confidential.ai

What It Is

Confidential AI positions itself as a “confidential computing stack for AI.” Its core focus is not simply providing a large-model chatbot, but running inference, training, and agent workloads inside hardware-encrypted trusted execution environments (TEEs). Its value proposition is that data and model weights remain private, tamper-resistant, and verifiable during processing.

Core Capabilities

On the AI side, the platform covers private inference, confidential agents, weight protection, and private training. Private inference emphasizes that customer prompts, responses, and model interactions remain invisible. Confidential agents isolate credentials such as tokens and API keys inside the TEE. Weight protection is designed to prevent proprietary weights from being extracted during inference or fine-tuning. Private training focuses on cryptographically proving which data was used. Listed inference models include the GLM, Qwen, and DeepSeek series, with other models available by request.

Pricing and Integration

Pricing is pay-as-you-go. Inference is charged per million input/output tokens. For example, DeepSeek V4-Flash is $0.20/$0.40, Qwen 3.5 35B is $0.25/$2.00, and DeepSeek V4-Pro is $2.50/$5.00. Enterprise plans, on-prem deployment, and high-volume usage require contacting sales. The platform also provides a Confidential Agents REST API, along with documentation for confidential Kubernetes and provable builds, suggesting that it is positioned more as an engineering infrastructure product.

Pros and Cons

Its strengths are a clearly defined security boundary and coverage across data, credentials, model weights, and the training process, making it suitable for AI use cases with strong compliance requirements. Some inference pricing is publicly available, which also helps with initial cost estimation. Limitations include the lack of disclosed free trial, SLA, customer case studies, compliance certifications, specific hardware, and available regions in the main materials. It also does not explain the impact of TEE on latency and throughput. GPU VM pricing information is likewise incomplete.

Who It’s For and China Access

It is suitable for AI labs, infrastructure providers, finance/healthcare/internal enterprise AI teams, and developers who need to protect proprietary model weights or sensitive data. Chinese-language support is not clearly stated, but the model list includes Qwen, DeepSeek, and GLM, so there may be a model foundation for Chinese-language tasks. Access from mainland China, payment methods, and available regions are not specified in the main materials, so they need to be tested directly or confirmed with the official team.