Articles
Research
Reverse-Engineering the B200's Tensor Core, One Bit at a Time
We formally prove optimized ML kernels correct instead of testing them: symbolic execution shows a FlashAttention tile is equivalent to naive attention at d=64
July 11, 2026
Research
One Model Per Customer Shouldn't Mean One GPU Per Customer
Every customer wants their own model. Here's how to hand them one without buying a GPU per customer.
May 17, 2026
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
