The work
changes hands.

Before

Provision, scale, and operate servers for a web application.

After

Deploy application code to a managed runtime.

Developers focused on the application while Google managed the runtime.

Before

Build and maintain infrastructure for large analytical queries.

After

Submit queries to a managed data warehouse.

Teams wrote queries instead of maintaining analytics clusters.

Before

Manually coordinate container placement, recovery, and scaling.

After

Declare the state that Kubernetes controllers should maintain.

Controllers kept applications close to their declared state.

Before

Join separate tools for building, training, deploying, and monitoring models.

After

Manage more of the model lifecycle through one platform.

Training, deployment, and monitoring moved into one managed workflow.

Before

Use separate processing paths for text, code, images, audio, and video.

After

Use one natively multimodal model family across those inputs.

The model handled more of the work between input types.

GPU cluster

Scale general-purpose parallel processors across an accelerator and network stack.

Ironwood pod

Co-design TPU matrix compute, HBM, ICI networking and XLA as one scale-up system.

Lower serving cost. Faster inference. Built to scale.