<p>A multi-cloud environment provides resources for complex applications but poses challenges, especially for embedded artificial intelligence. When task data for training such models is stored in data centers, far-off offloading points or poor network quality can increase offloading time and cost. Thus, how to offload training tasks based on global resources in a multi-cloud environment is crucial. This paper proposes an algorithm named <i>a global collaborative scheduling method for embedded artificial intelligence task offloading in a multi-cloud environment</i>. This algorithm enhances task offloading efficiency by leveraging three distinct strategies: the same-data-center storage and computing integration strategy, the remote mounting strategy, and the integrated storage and computing strategy based on multi-cloud data flow. It constructs a global collaborative scheduling model aiming to minimize both the cost and completion time of tasks and subsequently employs a multi-objective optimization algorithm to derive the optimal solution. Finally, the effectiveness of the proposed method is validated through a series of experiments carried out in Kubernetes and the real world. The experimental results demonstrate that, compared with remote mounting methods in a multi-cloud environment, this algorithm exhibits higher efficiency and stability in task offloading.</p>