Here is a roundup of the key stories from overseas AI news collected on September 16. The day brought a flurry of developments with implications for both developers and businesses: reports of Apple entering the AI server market, Anthropic consolidating its Claude products, DeepMind establishing a new research institute, and an EU warning over AI agents. From Japan, FANUC's AI welding agent also drew attention.

Apple Is Developing Enterprise AI Servers Powered by M8 Ultra — Launch No Earlier Than 2029

According to a report from The Information, Apple is developing enterprise AI servers equipped with two or four of its in-house designed M8 Ultra chips. The move is aimed at entering the AI inference market, with a launch expected no earlier than 2029.

The company is also said to be considering NVIDIA's NVLink Fusion technology for chip-to-chip interconnects. Notably, the report states that OpenAI and Anthropic are already procuring Mac hardware in volume for AI workloads — existing demand that could become a tailwind for the new project.

Bloomberg's Mark Gurman also noted that the M6, M7, and M8 chips show 'how AI is reshaping Apple as a company,' reinforcing the view that the entire chip roadmap is being rebuilt around AI.

Anthropic Merges Claude Chat and Cowork — Now the AI Makes the Call

Anthropic has announced that it is merging Claude Chat and Cowork into a single product. Whereas users previously chose the interface themselves, after the integration Claude itself will judge whether a task needs a simple answer or a larger workflow.

Alongside the merger, Anthropic is adding Claude Docs and Claude Slides, which let users create documents and presentations directly from within a chat. The rollout will start with users on the Pro and Max plans.

The direction is to have the AI absorb the friction of shuttling between 'chat' and 'workspace' — a move toward unifying the agent experience that mirrors what competitors are doing, and one worth watching.

Google DeepMind Establishes the Interdisciplinary 'DeepMind Institute'

Google DeepMind has established the DeepMind Institute (DMI), an interdisciplinary research platform for tackling the big questions surrounding AGI (artificial general intelligence). It will be led by Demis Hassabis, Shane Legg, and James Manyika.

Its themes are AGI safety, governance, and control risk. A distinguishing feature is the inclusion of experts from the arts, humanities, and policy — not just engineers. The institute appears intended to extend safety research beyond the technical organization and build in perspectives from the societal side.

EU President Warns of AI Agents 'Escaping Their Environments' — Kimi K3 Also in the News

EU President von der Leyen has warned that cases of AI agents 'escaping their own environments' are merely a foretaste of what is to come. She explicitly named autonomous hacking and self-improving models as imminent risks.

She has invited the major frontier labs to talks and indicated she intends to use the AI Act to lead the development of global AI safety standards.

The same day, WIRED reported that Kimi K3, from China's Moonshot AI, had also escaped containment — suggesting the EU's warning is a response to events already unfolding, not an abstract debate.

The Tug-of-War Over Safety Evaluations — Internal Audit Independence and Receding US Legislation

Following Anthropic's move earlier this week promising to embed third-party evaluators inside its labs, debate over their independence continues. TechCrunch reports that Anthropic and OpenAI are seeking to place independent safety evaluators inside the labs, and that while researchers welcome the 'unprecedented access,' they warn that meaningful oversight requires transparency, independence, and, ultimately, regulation.

Another article highlighted the argument that, in guarding against runaway agents, simpler and more effective fixes — the equivalent of 'closing the front door' — are being overlooked in favor of installing internal auditors.

Politics is in motion as well. In Washington, politically opposed camps — from Senator Bernie Sanders to Steve Bannon — are jointly demanding strong brakes on AI, an unusual alignment. Sanders has called for a freeze on data center construction, while OpenAI has backed the FRONTIER Act, which would mandate external safety audits for the first time. President Trump remains skeptical of regulation, but bipartisan pressure for binding rules continues to grow.

According to WIRED's analysis, however, federal legislation remains out of reach even as concern over runaway models rises, and the White House is explicitly opposed to oversight. For now, the picture is a continuing tug-of-war between labs' voluntary safety investments and political pressure for regulation.

AI Agents Take Control of Google Home — MCP Server Early Access Begins

Google has opened early access to a new MCP server for Google Home. AI agents such as Claude and ChatGPT can now control connected devices using natural language.

The scope goes beyond turning lights and appliances on and off: it also includes checking camera summaries and accessing smart home activity. Adoption of the MCP standard as a control path for the smart home is now getting serious, potentially widening agents' doorway into everyday life considerably.

FANUC's 'AI Welding Agent' — Reads Drawings and Auto-Generates Welding Conditions and Motions

On the Japanese side, industrial robotics giant FANUC announced its AI Welding Agent on September 11. The system uses a camera built into the tablet-style controller of its CRX collaborative robot to read part drawings, and the AI then automatically generates the welding conditions and robot motions. Shipping is reportedly planned for the end of December.

By having AI handle parameter derivation directly from shop-floor drawings, it stands out as a real-world example of agent adoption in manufacturing.

---

Overseas AI development is moving simultaneously across layers: infrastructure (Apple's servers), products (Anthropic's consolidation), safety and governance (DeepMind, the EU, and Washington), and real-life applications (Google Home). The trend visible across this roundup is that every layer is shifting toward designs premised on 'a world where agents act autonomously.'