Beyond GPUs: why we built Tsavorite
By Shalesh Thusoo, Founder and Chief Executive Officer
Today we are launching our updated website, and with it this blog. I want to use the first post to explain why we started Tsavorite, what we are building, and why the timing matters.
Our view is that AI has outgrown the way the industry builds computers for it. Adding more of the same hardware is no longer the answer. AI needs an architecture designed for it from the start, one that becomes more efficient as it grows rather than less.
Yesterday: scaling by adding GPUs
From about 2018 to 2023, the industry’s answer to bigger models was more GPUs. They were the best tool available, and connecting thousands of them made the era of large language models possible.
That approach carries a cost that grows with size. As clusters get bigger, more time and energy go into moving data between chips and waiting for it, and less into useful computation. Utilization falls as the cluster grows, so each added GPU buys less useful work than the one before.
Today: a different workload
From 2023 to today, the workload itself has been changing. Inference now runs around the clock, AI agents hold long multi-step sessions, and physical AI in robots and machines has to respond in real time.
These workloads are bound less by raw compute than by memory capacity, memory bandwidth, and power. What matters is how much useful work a system delivers per watt and per dollar at the scale a customer actually runs, and a peak FLOPS number on a datasheet does not answer that.
Where AI runs is changing as well. Enterprises want large models working on their own data inside their own walls, and machines at the edge cannot wait on a round trip to the cloud.
Tomorrow: composable by design
We founded Tsavorite on the belief that AI infrastructure from 2027 onward will be composed from coherent building blocks instead of assembled from discrete accelerators. That belief became the Omni Processing Unit (OPU), a coherent hardware and software architecture for every AI workload, from the edge to exascale.
Three parts define it. AI OmniCore is the compute engine: thousands of efficient AI Omnicores closely coupled with Arm cores, that run any AI model with confidential computing built in. MultiPlexus is the coherent fabric; it connects cores on a die, chiplets in a package, packages in a system, and systems across racks into unified memory pods. TAOS, our software stack, is PyTorch-first and CUDA-compatible, so developers bring their models without rewriting them.
Because MultiPlexus keeps data close to compute, utilization climbs as systems grow instead of falling, and scale becomes a feature rather than a tradeoff. The same architecture runs in a compact Helix-M at the edge, a Helix-D on a desk, and Helix-R in the datacenter rack.
Why this team
We have built this kind of technology before. Our leadership has engineered CPUs, accelerators, interconnects, and software stacks that the industry runs on today, at top companies including Intel, Cisco, Apple, AMD and others.
Several of us founded Tanzanite Silicon Solutions, which Marvell acquired in 2022. Tsavorite applies that experience to a larger problem, with engineering centers in Silicon Valley, California and Bangalore, India.
What comes next
This blog is where we will go deeper. In the next post, our CTO Mike Uhler explains how the OPU architecture works, and posts on TAOS and the Helix systems will follow.
If you are building AI infrastructure and running into the limits I have described, we would like to talk with you. If you want to help build what comes next, we are hiring.