The AI Week in Review: Nvidia Buys Hugging Face, Claude Gets a Browser

The past week moved the AI industry in several places at once. One chipmaker is reaching for the central platform of open AI, a second major provider is turning its assistant into a standalone browser, and a broad corporate coalition is warning about AI-enabled cyberattacks. Add a remarkably cheap model from China and an ongoing leadership carousel at OpenAI. Here are the six stories that matter.

Key takeaways

• Nvidia is reportedly acquiring Hugging Face for $12.9 billion, which would give it control of the most important exchange for open AI models.
• Anthropic is building a dedicated browser into the Claude desktop app, letting the AI agent click and type on the web itself.
• More than 100 companies warn in an open letter of a wave of AI-enabled attacks on critical infrastructure.
• The Chinese model GLM-5.3-Flash delivers top performance at a fraction of the cost and ran, per the maker, entirely on Chinese chips.
• Salesforce and Anthropic are bundling their products under the name Claudeforce, while executives keep leaving OpenAI.

Nvidia reaches for Hugging Face

According to the trade outlet The Information, Nvidia has agreed to a purchase price of $12.9 billion for Hugging Face; no signed contract exists yet. Hugging Face is known as the GitHub of AI, with around three million open models on the platform. For Nvidia, owning that access secures demand for its own compute chips, because downloaded models have to run somewhere. The risk: if the neutral exchange belongs to the largest chipmaker, its openness is in question. We covered the deal and its consequences in detail.

Claude gets its own browser

Since August 26, Anthropic has given its AI Claude an integrated browser. It appears as a side panel in the desktop app and handles web tasks on its own: pulling data from dashboards, collecting invoices from portals, filling in forms. Unlike the Chrome extension, this browser sees none of the user’s tabs, bookmarks, or passwords. The basic problem remains: hidden instructions on web pages can redirect such an agent. For enterprise customers, the feature turns on by default starting September 10. More in our full report.

Over 100 companies warn about AI cyberattacks

OpenAI, Anthropic, Google, Microsoft, Cloudflare, CrowdStrike, and dozens of other companies published an open letter on AI-enabled cyber defense on August 27. They see only a short window to protect hospitals, water utilities, and internet exchange points against automated attacks, and they call for coordinated action and government funding. Evidence includes an 89 percent rise in AI-enabled attacks in 2025. Criticism remains: the letter names no deadlines, and many signatories sell security products themselves. Our analysis of the letter goes into detail.

A cheap China model without Nvidia chips

The Beijing company Z.ai released the open model GLM-5.3-Flash under the MIT license on August 26. It has 320 billion parameters, of which only 18 billion are active, processes up to one million characters of context, and lands tenth on Artificial Analysis’s Intelligence Index, ahead of DeepSeek V4 Pro Max. The price is about $0.045 per task, roughly one-tenth of comparable offerings. Notably, Z.ai says it ran the model entirely on about 100,000 Chinese chips, with no Nvidia hardware; that has not been independently confirmed. The company’s Hong Kong-listed stock rose sharply in response. The trend toward cheap, capable open models from China continues, as Alibaba’s Qwen3.8-Flash-Next also showed recently.

Salesforce and Anthropic launch Claudeforce

CRM provider Salesforce and Anthropic expanded their partnership under the name Claudeforce on August 26. Claude moves into the Salesforce platform as a reasoning model, and in return Salesforce comes to Claude as a plugin with 37 prebuilt sales skills, from meeting prep to pipeline review. The open beta is planned for September 2026, with more skills to follow by year’s end. Salesforce had already invested $50 million in Anthropic in 2023. The move shows how heavily enterprise software is being rebuilt around agentic AI.

Leadership carousel at OpenAI

By TechCrunch’s count, 13 executives have already left OpenAI in 2026, most recently the head of data centers. Product strategy and infrastructure are increasingly consolidating under co-founder and president Greg Brockman while the company prepares an IPO. Brockman concedes the optics are unusual, saying every departure is currently scrutinized especially closely. In parallel, it emerged that Barret Zoph, a co-founder of Thinking Machines who was without a role after a brief second stint at OpenAI, is moving to Google as a vice president of research.

The week in context

One pattern runs through almost every story: the large providers are trying to occupy more of the value chain. Nvidia wants the model marketplace, Anthropic wants the browser and, through Salesforce, enterprise processes, and OpenAI is consolidating power internally. Against this concentration stands the open side, which with models like GLM-5.3-Flash shows that top performance is not tied to one provider or one chip architecture. For users, the week is a reminder to look closely at who owns the tools they work with every day. Next Friday brings the next review.