Google makes its Pathways training system generally available on Google Cloud

On November 12, 2026, Google Developers Blog announced that the Pathways training system has reached General Availability on Google Cloud Platform.

According to the post, Pathways is a distributed training system built around a single controller. It represents the full computation as a global dataflow graph, which allows it to coordinate large-scale model training across multiple groups of accelerators. The design departs from traditional SPMD approaches such as DDP and FSDP, which depend on synchronous collective operations. Instead, Pathways relies on asynchronous dispatch and a central scheduler to keep tensors moving and maximize hardware utilization. Google states that this setup avoids the mutual blocking and deadlocks that can arise under uneven workloads, including pipeline and reinforcement learning tasks, and that it decouples model definition from physical hardware through virtual devices.

Google also notes that Pathways is the system it built to train its own large models, including Gemini. As of the announcement, no further rollout details, benchmarks, or follow-up developments have been published, so the public record currently stops at the GA declaration itself.