SF Tensor Company
Early-stage startup building cross-cloud GPU orchestration and a GPU-aware compiler for AI training.
On the record
Checkpoint: Series A fundraise decision given roughly 6 to 12 months of runway on the $500K pre-seed
The call
SF Tensor's main hope, the Emma Lang compiler, remains unshipped while better-funded rivals and Nvidia's free tools already occupy the space it needs to win.
SF Tensor builds software to make GPUs run faster and cheaper across different cloud providers, and to automatically tune how code runs on them. It is getting hard to copy because Nvidia released free software called Run:ai that does similar work, and a well-funded startup called SkyPilot (backed by Databricks) just launched with the same idea but way more resources. The one thing that could save SF Tensor is if their unfinished compiler called Emma Lang works much better than what Nvidia and others offer, but they only have 4 people and very little money, so they have maybe six to twelve months to prove it before they run out of cash.
Why
01
AI pressure: Nvidia CUDA, OpenAI Triton, Modular Mojo compilers. These compilers already optimize GPU kernels at scale, leaving little room for the unbuilt Emma Lang to win developers. Could AI do the core job? Yes, likely.
02
Bigger players: Nvidia (via open-sourced Run:ai). Nvidia open-sourced Run:ai's orchestration tech, giving away for free the layer SF Tensor charges for. Could a giant copy it? Yes, easily.
03
What protects it. A savings-based pricing model and early customer relationships, though the GPU price gap it depends on is shrinking.
What would change this call
SF Tensor ships Emma Lang compiler with an independently verified benchmark within 6-9 months. Would show the company built its only defensible asset before running out of runway.
SF Tensor closes a Series A within 6-12 months at current $500K pre-seed burn. Would show investors believe the compiler bet can outrun SkyPilot and Nvidia's Run:ai.
SkyPilot announces named production customers or deeper cloud integrations. Would show the better-funded rival is converting distribution into locked-in usage.
SF Tensor closes a Series A within 6-12 months at current $500K pre-seed burn. Would show investors believe the compiler bet can outrun SkyPilot and Nvidia's Run:ai.
SkyPilot announces named production customers or deeper cloud integrations. Would show the better-funded rival is converting distribution into locked-in usage.
Key numbers
- $500K pre-seed round led by Y Combinator
- Rival SkyPilot raised $20M Series A
- Website traffic down 22.8% QoQ to ~2,000 visits
Named risks
- Core differentiator Emma Lang and AI-driven compiler remain unbuilt as of January 2026, still in hiring phase.
- Only $500K raised against $225K-$315K salaries implies 6 to 12 month runway before needing a new round.
- Thin cross-cloud orchestration layer being commoditized by Nvidia's open-sourced Run:ai and SkyPilot's $20M raise.
- GPU spot prices down ~60% from 2024 peak compress the arbitrage margin underlying SF Tensor's savings-based pricing.
- No independently verified customers or performance benchmarks published as of 2026.
Rivals, as read in July 2026
| Rival | Direction | Evidence |
|---|---|---|
| SkyPilot | gaining on | $20M raise July 21, 2026, live integrations across Nebius, RunPod, AMD, identical pitch |
| Nvidia (Run:ai) | losing to | Open-sourced ~$700M 2024 acquisition, commoditizing GPU orchestration layer |
| Applied Compute | contesting | $80M raised Oct 30, 2025, named customers DoorDash, Cognition AI, Mercor, overlapping elastic compute |
| OpenAI (Triton) | contesting | Owns established open-source GPU compiler competing with unbuilt Emma Lang |
| Modular (Mojo) | contesting | Chris Lattner-led, hundreds of millions in funding, hardware-aware language rival to Emma Lang |
Questions people ask
- Is SF Tensor profitable?
- No. SF Tensor has disclosed no audited revenue as of 2026, and its only funding is a $500K pre-seed round led by Y Combinator in October 2025. The company is pre-revenue and still building its core compiler product as of January 2026.
- Could AI replace SF Tensor's product?
- SF Tensor's own pitch is AI-driven kernel optimization and compilation, but that same layer is being commoditized: Nvidia open-sourced its $700M Run:ai acquisition (2024) and SkyPilot raised $20M in July 2026 with an identical cross-cloud orchestration pitch and stronger distribution.
- Who is SF Tensor's biggest competitor?
- SkyPilot, backed by Databricks co-founder Ion Stoica, raised $20M in a Series A led by Lux Capital on July 21, 2026, targeting the same cross-cloud GPU orchestration niche with far more capital, brand backing, and live integrations across Nebius, RunPod, and AMD.
- Who owns SF Tensor?
- SF Tensor (legally San Francisco Tensor Company) was founded in 2025 by three brothers, Ben, Tom, and Luk Koska, and went through Y Combinator's Fall 2025 batch. California state filings list Ben Koska as CEO, CFO, and Secretary as of April 2026.
- What is Emma Lang and has SF Tensor shipped it?
- Emma Lang is SF Tensor's planned hardware-aware programming language paired with an AI-driven GPU compiler meant to convert PyTorch, JAX, or TensorFlow models into optimized binaries. As of January 2026 job postings, it remains unbuilt and is still in the hiring and development phase.
- How much funding has SF Tensor raised?
- SF Tensor has disclosed $500K total funding, a pre-seed round led by Y Combinator in October 2025, with additional backing named from Susa Ventures and angels including Max Mullen and Paul Graham as of a January 2026 job listing. No valuation has been disclosed.
Timeline
- 2026-07-21Rival SkyPilot raises $20M Series A led by Lux Capital for identical cross-cloud GPU orchestration pitch
- 2026-06-25Crustdata reports SF Tensor at 4 employees, +300% YoY headcount growth, site traffic down 22.8% QoQ
- 2026-04-01California filing lists Ben Koska as CEO, CFO, and Secretary of San Francisco Tensor Company
- 2026-01-01SF Tensor posts founding engineer roles for GPU compiler, kernel, and AI-driven compilation, $225K-$315K base
- 2025-11-04SF Tensor publicly launches on Y Combinator as Fall 2025 (F25) batch company
- 2025-10-30Competitor Applied Compute launches with $80M funding from Benchmark, Sequoia, and Lux
- 2025-10-01SF Tensor closes $500K pre-seed round led by Y Combinator
- 2024-01-01Nvidia acquires Run:ai for approximately $700M, later open-sourcing it
Revision history
27 July 2026
First assessed. Losing its protection. The text above is fixed as of this date and is not edited afterwards.
In the same band
Sources
- Launch YC: SF Tensor - Infrastructure for the Era of Large-Scale AI Training (2025-11-04)
- SF Tensor - Headcount, Funding & Open Jobs (2026-06-25)
- Founding Product Engineer at SF Tensor (2026-01-01)
- SF Tensor - AI - Reviews & Alternatives (2026-04-12)
- Founding GPU Kernel Engineer at SF Tensor (2026-01-01)
- Founding GPU Compiler Engineer at SF Tensor (2026-01-01)
- SF Tensor (2025-11-04)
- SkyPilot, from Databricks' cofounder, raises $20M to be the Switzerland of AI compute (2026-07-21)
- SkyPilot raises $20M to simplify AI compute orchestration across clouds (2026-07-21)
- Former OpenAI researchers launch Applied Compute with $80M in funding (2025-10-30)
- GPU Clusters for Startups: Renting Power on a Budget (2026-07-20)
- San Francisco Tensor Company San Francisco, CA (2026-04-01)