Carpathian is a software and cloud company that runs on hardware the industry wrote off.
We build custom applications, host them on infrastructure we own and operate, and run an open research program into AI that fits on ordinary CPUs. Software has grown less efficient as compute got cheaper, and we work the other direction.
What we do
Five lines of work, all sitting on the same infrastructure and run by the same team.
Cloud servers
Virtual machines on hardware we own, with full root access, a flat monthly rate, and no per-request billing. Deploy in minutes and scale RAM, vCPU, and disk independently.
AI inference
Hosted open source models behind an OpenAI-compatible endpoint, so applications can call AI without anyone provisioning or maintaining a GPU.
Software development
Custom web and mobile applications, internal tools, and integrations, built by the same engineers who run the infrastructure they deploy onto.
Network and storage
Private networking, block storage, and object storage across our own infrastructure, included as part of the platform rather than sold as separate products.
Veritate research
An open research program into efficient, inspectable AI: a quantization-aware training stack and a hand-written CPU inference engine, published for anyone to read.
Why reclaimed hardware
We started on reclaimed hardware because we were bootstrapped and had no interest in taking on debt. We kept doing it after we no longer had to, once it was clear how much serviceable equipment gets written off every year, and how much of what does get bought sits massively overprovisioned.
Equipment is usually retired on a depreciation schedule rather than because it failed. Rebuilding it costs us less to stand up and less to run than buying new, and because we are not servicing the cost of new hardware, that difference shows up in the price list rather than in a margin.
- 2023Founded on computer vision workThe company started out building computer vision applications for businesses.
- 2023Moved into cloud and softwareClient work made a different gap obvious: applications that were not painful to use, and hosting a small team could manage without a platform engineer.
- TodayOur own cloud platformCustom applications, cloud servers, and AI model hosting through an inference API, all running on hardware we own.
- OngoingVeritate researchAn open program into inference that runs on ordinary CPUs, published as we go rather than at the end.
How we got here
Carpathian started in 2023 building computer vision applications. Working with businesses made a different gap obvious: they needed enterprise applications that were not painful to use, and hosting that was simple enough to manage without a platform team.
So we moved into cloud infrastructure and software development, and built the hosting we wanted to use. Today we build custom applications, host them on our own cloud platform, and provide AI model hosting through an inference API, with a research branch working on models that run lean.
How we operate
We own the stack
Our hardware sits in our racks. Nothing is resold capacity behind a white-labeled dashboard, so one team stays accountable from the metal up to the application.
Pricing is flat and published
Monthly rates with published upgrade tiers. Traffic and compute are included in the rate rather than metered on top of it, so the invoice does not move when your usage does.
An engineer answers
Support reaches the people who build and operate the platform, with no tier-one queue sitting in between.
The research is public
We publish what we measure, including the approaches that did not work, and release the tools we build as open source.
What the research is for
Carpathian is not a frontier AI lab and is not trying to be one. The research addresses a narrower and more measurable problem: the strain large-scale AI places on energy and compute, and how far the footprint of inference can be reduced without giving up capability.
That work starts at the inference engine. Rather than build on a framework that assumes a GPU and carries a runtime to match, we wrote ours from scratch in C, with hand-written SIMD kernels and low-precision math built directly into the decode path.
The model is designed the same way: a small reasoning core that holds language, logic, and behavior rather than memorizing the world, with knowledge kept outside the parameters and retrieved as needed. Because the core reads raw bytes instead of relying on a heavyweight tokenizer, it trains and runs end to end on a single consumer-grade machine.
This is research, not a product. We publish it so anyone can build on it, and we release the tools as open source.
What we build and run
Three things, on hardware we own and operate: an AI chat and inference API, custom software built by the engineers who host it, and cloud VPS hosting with flat monthly pricing.
The people behind it
Carpathian is a small team with backgrounds in defense, financial systems, and cloud engineering. The person who provisions your server is the person who answers when you write in.