Semiconductor startup Tsavorite Scalable Intelligence, Inc. is unveiling a new compute architecture aimed at artificial-intelligence workloads: the so-called Omni Processing Unit (OPU), which integrates GPU, CPU, memory and connectivity in a composable chiplet form. The company claims this removes performance and cost bottlenecks common in conventional AI infrastructure.
A key innovation in Tsavorite’s approach is its “MultiPlexus” interconnect fabric, which enables chiplets to be deployed at various scales—from individual devices at the edge to large server racks in the cloud. The OPU is built on top of Arm Neoverse compute subsystems, bringing in Arm’s high efficiency per watt alongside Tsavorite’s modular architecture.
On the software side, Tsavorite introduces the open-source stack called TAOS (Tsavorite Agentic Operating Stack), which enables developers to run existing CUDA-based workflows (originally built for Nvidia Corporation) across Kubernetes, PyTorch, Ray, Triton and HuggingFace frameworks with minimal changes. The goal is to ease migration and adoption of the new hardware platform.
The company says it has already secured over $100 million in pre-orders from major customers, including enterprises and sovereign cloud infrastructure providers across the U.S., Europe and Asia. Production of the OPU and a Helix AI appliance is targeted for early next year, marking Tsavorite’s entry into a rapidly evolving AI compute market.
