Zima Security Architecture

What is Zima?

Zima is a secure AI inference API with zero data retention. It provides OpenAI compatible endpoints for over 100 AI models including GPT, Claude, Gemini, Llama, Mistral, DeepSeek, Qwen, and more. Developers can switch their API base URL and key with no code changes needed.

Zima is designed for companies and developers who need to use AI models without exposing sensitive data to model providers or cloud infrastructure operators.

How Zima Protects Your Data

Zima provides two levels of protection depending on the model type:

Open Source Models (Llama, Mistral, DeepSeek, Qwen)

Open source models run directly on Zima encrypted hardware infrastructure. This uses Intel TDX (Trust Domain Extensions) and NVIDIA Confidential Computing to encrypt data at the processor and GPU level. Your prompts and completions are encrypted even while being processed. No one, including Zima infrastructure operators, can access the data during execution. All requests have zero data retention.

Centralized Models (GPT, Claude, Gemini)

Centralized models from OpenAI, Anthropic, and Google are routed through Zima with zero logging and zero data retention. These models do not run on Zima secure hardware because they are hosted by their respective providers. However, Zima ensures that no request data is stored, logged, or used for training on the Zima side.

Confidential Computing Technology

Confidential computing refers to hardware security features that encrypt data while it is being processed, not just at rest or in transit. Traditional cloud computing only encrypts data when it is stored (at rest) or when it is being sent between systems (in transit). Confidential computing closes the gap by encrypting data during computation.

Intel TDX (Trust Domain Extensions)

Intel TDX creates isolated execution environments called Trust Domains at the CPU level. Code and data inside a Trust Domain are encrypted by the processor hardware and cannot be accessed by the host operating system, hypervisor, or other virtual machines. Zima uses Intel TDX to isolate AI model inference from the rest of the infrastructure.

NVIDIA Confidential Computing

NVIDIA Confidential Computing extends hardware encryption to the GPU. AI model inference requires GPU processing, and without GPU level encryption, data would be exposed during the computation phase. NVIDIA Confidential Computing ensures that data sent to the GPU for inference remains encrypted throughout the entire processing pipeline.

Cryptographic Attestation

Before any data is processed, the hardware generates cryptographic proofs that verify the system is running approved software. This attestation process ensures that no unauthorized code can access your data. The system continuously verifies its state during execution and blocks access if verification fails.

Zero Data Retention Policy

Zima enforces zero data retention across all models and all plan tiers. This means:

• Your prompts are never stored after the request completes
• Model outputs are never logged or cached
• Your data is never used to train or fine tune any model
• No request metadata is retained beyond billing records

This applies to both the API and the Zima Chat interface. Whether you are using the free tier or an enterprise plan, zero data retention is the default.

Frequently Asked Questions

Does Zima store my prompts or completions?

No. Zima enforces zero data retention. Your prompts and model outputs are never stored, logged, or used for training.

Can Zima infrastructure operators see my data?

For open source models running on Zima encrypted hardware, no. Intel TDX and NVIDIA Confidential Computing ensure that data is encrypted even from infrastructure operators. For centralized models, Zima does not log or store any data, but the model providers (OpenAI, Anthropic, Google) operate their own infrastructure.

Is Zima compatible with the OpenAI API?

Yes. Zima provides OpenAI compatible API endpoints. You can switch your API base URL and key in any OpenAI SDK or framework with no other code changes.

What is the difference between Zima and using OpenAI directly?

When you use OpenAI directly, your prompts and completions may be logged and used for model improvement. With Zima, your data is never stored. For open source models, Zima adds hardware level encryption so data is protected even during processing.