SF Tensor builds infrastructure that lets AI labs focus on research rather than DevOps. Its managed platform offers automatic GPU kernel optimization across NVIDIA, AMD, and Vulkan, finds the cheapest compute across cloud providers, handles spot-instance migration, and scales distributed training from 1 to 10,000 GPUs without code changes. The founders claim the solution can cut compute costs by up to 80%. Founded by brothers Ben, Luk, and Tom Koska, the company is part of Y Combinator's Fall 2025 batch.
SF Tensor
ActiveInfrastructure for AI labs to focus on research
Total raised
$500K
1 round
Stage
Seed
Jan 2025
Team
1-10
since 2025
Pricing
—
Founded
2025
San Francisco, United States
Agent-ready
—
Managed AI infrastructure so labs can focus on research over DevOps
Automatic GPU kernel optimization across NVIDIA, AMD, and Vulkan
Cheapest-compute routing across multiple cloud providers
Automated spot-instance migration to reduce cost
Distributed training scaling from 1 to 10,000 GPUs
Scaling without code changes to existing training jobs
Cross-hardware support spanning multiple GPU vendors
Backed by Y Combinator (Fall 2025 batch)
12/100
Early
MCP server
Public API
Webhooks
OAuth 2.0
SDKs
No public agent surfaces detected yet.
Jan 2025 Seed $500K ● Y Combinator
Capital network
$500K raised ·1 backer·10 network links
- Backers1
- Shared portfoliocompanies these backers also fund
- Extended networkfunds that co-invest alongside them
Databricks
The data + AI company
AI AgentsAI Infrastructure
Figure AI
General-purpose humanoid robots
AI InfrastructureAI Robotics
Upscale AI
Pure-play AI networking infrastructure
AI Developer ToolsAI Infrastructure
Dash0
AI-native observability platform built on OpenTelemetry
AI InfrastructureAI Data Engineering
Noma Security
End-to-end security for agentic AI
AI InfrastructureAI for Cyber Defense
Ineffable Intelligence
An AI research company building a superlearner to achieve superintelligence through reinforcement learning
Foundation ModelsAI Infrastructure
- What hardware does SF Tensor support?
- It offers automatic GPU kernel optimization across NVIDIA, AMD, and Vulkan.
- How does it reduce compute costs?
- It finds the cheapest compute across cloud providers and handles spot-instance migration; the founders claim up to 80% savings, which varies by workload.
- Do I need to change my code to scale?
- No. SF Tensor scales distributed training from 1 to 10,000 GPUs without code changes.
- Who built SF Tensor?
- It was founded by brothers Ben, Luk, and Tom Koska and is part of Y Combinator's Fall 2025 batch.
Discussion
Sign in to join the discussion.
Sign inExplore more around SF Tensor
Contextual paths to related AI startups, deals and rankings.