Google Cloud’s Confidential AI Gets a Privacy Upgrade

Google Cloud is making waves in the swamp with a fresh batch of Confidential Computing updates. Think of it as building a secure, private channel through the murky waters of AI—where your data stays protected even while it’s being processed. Let’s dive into what’s new.

Confidential AI at Global Scale

Google Cloud is rolling out Confidential VMs and GKE nodes on the G4 machine series, powered by NVIDIA RTX PRO 6000 Blackwell GPUs. These are available in every region where standard G4 instances live, across multiple consumption models like On Demand and Spot. That means you can run AI inference and fine-tuning with hardware-based privacy guarantees, no matter where you are.

Powered by 5th Gen AMD EPYC CPUs with AMD SEV, these machines create a Trusted Execution Environment (TEE) that encrypts data between CPU and GPU. Perfect for sensitive workloads like private prompts or restricted data—with minimal performance hit.

Open-Source Prompt Encryption SDKs

Even the best TEE can’t protect prompts once they leave your client. That’s why Google Cloud open-sourced Prompt Encryption SDKs on GitHub. These SDKs establish an end-to-end encrypted channel: prompts are encrypted on the client, decrypted only inside the TEE, and responses are encrypted back. It’s like sending a secret message through a secure tunnel—only the intended server can read it.

Get started with the GitHub repo and Codelab.

Apple Private Cloud Compute on Google Cloud

In a surprising alliance, Apple is using Google Cloud to expand its Private Cloud Compute (PCC) service. This leverages Intel TDX, NVIDIA Confidential Computing, and Google’s Titanium security architecture to meet Apple’s stringent privacy requirements. It’s a testament to how far Confidential Computing has come—even the most privacy-obsessed companies are diving in.

Advancing Confidential Foundations

Google Cloud is also beefing up its core infrastructure:

• Intel TDX on C4 machines (preview soon): Hardware-isolated Trust Domains on 6th Gen Intel Xeon processors, with no code changes needed.
• Live Migration for C3D Confidential VMs (GA): Keep your workloads running during maintenance without exposing encrypted memory.
• Confidential Space updates: Independent attestation via Intel Trust Authority (GA) and H100 GPU support (GA) for secure multi-party computation.

Confidential Space now lets multiple parties collaborate on sensitive data—like federated learning—without revealing their data to each other or to Google Cloud. It’s like a secure meeting room where everyone brings their own data but leaves with only the insights.

Next Steps

Confidential Computing is becoming the bedrock of AI privacy. Check out Google Cloud’s Confidential VMs and tools to start building secure AI workflows. And if you’re comparing cloud options, see how Google Cloud stacks up against Vercel or Supabase for your next project.

Original announcement published on Google Cloud.