This week in AI: Nvidia targets Hugging Face

Good morning,

Nvidia reportedly agreed to buy Hugging Face for $12.9 billion, OpenAI announced plans to pull its models from Cursor following SpaceX's acquisition, OpenAI published the first benchmark results for its custom Jalapeño chip, and Anthropic gave Claude native browser control across Chrome and desktop.

Let's dive in!

Nvidia reportedly agrees to buy Hugging Face

Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, bringing the central repository for open AI weights, datasets, and demos under the world's dominant chipmaker.

• The reporting. Reuters reported the agreement based on sources from The Information, though neither company has officially confirmed the deal. TechCrunch noted that separate reports describe the agreement as unsigned and still subject to change.
• The strategic logic. Nvidia already controls the hardware layer. Owning the default hub where developers discover, benchmark, and host open models extends its influence deep into distribution as rivals like OpenAI, Google, and Amazon build proprietary silicon.
• The developer impact. Hugging Face's core value relies on hardware neutrality. If the deal closes, developers will need to watch platform neutrality, hosting costs, and potential regulatory scrutiny closely while maintaining local backups of critical weights.

OpenAI cuts off Cursor following SpaceX acquisition

OpenAI announced plans to terminate model access for Cursor on November 12, marking a massive escalation in the corporate rivalry between OpenAI and Elon Musk's ecosystem.

• The catalyst. The decision follows SpaceX’s acquisition of Cursor. OpenAI cited contract and terms-of-service concerns as the reason for withholding future models and ending OpenAI models access in Cursor.
• The transition. Existing access remains active during a 90-day notice period, though OpenAI has not announced any migration path for affected developers.
• The fallout. Cursor users will soon have to rely entirely on alternative models like Grok 4.6, Claude, or open-weight alternatives like GLM.

Elon Musk co-founded OpenAI in 2015 and later accused Sam Altman and the company of abandoning its nonprofit mission. A federal jury rejected his lawsuit in May because he waited too long to sue; Musk said he would appeal. The Cursor cutoff is the latest commercial consequence of that feud.

Want the practical version? Our new Grok Bot guide explains how SpaceX’s always-on agent works, what it costs, and how it compares with self-hosted options such as Hermes Agent and OpenClaw.

OpenAI's Jalapeño chip posts first benchmark results

OpenAI published the first measured benchmarks for Jalapeño, its custom inference processor co-developed with Broadcom for large models and agentic workloads.

• Performance gains. Tested on GPT-OSS 120B, DeepSeek R1, and Kimi K2.5, OpenAI claims Jalapeño delivered 1.5 to 1.9 times more work per watt and 1.7 to 3.6 times lower latency than Nvidia GB200 and GB300 systems.
• Why latency matters. Multi-step agents accumulate delays across every tool call and sub-agent loop. Shaving off latency directly speeds up long-running tasks.
• Deployment timeline. Internal deployment is planned by year-end to serve ChatGPT and API workloads, even as OpenAI continues purchasing Nvidia GPUs.

The immediate win is internal leverage over serving costs and capacity. If these numbers hold in production, developers will experience the benefits through faster responses and lower API pricing rather than renting Jalapeño instances directly.

Claude can now work in its own browser or yours

Anthropic released two major browser capabilities designed to let Claude navigate the web independently:

• Claude in Chrome (GA). Now generally available on all paid plans, this browser extension allows Claude to read web pages, switch tabs, click buttons, and fill out forms using active session logins. Source
• Cowork built-in browser. Claude Cowork added an isolated, built-in browser rolling out to Pro, Max, and Team desktop users. It keeps long-running background web tasks completely separate from personal bookmarks, tabs, and saved passwords. Source

Security reality: Prompt injection remains the biggest hurdle for web-operating agents. Malicious pages can hide instructions designed to hijack an agent. While Anthropic has introduced automated action checks, teams should keep human confirmation toggled on for financial transactions, account settings, and destructive actions.

Quick hits 🗞️

• Perplexity launched a local-first computer agent. Portable Computer runs its reasoning model, planner, and search index locally on an Nvidia DGX Spark, keeping data off the cloud without drawing usage credits. It requires Pro or Max on Linux at launch, with Windows and RTX PC support coming soon. Source
• Google released Gemini 3.5 Transcribe. The updated audio model brings live streaming transcription, recorded file processing, automated speaker labeling, and support for over 85 languages. It is currently in public preview for developers. Source
• Z.ai open-sourced GLM-5.3-Flash. The multimodal model features a 1-million-token context window and ships under an MIT license. Artificial Analysis benchmarked it at a cheap $0.09 per task, though noted it runs slower than competing flash models. Source
• Anthropic previewed a hardware standard for agents. The Model Hardware Standard (MHS) framework lets AI agents discover and operate laboratory equipment like liquid handlers, robotic arms, and microscopes via shared drivers and MCP. Source

See you next week!

Got feedback on the newsletter? Just hit reply, I read every response.

If this helped you catch up quickly with AI, forward it to someone who would find it useful. They can subscribe here.