
1. Technology overview
San Francisco Tensor Company, trading as SF Tensor, is a San Francisco company in the Y Combinator F25 batch, founded in 2025 by Ben, Tom and Luk Koska. It targets AI labs and enterprises that need to train models without managing the infrastructure layer.
Its Model Foundry optimizes kernels automatically, finds the cheapest GPUs across providers and chip types (NVIDIA, AMD, TPU, Trainium), and recovers jobs when spot instances fail. The site describes scaling from 1 to 10,000 GPUs without code changes, and an option with embedded post-training researchers for enterprises training on private data.
2. Highlighted features & specifications
- Kernel optimizer: Searches for faster kernels as the architecture changes.
- Multi-silicon scheduling: Prices compute across NVIDIA, AMD, TPU and Trainium.
- Automatic recovery: Resumes jobs when spot capacity fails.
- Frontier scale: Runs from 1 to 10,000 GPUs without code rewrites.
- Post-training support: Evals, SFT and RL on customer data.
- Deployment options: In the customer's cloud or the vendor's.
3. Potential application
A Vietnamese AI lab can train a Vietnamese-language model across mixed GPU providers at lower cost. A bank can build a domain model on private transaction text with an embedded research team. A university group can scale experiments without a dedicated infrastructure engineer.
4. Benefits
Pricing is a share of the savings, so nothing is owed if the bill does not fall, according to the site. Failure recovery is automatic and no dedicated infrastructure team is needed. The site cites a first place on an NVIDIA benchmark and a production model in about six weeks for enterprise engagements.
5. Market trends & opportunities
Training costs and GPU availability limit AI research and domain-model projects. Demand is rising for specialist models trained on private data. Vietnamese banks, telcos and AI labs exploring local-language models may consider outcome-priced training support.
6. Other information
Y Combinator F25 batch, San Francisco per YC profile, founded 2025, 6 people. Engagements are scoped per customer and priced on savings or fixed engagement; part of the offer includes embedded researchers.
One platform connecting startups, corporations, research institutions, regions and investors











Playbook & Handbook · February 6, 2026
Winning Accelerator Playbook 2026 by BambuUP
RETAIL E-COMMERCE: AI AND REAL-TIME DATA ARE TRANSFORMING HOW RETAIL BUSINESSES OPERATE








































