September 15 was a day that delivered both moves from the "make AI faster" camp and voices worried about where AI is headed. NVIDIA was reported to be in talks with Rebellions, a startup specializing in inference NPUs; in US-China relations, a framework for a "strategic AI dialogue" surfaced; and The Guardian published a warning from someone who had worked at Google DeepMind. On the product and research front, Vidu S2—a video generation system you can talk to and edit on the fly—and LLaDA-UI, which turns diffusion language models into GUI agents, are making their mark. In Japan, physical AI continued to spread across manufacturing floors, including FANUC's AI welding agent.

NVIDIA in Talks with Rebellions, an Inference NPU Startup—Investment and Acquisition on the Table

According to EE Times Japan, NVIDIA is in talks with the Korean startup Rebellions. Rebellions specializes in NPUs (Neural Processing Units) for inference, and the talks reportedly go beyond a technical partnership, with investment and even an acquisition on the table.

As the processing demands of generative AI shift their center of gravity from training to inference, specialized accelerators that determine inference efficiency are rising in value. NVIDIA—whose CUDA ecosystem gives it overwhelming dominance—reaching outside to pull in inference-specialized technology can be read as groundwork for the next phase of competition. That said, reports indicate the discussions are still at an early stage, and the outcome remains to be seen.

US-China "Strategic AI Dialogue" Emerges—Open Source Community Cool Toward "AI Deceleration"

Sinocism, a newsletter specializing in China, reported that the US and China are discussing the establishment of a "strategic AI dialogue" (via Hacker News). Details remain at the reporting stage, but the fact that an official dialogue framework is taking shape between two countries whose regulation and competition are deeply entangled is worth noting.

Meanwhile, in developer communities, skepticism toward the "AI deceleration" argument itself is conspicuous. On Reddit's LocalLLaMA, a user posted that "the narrative of AI taking over is just marketing that has been around since 2019," and the thread also cited a Reuters report in which Chinese state media criticized Anthropic's deceleration calls as "Cold War tactics." The divide between acceleration and caution is spreading beyond the political world into the developers on the ground.

"I Worked at DeepMind. Heed the Warnings"—A Former Employee Writes in The Guardian

The Guardian published a contribution from someone who worked at Google DeepMind, arguing that warnings about AI should not be dismissed (via Hacker News). Framed as a warning grounded in insider experience, it has been spreading as a counterpoint amid the ongoing wave of acceleration coverage.

As we have covered on this blog, the past few days have seen intense clashes between voices demanding "deceleration" and voices pushing "acceleration" through. With one of the insiders now speaking out in a major paper, this debate seems set to take on even more color in the days ahead.

Vidu S2—Real-Time Video Generation You Can Talk to and Edit on the Spot

The video generation model Vidu S2 has been unveiled as a system that supports real-time conversation, instruction, and editing (via Hugging Face Daily Papers). The key points:

S2-Avatar**: An interactive character running at 720p and 25–42 FPS. Mid-conversation, you can hand it a new reference image to change its outfit, have it hold objects, or switch scenes

S2-Editing**: Performs style transfer, virtual try-on, and real-time swapping of characters or backgrounds while preserving the motion and timing of the input video

Spatial video**: Brings generated characters and live editing into stereoscopic video, aiming for immersive VR experiences

In evaluations, the system reportedly achieves SOTA on all five public benchmarks spanning avatar generation and video editing. Video in which generation and response proceed simultaneously widens the horizons of streaming and avatar applications. Quality and latency in real-world operation remain to be verified.

Diffusion Language Models as GUI Agents—LLaDA-UI

LLaDA-UI, which extends diffusion large language models (dLLMs) into multimodal GUI agents, also debuted the same day. It uses a 16.7B-parameter MoE configuration, pairing a native-resolution vision encoder with the diffusion language backbone LLaDA2.0-mini-base, followed by a two-stage training pipeline that fine-tunes on GUI data from mobile, desktop, and web.

The results are strong: it substantially outperforms Qwen2.5-VL-7B on grounding and multi-platform navigation benchmarks, and reportedly exceeds even Qwen3-VL-8B on four of the six GUI benchmarks reported. GUI operation is a job of "repeatedly perceiving screen states and instantly emitting spatially grounded actions," which is a natural fit for dLLMs and their low-latency block-wise parallel generation. The significance of diffusion-based language models extending their application range from text generation to UI operation is not small.

Physical AI Spreading into Japanese Manufacturing—FANUC, UR, and GMO

There are developments in Japan as well. FANUC has developed an "AI welding agent" that reads part drawings and has a robot perform arc welding. It automatically generates welding conditions and robot motions, aiming to automate skilled work.

Universal Robots (UR) announced Gen 7, a new collaborative robot platform built for physical AI. Shipments will begin sequentially from October 2026, with plans to add onboard GPUs in 2027. It is designed as a foundation for smoothly deploying AI capabilities to manufacturing floors.

Meanwhile, GMO Tenbin AI of the GMO Internet Group made waves by announcing it would "provide AI for free to students holding email addresses issued by the University of Tokyo." The company also explicitly notes that "this service is unrelated to the University of Tokyo," and as a private initiative without official university endorsement, it has been drawing attention on social media.

Over 90 TPS on an RTX 3090—An Ampere-Specialized llama.cpp Fork and a "Time-to-Break-Even" Calculator

The local LLM scene is lively too. llamAmpere, a custom llama.cpp fork posted on Reddit's LocalLLaMA, publishes a configuration that—through Ampere-architecture-specific optimizations (said to benefit Blackwell and Lovelace as well)—achieves over 90 TPS with quantized 27B-class models and contexts of up to 240K tokens. The poster says that "running 27B on a 3090 at API-level speed has made both the model and the card significantly more useful."

Around the same time, "Sunk Cost" appeared on Show HN—a tool that calculates when a local LLM rig pays for itself. Enter your machine, model, and daily token usage, and it estimates how long it takes to recoup the investment compared with token-billed APIs. A simple but practical creation that lets you put numbers on the claim that "buy a Mac and run things locally and it will pay for itself."

---

Investment for acceleration, dialogue over acceleration, and diffusion into the field. Three currents at different layers advancing at once—if that is the air defining this week in AI, then this day's news looks like a microcosm of it.