Is SF Tensor profitable?

No. As of the most recent disclosed round in late 2025, SF Tensor has raised only $500K total and reports no revenue, ARR, or named customers. Its website shows a demo pricing widget rather than customer logos, indicating pre-revenue status as of mid 2026.

Who is SF Tensor's biggest competitor?

NVIDIA's DGX Cloud Lepton, launched via the April 2025 Lepton AI acquisition and rebranded at Computex in May 2025, directly competes on cross-cloud GPU arbitrage. Better-funded peers Modal ($1.1B valuation, Sep 2025) and Applied Compute also target the same segment.

Could AI replace SF Tensor's product?

SF Tensor's core offering, automated kernel optimization and compiler tuning, is exactly the systems work frontier AI code-generation research increasingly automates internally. No proprietary data or workflow lock-in has been publicly verified to insulate the product as of mid 2026.

SF Tensor is privately founder-owned, started in 2025 by three brothers, Ben, Tom, and Luk Koska. It is backed by Y Combinator and Susa Ventures, plus angel investors including Max Mullen, Paul Graham, and executives from Neuralink, Notion, and AMD, per its own job listings.

What does SF Tensor actually build?

SF Tensor sells infrastructure so AI labs can train models without managing compute: a Kernel Optimizer for NVIDIA/AMD/TPU hardware, 'Tensor Cloud' for cross-cloud GPU sourcing up to 10,000 GPUs, and 'Emma Lang,' a hardware-aware programming language, launched publicly on YC on November 4, 2025.

Has SF Tensor proven its compiler beats CUDA?

No independent verification exists as of June 2026. The company claims (vendor-stated, unverified) 'up to 80% compute cost savings' from its kernel optimizer, but no third-party benchmark, case study, or named design partner confirming this against hand-tuned CUDA has been published.

YC Fall 2025 · United States

SF Tensor

Early-stage startup selling a cross-cloud GPU cost and kernel optimization layer for AI training.

Protection under test

Assessed 27 July 2026 · Evidence as of 27 July 2026

Something worth protecting exists and is being tested right now.

On the recordSince 27 July 2026 · Checkable from H2 2026 · Every call on the record Checkpoint: Likely Series A or bridge round given current burn against $500K raised

The call

SF Tensor is chasing a real gap in cross-hardware AI infrastructure but is racing much better-funded players offering similar products today.

SF Tensor makes tools that let companies run artificial intelligence work on different types of computer chips from different cloud providers, instead of being locked into NVIDIA's dominant chips. The problem is that NVIDIA, AWS, Google, and other big companies are already building similar tools themselves, and they have far more money and engineers than SF Tensor's tiny four-to-six person team. The one thing that could change this is if SF Tensor finds a major customer or partner who publicly says their product works better than the alternatives, before they run out of money and have to raise more funds in the next six months.

Why

AI pressure: Frontier AI labs' internal auto-tuning and AI-assisted kernel/compiler generation. Labs increasingly automate kernel tuning in-house, shrinking demand for a third-party optimization layer like SF Tensor's. Could AI do the core job? Yes, likely.

Bigger players: NVIDIA (DGX Cloud Lepton) and Modal Labs. NVIDIA's Lepton and Modal's $1.1B-backed platform already offer cross-cloud GPU arbitrage, the same pitch SF Tensor makes with $500K. Could a giant copy it? Yes, easily.

What protects it. Hyperscalers and NVIDIA have a conflict of interest in building a truly neutral cross-hardware layer, leaving a narrow opening SF Tensor could still occupy if it proves the compiler works.

What would change this call

SF Tensor closes a Series A or bridge round within 6 months. Shows whether investors see enough traction to keep funding the compiler/arbitrage bet.
A named design-partner deal with AMD or Google TPU sales org is announced. Would show real movement toward breaking the CUDA moat, not just marketing.
An independently verified benchmark showing the compiler beats hand-tuned CUDA on AMD or TPU hardware. Would confirm the technical differentiation actually works, not just claimed.

Key numbers

$500K total funding raised (Dec 2025)
4 to 6 person team (Jun 2026)
Modal Labs peer valued at $1.1B (Sep 2025)

Named risks

No independently verified benchmark exists proving the compiler beats hand-tuned CUDA. Severity: high
Only $500K raised against $225K to $315K salary bands implies fast burn before next fundraise. Severity: high
Crowded competitive field with far better-funded rivals (Modal, Applied Compute, Modular). Severity: high
Structural incumbents (NVIDIA Lepton, Google direct TPU sales) narrowing the cross-hardware abstraction niche. Severity: medium
No disclosed customers or revenue as of mid 2026, limiting proof of product-market fit. Severity: medium

Rivals, as read in July 2026

Rival	Direction	Evidence
NVIDIA (DGX Cloud Lepton)	losing to	NVIDIA rebranded acquired Lepton AI into a GPU marketplace by May 2025, pre-empting SF Tensor's arbitrage pitch
Modal Labs	losing to	Modal reached $1.1B valuation after $87M Series B in Sep 2025 versus SF Tensor's $500K total raised
Applied Compute	losing to	Reports named enterprise customers including Cognition, DoorDash, Bridge, Mercor as of Jul 2026; SF Tensor has none disclosed
Google (direct TPU sales)	contesting	Google began selling TPUs directly to select customers, announced Apr 29 2026, reshaping the hardware-abstraction niche
Modular (Mojo)	losing to	Modular has raised over $250M for a comparable CUDA-alternative compiler effort

Questions people ask

Is SF Tensor profitable?: No. As of the most recent disclosed round in late 2025, SF Tensor has raised only $500K total and reports no revenue, ARR, or named customers. Its website shows a demo pricing widget rather than customer logos, indicating pre-revenue status as of mid 2026.
Who is SF Tensor's biggest competitor?: NVIDIA's DGX Cloud Lepton, launched via the April 2025 Lepton AI acquisition and rebranded at Computex in May 2025, directly competes on cross-cloud GPU arbitrage. Better-funded peers Modal ($1.1B valuation, Sep 2025) and Applied Compute also target the same segment.
Could AI replace SF Tensor's product?: SF Tensor's core offering, automated kernel optimization and compiler tuning, is exactly the systems work frontier AI code-generation research increasingly automates internally. No proprietary data or workflow lock-in has been publicly verified to insulate the product as of mid 2026.
Who owns SF Tensor?: SF Tensor is privately founder-owned, started in 2025 by three brothers, Ben, Tom, and Luk Koska. It is backed by Y Combinator and Susa Ventures, plus angel investors including Max Mullen, Paul Graham, and executives from Neuralink, Notion, and AMD, per its own job listings.
What does SF Tensor actually build?: SF Tensor sells infrastructure so AI labs can train models without managing compute: a Kernel Optimizer for NVIDIA/AMD/TPU hardware, 'Tensor Cloud' for cross-cloud GPU sourcing up to 10,000 GPUs, and 'Emma Lang,' a hardware-aware programming language, launched publicly on YC on November 4, 2025.
Has SF Tensor proven its compiler beats CUDA?: No independent verification exists as of June 2026. The company claims (vendor-stated, unverified) 'up to 80% compute cost savings' from its kernel optimizer, but no third-party benchmark, case study, or named design partner confirming this against hand-tuned CUDA has been published.

Timeline

2026-04-29Google announces plan to begin direct TPU sales, projected ~$13B revenue by 2027
2026-04-12SF Tensor founders publicly state goal to 'break the CUDA moat'
2026-02-08Follow-on press coverage of SF Tensor's YC launch
2026-01-01SF Tensor posts founding engineer roles at $225K to $315K salary bands
2025-12-30CBInsights records SF Tensor Seed round with Y Combinator as investor
2025-11-04SF Tensor publicly launches via Y Combinator's Launch platform
2025-10Crustdata dates SF Tensor's Pre-Seed round, led by Y Combinator
2025-05NVIDIA rebrands acquired Lepton AI as DGX Cloud Lepton at Computex
2025-04NVIDIA acquires Lepton AI, later forming a direct competitor to SF Tensor
2025SF Tensor founded by brothers Ben, Tom, and Luk Koska, joins YC Fall 2025 batch

Revision history

27 July 2026

First assessed. Protection under test. The text above is fixed as of this date and is not edited afterwards.

In the same band

Deeptrace Lightberry Unsiloed AI