The Future is Heterogeneous
CPUs, GPUs, TPUs, and domain-specific accelerators are proliferating. Software must treat them as first-class citizens.
Manifesto
Compute should get faster, cheaper, and more portable every time the world invents a new chip.
SF Tensor exists to make high-performance computing feel less like vendor archaeology and more like a mathematical certainty. These are the beliefs we are building from.
Hardware
Heterogeneous by default
Optimization
Automatic and measurable
Access
Fast compute at lower cost
Our Principles
A software layer for the hardware explosion.
The next decade will not standardize on one chip, cloud, kernel language, or accelerator strategy. The winning abstraction has to honor that complexity without passing it to every developer.
Performance Should Be Accessible
Only a handful of experts know how to squeeze every drop of speed out of modern hardware. We believe optimization should be automatic, measurable, and universal.
Cost is a Dimension of Performance
It's not just about FLOPS. It's about FLOPS per dollar. By arbitraging across clouds and hardware, we democratize access to affordable compute.
Mathematics is the Ground Truth
At the end of the day, performance is math. Memory distance, cache locality, algorithmic complexity: all quantifiable, all optimizable.
Code Should Outlive Hardware
Your ideas should not be trapped in CUDA or ROCm or a vendor's silo. A language like Emma lets you write once and inherit every new chip that emerges.
We Build Tools, Not Walled Gardens
Our commitment is to open, extensible, and developer-first infrastructure. We win only when developers do.
If your code can express the math, the system should find the machine.
We are building for people who want infrastructure to disappear without performance disappearing with it.
Train the models only you can build. One stack for enterprise post-training and frontier pre-training.
All System Operational© 2026 San Francisco Tensor Company
