San Francisco Tensor Company

The San Francisco Tensor Company

[Ben Koska @BenKoska]

We formalized PTX in Lean so that we can formally prove GPU kernels correct.

This way, we can be confident that the kernels our agents wrote are correct, without ever needing to read them. 🧵

[Ben Koska @BenKoska]

Today, there are 3 classes of compute: On-Demand is reliable but expensive, Reserved requires you to commit to your spend years out and Spot is cheap but can be reclaimed at a moments notice and anything running dies. What if we could just... continue a process on a new host? 🧵

[Ben Koska @BenKoska]

turns out when you have too many GPUs for your own good you can do stupid stuff such as pre-training a foundation model with constrained decoding on PTX -- that is, it does not speak natural language, it's tokenizer is PTX, it's thoughts are PTX, it's all PTX 🧵

[Ben Koska @BenKoska]

We're #1 on NVIDIA's SOL-ExecBench across all four tracks.

Across 235 real production kernels run on B200s we beat out 583 teams, including teams that have raised hundreds of millions of dollars to optimize kernels.

Along the way we learned a lot about agentic optimization 🧵

[Ben Koska @BenKoska]

We've been working on training our own models for large-scale, long-running optimization tasks.

The model is still training, but we asked an early checkpoint to optimize pre-training on 32 MI355x, surprisingly it manages to train faster than NVIDIA NeMo running on 32 GB200s 🧵

[Ben Koska @BenKoska]

for RL and agents you need fast, isolated sandboxes which is why we've used gVisor for CPU and NVIDIA GPU workloads.

as we started post-training models for AMD GPU kernels though, we hit a problem: gVisor has no AMD GPU support. or rather, it didn't, until we built it :)

[Ben Koska @BenKoska]

We have Day‑0 support for Xiaomi-Robotics-1 at @SFTensor

Let us help you specialize it for your use case

[Ben Koska @BenKoska]

The biggest advantage of owning the backend end-to-end is that instead of relying on ptxas, you can try different ways of scheduling, such as search & RL. While ptxas beats our beta backend without search, once we turn on search we beat it by ~8 points 🧵

[Ben Koska @BenKoska]

we've long been dissatisfied with the fact that the closest we can get to the hardware on NVIDIA is PTX, with ptxas then making all sorts of decisions (instruction selection, barriers, stall counts, etc.) for us without any way to influence it, so we've been building a custom LLVM backend that let's us create cubins directly from LLVM, which let's us do things such as write inline SASS and define custom scheduling. reply or reach out if you wanna give it a twirl

[Ben Koska @BenKoska]

The benefits of AI are dangerously concentrated. Today they accrue to a handful of labs and to one industry, software, whose work happens to look like the internet the models were trained on.

Changing that is the single biggest untapped lever for economic growth this decade.🧵

[Ben Koska @BenKoska]

And because we automatically target whichever hardware is cheapest or best for a given workload, vendors have to compete on performance and price instead of coasting on lock-in.

A chip that becomes uncompetitive can no longer hide behind switching costs. It has to win on the merits.

[Ben Koska @BenKoska]

We built a software model of NVIDIA Blackwell's Tensor Core MMA instruction (https://t.co/CnfMTbYdBK) that matches real B200 hardware bit-for-bit.

Not "close enough for numerics." Every output word, identical.

Here's how we reverse-engineered it. 🧵

[Ben Koska @BenKoska]

We've stopped testing our GPU kernels and started proving them correct instead.

A test suite checks the inputs you thought of. A proof covers every input that exists and hands you the exact one that breaks the kernel if it's wrong.

Here's how. 🧵

[Ben Koska @BenKoska]

The moment we saw this letter, we signed. An open ecosystem beats a closed moat every time and is how AI actually reaches every sector. Proud to stand with everyone helping build and accelerate open, American AI

[Ben Koska @BenKoska]

We built a benchmark where the AI agent has to build an inference chip. GPT-5.6 took 3.4 days and 1.7B tokens.

The agent has to design, verify and physically place-and-route a working inference accelerator from a single MAC unit to a routed die that decodes tokens bit-exactly.

[San Francisco Tensor Company @SfTensor]

made it to Times Square! now back to building

thanks @brexHQ for the support getting here

[Ben Koska @BenKoska]

We're hiring a Founding Product Engineer. If we hire someone you refer, I'll wire you $15k.

- Strong design taste
- Experience with React/NextJS
- Care deeply about user experience
- Deep curiosity and drive to learn through experimenting

💵 $225k-275k base salary + equity

[Ben Koska @BenKoska]

Excited to have powered the training infrastructure for @KalpaLabsAI's latest model.

Congrats to @JhaGauti and @pshishodiaa on the launch and excited to see where you take it!

[Y Combinator @ycombinator]

SF Tensor (@sftensor) is building infrastructure for training foundation models, letting researchers focus on research and forget that the infrastructure layer even exists while saving 70% in training costs.

https://t.co/Kcy4TeCjm3

Congrats on the launch, @benkoska, @tom_koska, and @lukkoska!

[San Francisco Tensor Company @SfTensor]

One interface for everything. Automatically finds cheap GPUs, handles spot instances (saves up to 80%), moves jobs around when needed.

Tell us what you need. Your training runs. That's it.

https://t.co/hlRc577Vmu

[San Francisco Tensor Company @SfTensor]

Running AI workloads across clouds is honestly a nightmare. GPU availability is random, prices swing wildly, and managing AWS + GCP + Azure + niche providers is basically a full-time job.

We got sick of it, so we built Elastic Cloud. (1/2)