What's new in Cloud Run

Key Concepts

• Cloud Run: A fully managed, serverless compute platform for deploying containers without infrastructure management.
• Vibe Coding: A paradigm where AI tools (like Gemini) generate full-stack applications from natural language prompts.
• AI Agents: Autonomous systems requiring persistent memory, tool orchestration, and secure sandboxed environments.
• MCP (Model Context Protocol): A standard for connecting AI assistants to systems and data.
• Crema (Cloud Run External Metrics Autoscaling): A mechanism powered by KEDA to scale services based on external events/metrics.
• Ephemeral Disks: Dedicated storage for Cloud Run instances to handle large file manipulation without relying on memory.
• Service Bindings: A private, secure way to route service-to-service communication using automated JWT authentication.

---

1. Vibe Coding and Application Development

The presentation highlights a shift in software development where AI models (Gemini) lower the barrier to entry.

• Process: Users provide prompts to AI Studio, which generates full-stack code (React, Three.js, Express).
• Deployment: Applications are deployed directly to Cloud Run via a "Publish" button, which handles dependency installation, security, and HTTPS endpoint generation.
• Spend Caps: To provide budget predictability for "vibe coders," Google introduced spend caps that automatically pause resources once a defined monthly limit is reached.

2. AI Agents and Sandboxing

Cloud Run is positioned as the ideal runtime for AI agents, providing the necessary "sandbox" for executing untrusted code.

• Cloud Run Sandboxes: Uses ephemeral, isolated micro-VMs to execute agent-generated code or scripts securely, preventing the code from interfering with the host application.
• Cloud Run Instances: A new primitive allowing for long-running, asynchronous background agents. Unlike standard services, these do not scale to zero and provide persistent compute.
• SSH Support: Users can now SSH into Cloud Run containers for advanced troubleshooting, protected by Identity-Aware Proxy (IAP) and OS logging.
• Dev Sync: A tool that synchronizes local folders with a Cloud Run instance, allowing developers to "hot-swap" code in the cloud without losing application state.

3. AI Models and Infrastructure

• GPU Support: Cloud Run now supports NVIDIA RTX Pro 6,000 Blackwell GPUs, designed for high-performance AI inference and fine-tuning.
• Jobs with GPUs: Cloud Run Jobs now support GPUs, enabling cost-effective fine-tuning of open models that scale to zero upon completion.
• Delayed Execution: A feature for jobs that allows scheduling execution within a 12-hour window to optimize costs.

4. Scalability and Enterprise Networking

• Custom Scaling Controls: Provides fine-grained control over minimum and maximum instance limits to balance cost and performance.
• Worker Pools: Now generally available, these are "always-on" instances designed for continuous background tasks (e.g., Temporal workers).
• Crema: Enables scaling based on external metrics (like backlog counts) without requiring a Kubernetes cluster.
• Service Bindings: Simplifies internal networking by injecting service account JWTs, allowing services to communicate securely using simple, context-aware short names.

5. Case Study: Anthropic’s Data Platform

Mimi from Anthropic detailed how they built a security-critical data platform on Cloud Run.

• The Challenge: Managing billions of messages with inconsistent encryption and fragmented access control.
• The Solution: A symmetric architecture using Cloud Run for both ingestion and serving.
  • Ingestion: Encrypts, compresses, and indexes data before writing to Bigtable.
  • Serving: Fetches and decrypts data on-the-fly.
• Key Results:
  • Handles 200,000 QPS at peak.
  • Operates 10,000–20,000 instances without manual infrastructure management.
  • Maintains a cold-start time of under 60 seconds.
  • Security: Plaintext data exists only transiently in memory; no data at rest in plaintext.

---

Synthesis and Conclusion

Cloud Run has evolved from a simple container-hosting service into a comprehensive, serverless ecosystem. By introducing specialized primitives like Worker Pools, GPU-enabled Jobs, and Secure Sandboxes, Google is targeting three distinct growth areas: the "vibe coder" (rapid prototyping), the AI agent developer (long-running, secure compute), and the enterprise (high-scale, secure data processing). The core value proposition remains consistent: zero infrastructure management, pay-per-use pricing, and the ability to scale from zero to massive production loads seamlessly.