AMD announced it would acquire Fei-Fei Li's World Labs for $8.2 billion in stock — according to CNBC, this is the company's second-largest acquisition in history after Xilinx. Since 2024, World Labs has been building spatial intelligence models: generating, reconstructing, and simulating interactive 3D scenes from text, images, and video. In the same news package, xAI launched the public beta of its team agents Team Bots, Kling 4.0 Flash became available to Ultra subscribers, and The Information reported that China is considering allowing new Nvidia chip purchases for ByteDance and Alibaba.

image
image
image

What happened

On September 28, AMD announced it would acquire World Labs for $8.2 billion in stock; the deal is scheduled to close by the end of 2026 and is subject to regulatory approval. Fei-Fei Li will take the position of EVP and Chief Scientist at AMD and will report directly to Lisa Su, while World Labs co-founders Justin Johnson and Ben Mildenhall will continue to lead the lab's team. Starting September 28, Kling 4.0 Flash is open to Ultra subscribers: it claims up to 30 seconds of video in a single pass, twice as much as Kling 3.0, plus up to 10 keyframes, up to 15 references, point editing, stereo sound, and multilingual lip-sync, with the full version expected to launch in October. In the same release, xAI opened the public beta of Team Bots on the Teams and Enterprise plans. Separately, The Information reported that China is considering allowing the purchase of new Nvidia graphics chips for ByteDance and Alibaba.

Context

World Labs has been working on "world models" since 2024, meaning systems that understand and reproduce space, not just text: the lab is engaged in generating, reconstructing, and simulating interactive 3D scenes from text, images, and video. The deal fits into a wave of consolidation in which major players are buying up teams around a new class of models: in the same field, deals included Nvidia (Groq assets, about $13 billion for Hugging Face) and OpenAI (about $6.4 billion for io). A third element of the background is the economics of open weights: Vercel's metric of 56% of tokens and OpenRouter data show that the shift to open models has become a way for companies to save money, not an ideology.

Why this matters for the industry

AMD is not buying revenue, but an understanding of model evolution: according to Lisa Su, the deal is needed to design computing platforms in advance for the next stage of AI beyond LLMs, namely robotics, simulation, and physical AI. The workload profile here differs from token generation, so if successful, world model research will directly set the specifications for AMD's future accelerators. Integration risk has been partially mitigated in advance: the partnership began with optimizing the training and inference of World Labs models on AMD GPUs, meaning the junction of models and hardware has already been tested before the deal was announced. Team Bots cement the shift from "one assistant per person" to shared agent infrastructure for a role or process, with shared skills and private memory for each employee, and this pattern will likely be copied by other agent platforms. If China approves the purchases, Nvidia could return to the Chinese market: RTX Pro chips are estimated at about $13,000, and at a pace of up to 500,000 cards per quarter, this is a market worth up to $6.5 billion per quarter.

Why this matters for users

You can try it now. Teams on the Teams and Enterprise plans get shared Grok bots Team Bots that get context from files, instructions, and skills, work with plugins for Salesforce, Notion, and GitHub, and get access to third-party APIs via credentials; ready-made bots for sales, product management, marketing, and analytics are available via the x.ai/bot/ links from the xAI announcement. Kling 4.0 Flash is available to Ultra subscribers — our own tests currently replace the lack of independent benchmarks. If bills for closed APIs are growing, a working pattern at the level of Tinder and AT&T is routing some requests to open models through gateways like Vercel or OpenRouter; the practical first step is to measure what share of requests can be moved to open models without a loss in quality. The AMD and World Labs deal itself does not affect current products, but it is worth keeping an eye on until it closes.

What is still unknown / limitations

The AMD and World Labs deal is not closed: it awaits regulatory approval by the end of 2026, and the final effect depends on whether the lab's research core is preserved. Kling 4.0's characteristics are vendor specifications without independent measurements; open questions are the coherence and physics of motion in long passes and the quality of lip-sync, which can be assessed after the full version is released. The Nvidia story in China is based on a report of a decision being weighed, and the pace of up to 500,000 RTX Pro cards per quarter by the end of December is rumored; neither the permission nor the delivery schedule has been confirmed yet. For Team Bots, the reliability of separate employee memory and shared skills can only be assessed based on accumulated beta statistics.

Sources

Author

Look at AI, editorial team