Vertex AI

Three things, mostly:

Hosted models. Gemini Pro, Gemini Ultra, Imagen, and Veo are all available via the Vertex API with the same per-token / per-image pricing as the public APIs but with enterprise extras — VPC peering, customer-managed encryption keys, regional residency, audit logs through Cloud Logging.

Model Garden. Open-weights models (Llama, Mistral, DeepSeek, Qwen, and a long list of HuggingFace transfers) deployable as managed endpoints. You pick the model, click deploy, get an API. Prices are per-second of GPU usage, which is usually cheaper than building your own GKE cluster and running a vLLM serving stack — but pricier than Replicate or Together for the same model.

Pipelines and training. The custom-training side: KubeFlow pipelines, hyperparameter tuning, distributed training on TPU v5p / v6, evaluation, and deployment to managed endpoints. This is the part most directly comparable to SageMaker.

The documentation sprawls. Vertex absorbed several previous GCP ML products (AI Platform, AutoML, Cloud AI), and the docs still show seams. A search for "deploy model" can land you in three different products; the right one is usually Vertex Endpoints but the legacy AI Platform docs remain top-ranked on Google.

Pricing is opaque until you sit down with the calculator. Per-token is clear; per-second-GPU on custom training depends on machine type, region, and committed-use discount; pipelines have their own per-step cost; and a real training job touches all three. The bill is rarely surprising in the bad direction once you've run a project, but the first project is hard to estimate.

GCP onboarding is the cost. If your team isn't already on Google Cloud, picking Vertex means picking GCP — IAM, billing, BigQuery, Cloud Storage, the whole thing. SageMaker has the same dynamic for AWS. The right choice is usually "wherever the rest of your data lives."

If you're on Google Cloud, Vertex is the default ML platform and a solid one — Gemini hosting alone is a sufficient reason to use it. If you're not on GCP, the case is narrower: you'd go to Vertex specifically for Gemini-with-enterprise-controls, and you'd use SageMaker, Anyscale, or self-hosted infrastructure for everything else.