Tsavorite Scalable Intelligence Sets New Frontier for AI Efficiency and Scale with the Omni Processing Unit

Integrated, Composable Platform Delivers Dramatic Improvements in Performance, Cost and Scalability for Training, Reinforcement Learning, Inference and Agentic Workflows

MILPITAS, Calif.-- November 10,2025 --Tsavorite Scalable Intelligence, a pioneering AI computing solutions provider, today announced its groundbreaking composable, compute architecture that sets a new industry standard for performance, scalability and affordability for AI workflows. Tsavorite’s Omni Processing Unit™ (OPU) unifies CPU, GPU, memory and scale up/scale out connectivity in a single device. The OPU can be configured for markets and applications with different power, performance and scaling requirements.

“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. “With production silicon and Helix™ AI appliances on track for 2026 and more than $100M in pre-orders, we’re turning the promise of agentic AI into practical, deployable infrastructure.”

“The semiconductor market is entering a new phase of architectural innovation driven by AI and data-intensive applications,” said Karl Freund, Founder and Principal Analyst at Cambrian-AI Research. “With prototype systems already in customer evaluation and multiple design-ins, Tsavorite’s chiplet-based architecture demonstrates real-world value and the potential to improve performance and efficiency at scale.”

Multiple Breakthrough Innovations

Tsavorite’s OPU is built on several industry-first technologies developed in-house by the company’s veteran team:

• MultiPlexus™ fabric: Tsavorite’s revolutionary interconnect spans from die to rack, offering petabyte-scale bandwidth and gigabyte-scale caches with ultra-low latency.
• Modular chiplet architecture: With flexible scaling and form-factor customization, Tsavorite’s OPU delivers optimized performance across edge, on-prem, and cloud deployments.
• Unified petabyte-scale memory: High-bandwidth memory across configurations, radically improves performance of large model operations.
• Simplified, seamless programmability: Tsavorite’s full-stack, open software platform allows developers to seamlessly run existing CUDA-optimized workflows on Tsavorite systems with zero switching cost.

“Tsavorite’s composable AI architecture embodies the innovation required to build a more sustainable digital future,” said Eisuke Takenaka, Executive Officer, Sumitomo Corporation. “By combining Tsavorite’s breakthrough design with Sumitomo’s global reach and strong business development capabilities, we aim to accelerate the deployment of energy-efficient AI infrastructure that drives economic growth while reducing environmental impact.”

“We’re pleased to see Tsavorite’s new Omni Processing Unit entering the market, offering a fresh and innovative approach to addressing the growing complexity of AI workloads,” said Eric Eppe, CTO for Advanced Computing and AI at Eviden, Atos Group. “By combining compute, acceleration, and memory into a single device, this innovation holds strong potential to enhance performance across learning, inference, and emerging agentic workflows. At Eviden, we remain committed to exploring and supporting all advancements that contribute to the evolution of AI technologies and help our clients navigate this dynamic landscape.”

“Enterprises are increasingly focused on achieving greater AI performance while optimizing the power available to them and keeping costs under control,” said Michael Kuperman, Chief Reliability Officer & SVP Cloud Operations, Zscaler, Inc. “Tsavorite’s approach to high-efficiency computing aligns with the broader industry shift toward sustainable, scalable AI infrastructure.”

Flexibility and scalability by design

Founded by chip and software industry veterans with deep experience building products and companies, Tsavorite has developed a completely novel approach to overcoming the performance and cost bottlenecks that hinder AI Implementations. Tsavorite’s chiplet architecture can be adapted to use cases across edge, on-prem, data center or cloud. The company’s MultiPlexus™ AI fabric, the industry’s first AI compute catalyst, connects chiplets, packages, systems and racks – up to thousands of devices – with ultra-low latency, distributed gigabyte caches and petabyte-scale bandwidth to form unified memory pods that maximize compute utilization. This unified fabric forms a single, high-efficiency compute domain that delivers unmatched performance across inference, training, fine-tuning, reinforcement learning, and agentic workflows.

Tsavorite’s OPU includes a compute subsystem built on Arm® Neoverse®, delivering high performance-per-watt and mature software compatibility across deployment environments.

“The growing complexity of AI workloads is pushing the industry toward purpose-built platforms that can scale efficiently,” said Dermot O’Driscoll, vice president of product solutions, Cloud AI Business, Arm. “By building on Arm Neoverse, Tsavorite is delivering a platform that brings together their own specialized innovation and Arm’s proven high performance and energy efficiency, enabling AI solutions that scale from edge to hyperscale.”

“We’re excited to collaborate with Tsavorite to bring its innovative Omni Processing Unit to life on our advanced SF4X platform solution,” said Margaret Han, Executive VP and Head of US Foundry at Samsung Electronics. “Tsavorite’s composable architecture is a remarkable example of how advanced design and Samsung Foundry’s leading process technology can come together to redefine efficiency and scalability in AI computing.”

Developer friendly from day one

With Tsavorite’s Agentic Operating Stack (TAOS) developers can deploy models instantly without code rewrites, quantization adjustments, or proprietary software dependencies. TAOS natively supports leading frameworks including PyTorch, vLLM, Triton, HuggingFace, Ray and Kubernetes enabling effortless integration with existing AI pipelines.

Customers have validated Tsavorite’s hardware and software platform, using the TAOS SDK to run AI models and end-to-end agentic workflows.

Availability and early market traction

Tsavorite expects to deliver silicon and Helix™ enterprise class AI appliances in 2026 that will be capable of supporting all agentic AI workflows with leadership performance.

Tsavorite has more than $100 million in pre-orders from customers, including Fortune Global 500 companies, sovereign cloud providers, and leading systems integrators across the U.S., Asia, and Europe, demonstrating strong product-market fit and customer confidence.

"Presidio Ventures is committed to investing in technologies and teams that have the potential to transform industries. For the world to truly experience the magnificence of what artificial intelligence can offer, it is essential to develop transformative solutions that address legacy challenges including power usage, scalability, and cost; factors which have traditionally limited the widespread adoption and true impact of AI,” said Masood Pirzada, CEO, Venture Investments at Presidio Ventures. “Tsavorite is poised to unlock a real, affordable takeoff velocity AI compute silicon to the masses, enabling all to access unprecedented computing power without prohibitive costs or constraints."

"Across diverse applications like humanoid robots, autonomous vehicles and datacenter-scale text and video generation, the next wave of progress in AI is unlikely to come from ever-larger monolithic chips but from modular systems tuned to each workload,” said Austin Lyons, Senior Analyst, Creative Strategies Inc. “Tsavorite’s chiplet-based architecture follows that direction, combining compute, memory, and AI processing into composable systems that let customers optimize for what increasingly matters most: tokens per second per watt.”

"Tsavorite enters the market at a time when AI architectures are moving from homogeneous GPU-based servers to heterogeneous rackscale systems optimized for emerging inference workloads,” said Bob Wheeler, Principal Analyst, Wheeler’s Network. “At the same time, millions of developers use the CUDA stack for AI-software development. Tsavorite aims to bridge the gap between the development community and the imperative to improve generated tokens per second per watt.”

“No one is better suited to creating and commercializing an entirely new approach to AI compute than Tsavorite’s deeply experienced team,” said Atiq Raza, Board Member, Tsavorite Scalable Intelligence, Inc. “It’s rare for a chip startup to emerge from stealth having gotten as far along in development and with as much early customer interest as Tsavorite. It’s a reflection of the team’s vision, technical competence and their ability to execute.”

About Tsavorite Scalable Intelligence

Founded in 2023 with simultaneous operations in Milpitas, California and Bangalore, India, Tsavorite Scalable Intelligence is developing the semiconductor industry's first Omni Processing Unit (OPU), delivering breakthrough performance for AI inference, training, and reinforcement learning. Representing a paradigm shift in AI compute, the OPU and Tsavorite’s integrated open software stack are designed from the ground up to eliminate the cost, efficiency, and complexity obstacles that limit the adoption and deployment of AI for edge, enterprise, and data center applications.

Complementing its U.S. headquarters, Tsavorite’s India Design Center plays a critical role in co-developing the company’s next-generation hardware, software, and system-level solutions. The center underscores Tsavorite’s commitment to global innovation and collaboration. With a substantial and rapidly growing engineering presence, Tsavorite’s India Design Center has become a cornerstone of its worldwide R&D network, accelerating the development of intelligent, efficient compute systems.

For more information, visit: https://www.tsavoritesi.com