeds root model in 6 out of 7" tests.

It's not news. It's what happens when model weights are open and compute is cheap enough that enthusiasts can afford to experiment. The output is mostly noise—models trained on undisclosed data, uploaded without reproducibility guarantees, downloaded by maybe a few hundred people who want to bypass content filters. Occasionally something useful surfaces. Most of the time it's just clutter in the model registry.

The interesting part is what it signals about where the open-weight ecosystem is heading. When the barrier to fine-tuning drops low enough that individuals can crank out custom variants in a weekend, the question stops being can we do this and starts being should we index it. Hugging Face doesn't curate. They host. That's fine until regulators start asking platforms to label or gate models that can be weaponized, at which point someone has to draw a line between a hobbyist project and a compliance headache.

We're not there yet. But the FSB memo about Mythos suggests we're closer than we were last quarter.