SF Tensor GTM brain

SF Tensor provides managed model training infrastructure and kernel optimization for AI labs and enterprises building proprietary models.

category

AI training infrastructure

market profile

  • text: Model training teams face high compute costs and distributed infrastructure maintenance that compete with research time.
  • text: SF Tensor positions itself across GPU and alternative accelerator hardware, from experiments to large training runs.
  • text: Model Foundry is priced as a share of savings; a separate embedded research engagement has a fixed fee plus compute. Neither maps cleanly to the available pricing models.

personas

  • title: Head of AI research
    pains:
    • Researchers spend weeks debugging distributed jobs rather than testing architectures
    • Limited compute budget restricts the number of model experiments
  • title: ML infrastructure lead
    pains:
    • Recovering long training jobs after hardware failures
    • Tuning kernels and managing capacity across providers
  • title: Enterprise AI leader
    pains:
    • Building a private domain model without assembling a large training team
    • Keeping model and data ownership while moving quickly

icp

  • text: AI labs running pre-training, post-training or reinforcement learning on substantial accelerator capacity.
  • text: Enterprises with private domain data and a goal to train or adapt a model they own.

message pillars

  • text: Free research teams to focus on model outcomes.
  • text: Package the infrastructure around the training outcome.
  • text: Use multiple hardware families without rewriting the model.
  • text: Let enterprises retain the assets they build.

objections

  • objection: Will migration to unfamiliar hardware slow our research?
    response: Benchmark the actual training workload on target hardware and compare throughput and recovery behavior before scaling.
  • objection: Will a provider own our proprietary model or data?
    response: Review deployment and ownership terms for the proposed engagement, including private cloud options.

competitors

  • name: CoreWeave
    why: Provides GPU cloud clusters for AI training workloads.
  • name: Together AI
    why: Offers GPU clusters and training or fine-tuning services.

keywords

  • distributed GPU training platform
  • multi cloud GPU training
  • AI training kernel optimization
  • model training infrastructure
  • GPU cloud cost optimization
  • private model training

voice

  • technically confident
  • researcher oriented
  • outcome focused

signals

  • text: A lab announces a large model training run or hires distributed training engineers.
  • text: An enterprise is building domain models on proprietary data.

stack

  • NVIDIA
  • AMD
  • TPU
  • Trainium

hiring

  • Member of Technical Staff, GPU Compiler
  • Member of Technical Staff, GPU Kernels
  • Member of Technical Staff, Post-Training & Applied Research

pricing

model: unknown

motion

sales_led

channels

  • LinkedIn
  • engineering blog
  • YC launch

confidence

high

Founders

  • name: Ben Koska
    title: Founder/CEO
    role: CEO
    bio: Optimizing GPU Kernels @ SF Tensor.
  • name: Luk Koska
    title: Founder/CTO
    role: CTO
    bio: Optimizing AI infrastructure @ SF Tensor. Finished BSc in CS @ age 16
  • name: Tom Koska
    title: Founder
    role: Founder
    bio: Optimizing AI infrastructure @ SF Tensor (F25). Finished BSc in CS @ age 16

Back to YC companies