Edition 006 — The hub gets bought, the frontier gets cheaper, the PC gets an agent

NVIDIA just bought the open-model commons for $12.93B, Anthropic shaved a quarter off the frontier bill, and local AI quietly became a shipping product category. Three of the four stories this week are about who controls the rails — and the fourth is about who edits your screenshots.

NVIDIA acquires Hugging Face for $12.93 billion

NVIDIA and Hugging Face partnership

NVIDIA's largest acquisition ever is the open-model hub itself — and Jensen Huang is promising it stays open. Definitive agreement signed September 2, 2026, announced September 3; SEC 8-K filed as nvda-20260902. The price is $12,930,300,000, and it buys a platform of 18+ million developers, 3+ million models, 500k datasets, 1M applications, and 200k+ companies.

The awkward part: NVIDIA is already Hugging Face's largest open-model contributor, with 500+ models and 250+ open datasets. Huang's post is explicit on the neutrality question — "Hugging Face will remain an open platform for the entire AI ecosystem. Developers will choose the models they want, frameworks, clouds, inference providers, and computing platforms. NVIDIA compute will not be required." Whether that holds post-close is the entire story. Regulatory review is pending; NVDA rose ~1% in early trading. Hugging Face was founded in 2016 by Clément Delangue, Julien Chaumond, and Thomas Wolf.

Anthropic ships Claude Fable 5.1 GA and Claude Mythos 5.1 by invitation

Claude Fable and Mythos 5.1

Anthropic's frontier gets cheaper, not bigger — and the dangerous tier is gated behind trusted access. Both shipped September 1, 2026: Fable 5.1 is GA, Mythos 5.1 is invitation-only via trusted access programs for cybersecurity and life sciences. Same underlying model, different safeguard tiers.

Pricing is the headline: roughly 25% cheaper than Fable 5 for typical token-billed workloads, up to ~45% cheaper for highly agentic work, driven by reduced cache-read prices. Safeguard false positives in cybersecurity drop 60%. Fable 5.1 can discover software vulnerabilities but will not develop exploits; a biology access program runs with the US government. New Enterprise Frontier Safeguards for zero data retention roll out later this fall. Defaults: High effort in Claude Code, Medium in Claude Cowork and Claude.ai. The pitch is incremental — coding, knowledge work, long-horizon problem solving — but the price cut is the part competitors will have to answer.

NVIDIA pushes local AI into the PC mainstream at IFA 2026

NVIDIA local AI at IFA

Local inference is no longer a hobbyist project — NVIDIA is shipping it through Microsoft, llama.cpp, vLLM, and a new personal router. Announced September 3 at IFA 2026, with Microsoft and partners. Simplified local AI support for NVIDIA GPUs is coming to Hermes Agent, OpenClaw, and the Perplexity Portable Computer.

The performance claim is concrete: up to 1.9x faster local inference via new llama.cpp and vLLM optimizations, live now and also accessible through LM Studio and Ollama. NVIDIA PAIR — Personal AI Router — distributes inference across the PCs on a user's local network, which is the part that actually changes the shape of the market. New RTX Spark Windows PCs arrive in October. The story isn't one box; it's a household of boxes routing among themselves.

Google Pics reaches GA in Workspace

Google Pics in Workspace

Google's image editor is now a Workspace native — object-level editing powered by Nano Banana, one click from Docs and Slides. GA September 1, 2026, announced earlier at I/O May 2026. Follows Google Vids into the Workspace suite. Click an image in Docs or Slides and the Pics editor opens as an overlay; pics.new works too.

The toolset is the real pitch: select and edit individual objects, edit or reformat or translate text inside images, crop for web/social/print/digital, upscale to 2K or 4K, batch edits with revert. Powered by Nano Banana, the Gemini imaging model. Editing images in Drive is coming in a follow-up update.