Duality & Google Cloud launch confidential AI with GPU power

Duality Technologies has announced support for Google Cloud's Confidential Computing portfolio, including NVIDIA GPU-powered confidential virtual machines, enabling enterprises to process sensitive artificial intelligence workloads securely at scale.

Confidential AI

The latest capability from Duality allows customers to deploy large language model (LLM) training and inference as well as encrypted retrieval-augmented generation (RAG) inside confidential environments powered by NVIDIA H100 GPUs on Google Cloud. Previously, confidential AI was limited to CPU-only instances, which limited performance and scalability.

By combining full-stack data confidentiality with GPU acceleration, enterprises can process private or regulated data with performance levels suitable for production environments, addressing key concerns in sectors such as finance, defence and healthcare.

Performance gains

GPU-backed confidential computing represents a performance increase over earlier solutions. Enterprises are able to meet the latency and throughput requirements needed for large-scale AI applications, such as searching and summarising confidential documents or extracting insights without exposing the underlying data.

"This changes the game. Our customers can now run privacy-preserving AI with LLMs at production scale. With GPU acceleration, the performance bottlenecks of secure computing are gone-making secure LLM training and inference practical," said Dr. Alon Kaufman, CEO and Co-Founder, Duality Technologies.

Trusted environments

The solution leverages Google Cloud's Confidential Space and NVIDIA H100-powered confidential virtual machines, with additional support for Intel TDX and integration with Cloud Key Management Service. Duality has successfully validated the platform by running a Mistral-7B model using encrypted vector RAG in a fully confidential pipeline.

Workloads are run entirely within trusted execution environments (TEEs), which reduce the risk of data leakage during processing. The integration is now available to enterprises through Google Cloud's Dynamic Workload Scheduler within Confidential Space, meant to support a range of regulated industries and AI-native companies.

Enterprise focus

According to a recent survey by financial institution UBS, almost half of respondents cited compliance and regulatory issues as the main barrier to adopting AI. By providing scalable, confidential computing options, the Duality solution aims to address these regulatory and privacy constraints.

"With Confidential GPUs, organizations can process sensitive AI workloads entirely within trusted execution environments without giving up performance. Pairing NVIDIA H100-powered confidential VMs with Duality's encrypted workflows allows LLM training and inference to happen at scale, with end-to-end protection from data leakage," said Nelly Porter, Director of Product Management, Google Cloud.

Industry impact

The launch follows industry demand for confidential AI solutions that meet enterprise security requirements. The initial rollout is available on the Google Cloud Confidential A3 virtual machine type in preview. Broader availability is expected to follow later in the year.