
Note: pricing and features may have changed since publication. Check each provider’s documentation for current details.
Choosing between Upsun Cloud and Google Cloud Run for AI comes down to where the hard part of the workload sits: in the model, or in the application around it. Google Cloud Run and Upsun Cloud both run containerized AI workloads, but each is organized around a different unit.
Cloud Run is organized around the container. Services, jobs, and worker pools run on Google Cloud, with NVIDIA L4 and RTX PRO 6000 Blackwell GPUs, services that scale to zero, and native access to Vertex AI and Gemini.
Upsun Cloud is organized around the application. The app, its databases, workers, and on-demand task containers are defined together in one .upsun/config.yaml; every Git branch gets its own environment, and the same setup deploys on AWS, Azure, Google Cloud, or OVHcloud. Upsun Cloud has no GPUs and does not host models.
This comparison covers how each platform handles AI agents and background jobs, GPUs, data access, preview environments, isolation, compliance, and pricing.
Capability | Google Cloud Run | Upsun Cloud |
| Runs on | Google Cloud only | AWS, Azure, Google Cloud, OVHcloud |
| Workload types | Services, jobs, worker pools, functions | Apps, workers, cron, task containers |
| Longest request or run | 60 min per request; 168 h per job task | 24 h per task run |
| Scale to zero | Yes | Task containers only |
| Databases and caches | Separate products (Cloud SQL, Memorystore) | Services in the same project |
| Per-branch environments | Revision tags; no data cloning | Each branch inherits data and services |
| Untrusted-code isolation | gVisor or microVM; sandboxes in preview | Containers; bubblewrap recommended |
| Configuration | gcloud, YAML, or Terraform | .upsun/config.yaml in Git |
| Billing | Request-based or instance-based | Per second of allocated resources |
Google Cloud Run runs containers on Google's infrastructure as services (HTTP and gRPC), jobs (run to completion), worker pools, and functions. Databases, caches, secrets, and queues come from other Google Cloud products, such as Cloud SQL, Memorystore, Secret Manager, and Pub/Sub, connected through IAM and VPC configuration.
Upsun Cloud runs applications, background workers, cron jobs, and on-demand task containers next to managed services such as PostgreSQL, MariaDB, Redis, OpenSearch, and RabbitMQ, all defined in .upsun/config.yaml and deployed with a Git push. Upsun Cloud has no GPUs and no managed model-hosting service, so AI features typically call external model APIs.
Upsun Cloud runs agents as always-on workers, scheduled cron jobs, or task containers. A task container starts on an API call, runs one command for up to 24 hours (1 hour by default), and is removed when it exits. Three tasks run in parallel by default, running tasks are cancelled before a deploy on the same environment, and the caller handles retries.
Cloud Run runs agents as services (each request up to 60 minutes), jobs (each task up to 168 hours, or 1 hour with a GPU), or worker pools, which have no HTTP endpoint and scale manually or through an external autoscaler. Google's Agent Development Kit (ADK) deploys to Cloud Run with adk deploy cloud_run.
Cloud Run's strength is breadth: an agent can use Cloud SQL, Memorystore, Pub/Sub, BigQuery, and Vertex AI. The trade-off is assembly. Each is a separate product with its own IAM roles, networking, and provisioning, and each environment needs its own copies. Secrets come from Secret Manager, as environment variables or mounted volumes.
On Upsun Cloud, PostgreSQL, Redis, OpenSearch, and other services are declared in .upsun/config.yaml, and an app, worker, or task reaches them through relationships listed in PLATFORM_RELATIONSHIPS. Model API keys are stored as sensitive variables, part of Upsun's secrets and environment configuration.
Cloud Run revision tags give a new revision its own URL without sending it production traffic. Databases and other backing services are not cloned, so a tagged revision uses whatever services it is configured to reach.
On Upsun Cloud, each branch or pull request gets an environment that inherits its parent's data and services, including databases, network storage, and queues. For AI features, that means a prompt, retrieval, or agent change can be tested against realistic data before it merges.
Every Cloud Run instance is sandboxed: first-generation instances use gVisor, and second-generation instances, which jobs and worker pools always use, run as microVMs. Cloud Run sandboxes, in preview since July 2026, run generated code without access to the service's environment variables or the metadata server, at no extra cost.
Upsun Cloud runs workloads in Linux containers with namespace isolation, capability dropping, a seccomp profile, cgroup limits, and network isolation. For untrusted code, Upsun recommends running bubblewrap inside the task; it is not built in, and task containers do not yet offer egress scoping.
Cloud Run runs only on Google Cloud, where it integrates deeply with Vertex AI, Gemini, and IAM. Cloud Run is in scope for Google's ISO 27001, SOC 2, and PCI DSS programs, and Google Cloud's HIPAA Business Associate Agreement covers it.
Upsun Cloud regions run on AWS, Azure, Google Cloud, and OVHcloud, with IBM Cloud through its marketplace. For organizations whose applications and agents are spread across providers, each with its own pipeline, runbook, and audit process, that gives consistent configuration and audit evidence without manually reconciling separate consoles. Upsun holds ISO 27001 certification, a SOC 2 Type 2 report, and PCI DSS Level 1 compliance for its AWS, Azure, and Google Cloud regions; HIPAA workloads run only in the US-4 region.
Cloud Run services use request-based billing, which allocates CPU only while a request is processed and adds a per-request fee, or instance-based billing, which charges for the whole instance lifecycle. Services can scale to zero, and a monthly free tier applies.
Upsun Cloud bills resources per second on their allocation, with CPU and memory set per app, worker, or service. Apps and workers run continuously, while task containers bill only while a run is active.
Use case | Choose |
| Serves a self-hosted model on GPUs | Cloud Run |
| Is built on Vertex AI, Gemini, and other Google Cloud services | Cloud Run |
| Serves spiky web traffic that should scale to zero | Cloud Run |
| Runs spiky background jobs or agent runs | Upsun Clou |
| Is an agent that reads and writes the app's own database | Upsun Cloud |
| Needs each branch tested with the app's data and services | Upsun Cloud |
| Belongs to an organization running apps on several clouds | Upsun Cloud |
| Needs GPU inference and an app with agents | Upsun Cloud for the app, a GPU platform for inference |
Can Upsun Cloud run on Google Cloud?
Yes. Upsun Cloud has regions hosted on Google Cloud, including Switzerland (ch-1), Germany (de-2), and the UK (uk-1), alongside regions on AWS, Azure, and OVHcloud.
Can an Upsun app use Gemini or Vertex AI?
Yes. An app, worker, or task on Upsun Cloud can call Gemini or Vertex AI over their APIs, with credentials stored as sensitive environment variables.
Can Google ADK agents run on Upsun Cloud?
ADK is a code framework, so an ADK agent can run as an Upsun app, worker, or task container that calls Gemini or another model API. The adk deploy cloud_run shortcut is specific to Cloud Run.
Is Upsun cheaper than Cloud Run for AI workloads?
Neither is cheaper for every AI workload: Cloud Run usually costs less for services that are idle much of the time, while the comparison is closer for an always-on app with its own database, workers, and agents.