Anthropic’s 1M TPU Deal & AMD’s $1.3K AI GPU Launch Redefine AI Infrastructure
AI Infrastructure & Compute
Anthropic Expands Google Cloud TPU Partnership to 1M+ TPUs and >1GW Capacity by 2026
Anthropic is significantly scaling its AI infrastructure with Google Cloud, committing to over 1 million TPUs and 1+ Gigawatt of capacity by 2026. This multi-billion-dollar investment underscores the accelerating demand for high-performance AI training and inference hardware.
AMD Launches Radeon AI PRO R9700 (32GB VRAM) for $1,299 on October 27
AMD’s new Radeon AI PRO R9700 GPU, priced at $1,299, targets AI workloads with 32GB VRAM and launches on October 27, 2025. The card promises substantial performance gains for AI training and inference tasks.
AI Models & Research
New "Markovian Thinking" Technique Enables Million-Token AI Reasoning
Researchers introduced "Markovian Thinking", a method that decomposes complex reasoning tasks into smaller, interconnected steps. This could allow AI models to handle million-token contexts with improved coherence and efficiency.
Cerebras Releases Pruned GLM4.6 Models (25%–40% Pruning) on Hugging Face
Cerebras published pruned FP8 checkpoints of the GLM4.6 model (25%, 30%, and 40% pruning ratios) via Hugging Face. These versions reduce computational overhead while maintaining performance, leveraging their REAP pruning technique.
Qwen3 Next Support Added to llama.cpp for Local Deployment
The Qwen3 Next model family is now compatible with llama.cpp, enabling local execution. This update, contributed by Piotr, expands options for running advanced LLMs offline.
Undocumented OpenAI Alpha Model "GPT5-Mini Scout V4.2" Discovered
Users reported an unreleased OpenAI model, "GPT5-Mini Scout V4.2", capable of referencing files uploaded weeks prior—even outside the current chat session. This hints at persistent memory or advanced contextual retention.
Open-Source & Developer Tools
VT Code: Rust-Based Terminal Coding Agent with AST-Aware Edits
VT Code is a new open-source coding agent written in Rust, supporting AST-aware edits and local LLMs via Ollama. It integrates with multiple backends (OpenAI, Anthropic, Gemini, etc.) and enforces policy-controlled workspace safety.
Corporate & Industry Moves
Meta Lays Off Open-Source Llama Team Amid Safety Cuts
Meta is cutting staff from its FAIR team, including veterans working on the open-source Llama models, while preserving the TBD Labs group focused on next-gen AI. The layoffs coincide with broader safety and privacy team reductions.
OpenAI Acquires Sky Maker Software Applications Inc.
OpenAI acquired Software Applications Inc., the developer of Sky, a tool enhancing AI capabilities on computers. The acquisition aims to integrate Sky’s features into OpenAI’s ecosystem.
Ethics & Controversies
OpenAI Faces Backlash Over Request for Teen Suicide Victim’s Family Data
OpenAI is under fire for requesting sensitive information from the family of a teenager who died by suicide, raising ethical concerns about its data collection practices and motives.