Ben Koska on SF Tensor: GPU orchestration for AI training across clouds, $41K revenue in 2 weeks post-launch
Dec 3, 2025 with Ben Koska
Key Points
- SF Tensor generated $41,000 in usage-based revenue within two weeks of launch, with its funding round closing on day one.
- The company orchestrates GPU allocation across Nvidia, AMD, and TPUs for AI training workloads, deliberately excluding inference to focus on an underserved market.
- SF Tensor targets specialized training beyond text models, including protein folding and drug discovery, plus emerging continuous pre-training use cases where enterprises inject proprietary data into foundation model checkpoints.
Summary
SF Tensor is building GPU orchestration infrastructure for AI training runs, sitting above neoclouds and hyperscalers so researchers never have to think about the underlying hardware. Founded by Ben Koska, the company handles GPU allocation across Nvidia, AMD, and TPUs, letting customers focus on what they're training rather than how to provision it.
The pitch is training-only — deliberately. Inference is a solved market with established players; training, Koska argues, remains largely unaddressed infrastructure. The customer base runs from academic and hobbyist researchers training small models up to startups that have raised hundreds of millions and are building specialized models for drug discovery, protein folding, and other non-text domains.
The structural bet
SF Tensor's thesis isn't that every company will train a frontier LLM. It's that text is one modality, and most real-world problems require something else. Protein folding, text-to-speech, and domain-specific scientific models can't be solved by prompting GPT-4. The market for specialized training infrastructure follows however many of those use cases get productized.
A second growth vector is emerging: AWS recently launched a product that ships a foundation model checkpoint 80% through pre-training, letting companies inject proprietary data before the final training stages rather than relying on fine-tuning or prompt injection. Koska sees continuous pre-training as an underexplored area that could pull a much broader set of enterprises into training workloads, and SF Tensor is positioned to serve that demand regardless of what's being trained.
Early traction
Two weeks after launch, SF Tensor had generated $41,000 in usage-based revenue. The round closed on the first day of fundraising, with at least one prominent angel backing the company — the identity wasn't disclosed, but the investors on the show noted receiving a direct text recommending the deal.