Google has launched AI Edge Gallery, an open-source application that enables Android phones to download and run advanced AI models locally without requiring an internet connection or cloud services. This development marks a step toward on-device AI, offering instant AI capabilities with no lag. The app supports running Hugging Face models offline and is expected to be available for iOS devices soon. Concurrently, Google Cloud has introduced serverless AI inference powered by NVIDIA L4 GPUs through its Cloud Run platform, featuring pay-per-second GPU billing, scaling to zero instances, and a cold start time of approximately 19 seconds. In parallel, decentralized AI training is advancing, with teams like MacrocosmosAI and Pluralis demonstrating the ability to train large language models across consumer devices connected via the internet without loss of speed or performance. Pluralis notably achieved training a billion-parameter AI model over an 80kbps connection using consumer GPUs, highlighting progress in distributed AI model training. These innovations collectively indicate a shift toward more accessible, efficient, and decentralized AI deployment and training methods.