Alibaba Aims for 20GW by 2032, AI Announcements and Open Source Models

Alibaba has set an ambitious target for its data centers by 2032. Meanwhile, tools and models are becoming available, while various players are testing new versions and ramping up projects and analyses. Here’s a concise overview of the announced movements.

Alibaba unveils a chip and aims for 20GW capacity by 2032

Alibaba has introduced a new AI chip. The company states that this chip is expected to power 20GW of data centers by 2032.

Immediate releases: open source at Xiaomi, AWS agent, Grok 4.7

Xiaomi has released the MiMo-V2.6 Pro and Flash models as open source. AWS has launched Strands Harness, an agent that provides everything except the model. Grok 4.7 has been presented. Aikido Altar, an open-weight AI, has been announced for sovereign security.

Testing, projects, and analyses: Anthropic, Muse, Qwen, and initiatives

Anthropic is testing Fable 5.2 and Opus 5.5 ahead of their release. Meta's personal AI agent Muse has surpassed ChatGPT, Grok, and Claude in download numbers post-launch. Several projects and initiatives have been reported: Qwen’s Recreation is training agents to rebuild applications, and an update showcases the progress of the StepFun project at its stage 5. An advisory group on mathematics and artificial intelligence has been announced. Various publications and approaches are mentioned: Swarm Scaling for the scalability of AI systems, an article on the evolution of reinforcement learning towards self-improvement with MiMo-V2.6, an analysis of the new capabilities of AI systems in response to the If instruction, as well as analyses on the balance of power in open models and the business implications of AI.