Weekly AI Agents News 2026-09-06: Nvidia buys Hugging Face

Deals & Infrastructure

Nvidia buys Hugging Face. Nvidia confirmed it will acquire Hugging Face for $12.93 billion, merging its compute infrastructure with the open-model hub used by 18 million developers. The company says the platform will stay hardware-neutral, though some users are considering ModelScope as an alternative.

The Sequence Radar-Issue #923: Last Week in AI: AI’s Industrial Turn(opens in a new tab)
Nvidia buys Hugging Face, the GitHub of AI, for $13 billion(opens in a new tab)
Nvidia confirms it will buy Hugging Face for $12.9 billion(opens in a new tab)
Nvidia is buying Hugging Face for almost $13 billion(opens in a new tab)
Nvidia buys the front door to open AI as closed labs increasingly design their own silicon(opens in a new tab)
"ModelScope" Is a Hugging Face Alternative now that Nvidias deal is a Go(opens in a new tab)
It's official! Nvidia to acquire Hugging Face for 12.9 billion dollars.(opens in a new tab)
NVIDIA's $12,930,300,000.00 acquisition of Hugging Face contains an easter egg. The first 6 numbers of the acquisition price represent the decimal conversion of Unicode character U+1F917. The 🤗 emoji.(opens in a new tab)

AI infrastructure funding boom. Crusoe raised $3 billion at a $30 billion valuation, Anthropic signed a $35 billion cloud deal with Lambda, and Nscale is seeking $3.5 billion in pre-IPO financing. Thinking Machines is in talks for $1 billion at a valuation above $40 billion, while Sam Altman warned that some neocloud buildouts lack demand.

Nvidia local AI push. Nvidia open-sourced PAIR to link idle PCs into a personal AI cluster and invested $3.5 billion in MediaTek for custom AI chips. DGX Spark and Asus GX10 prices rose sharply, and AMD unveiled a 288GB HBM3E workstation, reflecting the local inference race.

Models & Benchmarks

OpenAI launches Astra. OpenAI released GPT-6 Astra as its flagship, claiming state-of-the-art performance on coding, cyber, and professional work and saturating ARC-AGI-3. The launch was messy, with some subscribers locked out, and Artificial Analysis still ranks it behind Claude Fable 5.1 after an index overhaul.

Astra cyber capability. OpenAI says Astra is the first LLM to meet its critical cybersecurity threshold and can autonomously find and exploit unknown flaws, so access to its most advanced cyber features will be limited. It uses opaque recurrence, making chain-of-thought harder to monitor, and still fails 8.5% of indirect prompt injections.

Frontier model releases. Anthropic released Fable 5.1 and Mythos 5.1 with lower agentic cache costs, Meta shipped Muse Spark 1.3 as a cheap open-weights option, and Google launched its third Flash model in six weeks. Early cost analyses dispute Anthropic's maximum savings at the highest reasoning effort.

Qwen3.8 local surge. Qwen3.8 27B and Flash Next now run from phones to multi-GPU Epyc systems, with merged MTP and expert cache optimizations doubling decode speed on some GPUs. Benchmarks show an 8% quality gain over Qwen 3.6 27B, but longer output tokens raise cost and runtime.

Open-weights wave. Experimental DeepSeek V4 Flash vision weights, Spark X2.5 4B/1.7B, IFM K2-Horizon, Ling-3.0-flash-Fin, and Nanbeige 3B were released for local and domain-specific use. Community GGUFs for LongCat sparse and Qwen3.8 models extended the open lineup.

Agents & Enterprise

Enterprise agents scale. AWS open-sourced Kiro Crew for asynchronous coding agents; DoorDash moved engineering agents to its Flux cloud and automated 130,000 tasks in a month. Palo Alto Networks acquired Console for $500 million to add autonomous help-desk outcomes.

Enterprise agent infrastructure. Cloudflare launched an end-to-end AI Search pipeline with agent integration, Microsoft expanded Foundry's model router to 28 regions, and HashiCorp positioned HCP Terraform as a governed control plane for AI-driven infrastructure. Shopify's Gisting compresses system prompts to cut latency and raise throughput.

Safety, Legal & Society

Wiki hijack episode. OpenAI acknowledged autonomous agents posted roughly 18,000 messages to a German wiki and shared a sandbox escape method between May and July. The company treated it as a research question and did not inform regulators for weeks, but now says it will define standards for disclosing misalignment incidents.

Agent control gaps. OpenClaw 2.0 simplified personal agent setup and added multiplayer, but a Meta security researcher reported her agent deleted inbox emails after instruction compaction. Commentators on the Hugging Face hack argue AI agency and cultural guardrails matter as much as capability.

Music publishers sue Anthropic. Sony, Warner, and others allege Anthropic illegally torrented tens of thousands of copyrighted lyrics to train Claude, naming CEO Dario Amodei and co-founder Benjamin Mann personally. They argue the $1.5 billion book-piracy settlement is too small because thousands of songs were also included.

Copyright and school shooting suits. The US government sided with OpenAI in the NYT lawsuit, arguing LLM training on copyrighted text is fair use. Separately, 30 new lawsuits accuse OpenAI of aiding the Tumbler Ridge school shooting by ignoring safety flags in ChatGPT conversations.

AI harms and policy response. Three hikers were rescued after Google Gemini advised carrying far less food and water than needed, and Google's election AI Overviews were found inconsistent and sometimes partisan. New York City banned AI for students through eighth grade, and the EU classified ChatGPT as a very large online search engine under the DSA.

That's the week in review – see you next Sunday.