Turn the GPUs
you already own
into one computer.
One machine is never enough. Instead of renting the cloud or spending weeks stitching infrastructure together, install a 1 MB agenton every box with a GPU. Clystria fuses them into one cluster that just works — you don't think about which machine does what. You just deploy.
install the agent · join the cluster · deploy a model — that's it.
You bought the compute. You're still renting the answer.
Stranded silicon
A 4090 here, an A6000 there, a rack of H100s no one can fully book. Each a locked island. Your building already holds a supercomputer — it just doesn't know it yet.
The YAML tax
Kubernetes, Helm, schedulers, operators, service meshes. A second full-time job just to run the first one. Most teams give up and rent the cloud back at 8× markup.
Models too big to fit
The best open models no longer fit on one card. Sharding them by hand across machines is a research project. So they sit undeployed, and you pay per token elsewhere.
Three moves from cold metal to live intelligence.
Drop the agent on any GPU box
One line. 1.2 MB. Zero dependencies. It fingerprints the GPU, VRAM, interconnect and thermals, then whispers itself onto the network. Works on Linux, Windows, or that Mac Studio in the corner.
$ curl clystria.sh | shMachines melt into one pool
Agents gossip over your LAN and self-assemble into a single fabric. No control-plane to babysit, no etcd to lose sleep over. VRAM, compute and bandwidth become one addressable pool that grows the instant a machine powers on.
$ clystria status
▸ 14 nodes · 512 GB VRAM · healthySummon any model, one command
Clystria reads the model, plans which layers live on which GPU, balances the KV-cache, and routes traffic. A machine dies mid-request? Layers re-forge onto survivors before the token stream even stutters.
$ clystria deploy mixtral-8x22b --replicas autoYou don't think about which machine does what. You just deploy.
Gaming PCs, workstations, old servers, whatever you have — Clystria binds them into one system. Add a GPU and the cluster uses it. Lose a machine and the workload moves.
ClystriaUnified compute. Shared memory. One network.
Kill a machine. Watch it heal.
A live simulation of the control loop: four machines serve a 70B model, a node dies — and the workload re-forges onto the survivors before you'd notice. Hit Kill a node yourself.
▸ Live · llama-3.1-70b sharded across 8 GPUs · KV-cache balanced · 0 spillover.
An entire cluster team, compiled into a daemon.
Unified memory pool
Every card's VRAM becomes one addressable space. Run models bigger than any single GPU could ever hold.
Automatic layer placement
Clystria reads each model and decides which layers live where — by VRAM, bandwidth, and thermals. You never touch a shard map.
Self-healing failover
A machine dies mid-inference and layers re-forge onto survivors. Requests drain gracefully. Uptime measured in nines.
Zero Kubernetes
No control plane, no etcd, no Helm charts. A gossiping mesh that assembles itself. Delete the YAML museum.
One-command deploy
clystria deploy <model>. Quantization, sharding, routing and autoscaling handled. From HF repo to endpoint in seconds.
Gloriously heterogeneous
Mix a 4090, an H100 and a Mac Studio in the same pool. Clystria weighs each by its true throughput. No card left behind.
OpenAI-compatible
A drop-in /v1/chat/completions endpoint. Point your existing app at your own hardware. Cancel the invoice.
Live observability
Per-layer utilization, token economics, thermal maps and traces — streamed in real time. See the metal think.
OpenAI-compatible. Point & go.
Every deployment exposes a /v1 endpoint that speaks the OpenAI protocol. Change one base URL and your existing app runs on hardware you own — no rewrite, no vendor SDK.
- —Chat, completions & embeddings
- —Streaming + tool calls
- —Cancel the cloud invoice
curl http://home.local:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "llama-3.1-70b",
"messages": [{"role": "user", "content": "hi"}]
}'You, versus the hyperscalers.
| Clystria | Kubernetes | Hyperscale cloud | |
|---|---|---|---|
| Time to first model | ~90 seconds | 2–6 weeks | minutes* |
| Config to maintain | None | Thousands of YAML lines | Vendor console sprawl |
| Where your data lives | Your building | Your cluster | Someone else's |
| 8×H100 for a month | Hardware you own | Hardware you own | $50k+ rented |
| Mix consumer + datacenter GPUs | By design | Fight the scheduler | Not allowed |
| When a node dies | Self-heals, no ticket | Page the on-call | Read the status page |
| Runs on the gaming PC upstairs | Absolutely | Please don't | N/A |
* until the bill arrives.
You already bought the GPUs. Stop renting them back.
Renting 8 datacenter-class GPUs at ~$2.4/hr. With Clystria that same compute runs on the hardware you already own — for the price of electricity.
Illustrative. Real savings depend on your hardware, power cost and usage.
Built by two engineers.
Two builders who got tired of renting compute they already owned — and built the operating system to fix it.
The compute already exists. It's humming under desks, idling in closets, sleeping in racks that run at nine percent utilization. Every building is already a supercomputer — scattered, unaware, waiting for something to bind it into one.
Clystria is the binding. A private AI cloud shouldn't need a platform team, a data-center lease, or a hyperscaler's permission. It should be as easy as plugging computers into the same network.
Light your
own forge.
Point us at the GPUs you already own. We'll turn them into a single AI computer you fully control — and never send a token to a cloud you don't.