An introduction to Sparset

An introduction to Sparset

Last week we announced our rebrand to Sparset. That post covered where the name came from and what changed. This one is about what we do and why we do it. Our co-founder and CEO, Marcel Slowikowski, explains it in a short video, and the rest of this post covers the same ground for anyone who would rather read.

Two options, and neither is great

Running your own AI model in production usually means one of two things.

You use the cloud, and you keep paying more as usage grows. Every new user is a bigger bill, and the bill is the one part of the product that scales perfectly.

Or you build the infrastructure yourself, and you hire people to deploy the model, optimise it, monitor it, and keep it alive. Your product team turns into a platform team, and the model still needs someone awake at three in the morning.

We ran into this ourselves while deploying models for our customers. It is what led us to Sparset.

What Sparset is

Sparset is the inference layer for AI running in production.

You bring your model. You bring your own infrastructure, or we help arrange it through our partners. Our agents then take the model, optimise it for the hardware it will run on, deploy it, operate it, recover it when something breaks, and keep looking for ways to make it faster and cheaper.

You keep the decisions. You draw the line between what runs on its own and what waits for your approval. The recovery actions you authorise up front, such as restarting a worker or shifting traffic, run inside the limits you set. Everything else is tested first and presented to you before it touches production.

There is no Sparset cloud. The deployment runs on infrastructure you own or rent, with a model you own, and it stays yours. We do not replace your engineers; we take the infrastructure work off them so they can work on the product.

What it looks like in production

We are already doing this for a customer. Roomates wanted to move away from an expensive third-party image generation API and run a model they own. They kept their model and their custom adapter. What changed is where it runs and who runs it.

On Gemini they paid 4 cents a generation. With Sparset they pay a fixed $400 a month for a single-GPU deployment. The price buys the GPU, not the generations, so it does not move with volume, and the cost per image falls as volume grows.

• At the GPU's projected capacity of 800,000 generations a month, that is $0.0005 a generation, 99% less than the 4 cents they paid before. Twenty images for a cent.

• At their current 50,000 a month, the monthly bill went from $2,000 to $400 the day they moved.

• Generation time is about three seconds, against a twelve-second baseline. Steady throughput is 1,180 images an hour, against 300 before.

Growth used to mean a bigger bill. Now it means a lower cost per image on a line they can plan against.

The idea behind it

Most AI infrastructure problems get solved by adding more compute. We think the better answer, most of the time, is to make the model use the compute it already has more efficiently. Choose the hardware carefully, make the model fit it, and then keep measuring, because the workload on day ninety is not the workload on day one.

That is the whole product in one sentence: a loop that plans, optimises, deploys, operates, and reviews your model, run by agents, bounded by you.

Where this goes

We believe running your own AI model should be a real option for more teams. More affordable, more accessible, and simpler to operate than it is today.

Our ambition goes further than data centres. AI should be able to run wherever it is needed: private servers, workstations, vehicles, robots, and eventually personal devices. The future of AI runs everywhere, and we are building the technology that makes that possible.

Come build it with us

We are a small team, and this is early. If the problem in this post is one you would rather work on than read about, the systems layer between AI models and the hardware they run on, we would like to hear from you. Open roles are on our careers page, and when there is not one that fits, write to careers@sparset.ai and tell us what you would build.