Tsavorite Scalable Intelligence, a start-up based in Milpitas, California, introduced its Omni Processing Unit (OPU), a new composable compute architecture designed to unify CPU, GPU, memory, and interconnect in a single device. The OPU targets the full spectrum of AI workloads—from training and reinforcement learning to inference and emerging agentic workflows—promising significant gains in performance, power efficiency, and scalability.
The company’s OPU integrates with its proprietary MultiPlexus fabric, which links chiplets, packages, systems, and racks into a unified compute domain. MultiPlexus provides petabyte-scale bandwidth, distributed gigabyte caches, and ultra-low latency interconnects, creating a cohesive high-efficiency platform that scales from edge to hyperscale. The architecture supports standard AI frameworks and CUDA-based workflows through Tsavorite’s Agentic Operating Stack (TAOS), eliminating software migration barriers. With production silicon and Helix AI appliances expected in 2026, Tsavorite reported more than $100 million in pre-orders from global enterprises and cloud providers.
Built on Arm Neoverse cores and fabricated using Samsung Foundry’s SF4X process, the OPU employs a modular chiplet design to enable flexible power-performance tuning across configurations. Industry partners—including Eviden (Atos Group), Sumitomo Corporation, Zscaler, and Presidio Ventures—have endorsed Tsavorite’s sustainable and scalable approach to AI compute. Early customers are testing prototype systems using TAOS for full agentic workflow deployment, validating the system’s interoperability and performance claims.
• Omni Processing Unit unifies CPU, GPU, memory, and interconnect in one device
• MultiPlexus fabric enables petabyte-scale bandwidth and distributed gigabyte caches
• Arm Neoverse–based compute subsystem optimized for power efficiency and scalability
• TAOS software stack supports PyTorch, vLLM, Triton, HuggingFace, Ray, and Kubernetes
• Samsung Foundry’s SF4X process to produce production silicon in 2026
• $100M in pre-orders from Fortune Global 500, sovereign cloud, and system integrators
• Partners include Eviden, Sumitomo, Arm, Samsung Foundry, Zscaler, and Presidio Ventures
• Unified petabyte-scale memory improves large-model performance
• Prototype systems already in customer evaluation and multiple design-ins underway
• Target markets: agentic AI, inference, reinforcement learning, and edge-to-cloud deployments
“With the Omni Processing Unit and our MultiPlexus fabric, we’ve built the first truly composable, developer-friendly AI platform that delivers step-change gains in efficiency, cost, and scale from edge to hyperscale,” said Shalesh Thusoo, Founder & CEO of Tsavorite.
🌐 Analysis: Tsavorite enters the crowded AI silicon market with a bold composable architecture that merges compute, memory, and interconnect into a unified domain—an approach reminiscent of Cerebras’ wafer-scale and Tenstorrent’s chiplet-based designs, but with greater emphasis on rackscale composability and open software compatibility. Backed by Sumitomo and Samsung Foundry, Tsavorite’s OPU aims to address the escalating energy and scaling limits faced by GPU-centric architectures. Its $100 million in pre-orders signals early confidence, though broad ecosystem adoption will hinge on delivering production silicon and maintaining CUDA interoperability through TAOS in 2026.
🌐 We’re tracking the latest developments in semiconductors. Follow our ongoing coverage at: https://convergedigest.com/category/semiconductors/
🌐 We’re launching the “Data Center Networking for AI” series on NextGenInfra.io and inviting companies building real solutions—silicon, optics, fabrics, switches, software, orchestration—to share their views on video and in our expert report. To get involved, send a note to jcarroll@convergedigest.com or info@nextgeninfra.io
