AI compute from racks in your neighbourhood
Run your models on NVIDIA RTX PRO 6000 Blackwell GPUs spread across many small sites — sold as one pool. One deployment, one HTTPS endpoint, one usage figure.
A GPU cloud built from many small sites
We own and run compact GPU racks at homes and properties across many regions, and sell their capacity as one cloud.
Rent GPU compute
Bring a container image. We run it on as many GPUs and copies as you ask for and give you one HTTPS endpoint.
- Up to 16 GPUs and 1.5 TB of GPU memory per copy
- Copies spread over several racks, so one site going down doesn't take you offline
- Your own servers — never shared with another buyer
Host a rack
Give one of our racks a home in your backyard, garage or outbuilding. We bring the hardware and the internet connection, and run it remotely.
- We install, monitor and maintain the rack
- Kernel2Cloud provides the rack's internet connection
- No software or IT work on your side
Racks at many homes
Each rack is a 1U control plane and one GPU server with 8 to 16 GPUs, spread across districts and regions.
One scheduler
Copies of your container go wherever there are free GPUs, spread across racks.
One endpoint
You see one capacity, one link and one usage figure — never individual sites.
Ship a container, get an endpoint
Tell us the image, the port, how many GPUs each copy needs and how many copies to run. We handle placement, health and rollouts.
// one copy = one server's worth of GPUs, or less { "name": "chat", "image": "vllm/vllm-openai:latest", "args": ["--model", "Qwen/Qwen2.5-72B-Instruct", "--tensor-parallel-size", "2"], "port": 8000, "gpus_per_replica": 2, "replicas": 3, "health_path": "/v1/models" } # then call it with your API key curl https://chat-acme.apps.kernel2cloud.ai/v1/models \ -H "Authorization: Bearer $K2C_API_KEY"
HTTPS with API keys
Every deployment gets its own hostname. Requests need one of your API keys.
Copies across racks
Traffic is load-balanced over all ready copies. If a rack goes down, the others keep answering.
Health checks
Set a health path and a startup window — loading a big model can take minutes, and that's fine.
Staged rollouts
New versions replace copies one at a time and pause if the new version can't start.
Your logs
Read your containers' logs in the portal. You only ever see your own.
GPU usage, hourly
GPU-hours per deployment and GPU model, broken down by hour.
NVIDIA RTX PRO 6000 Blackwell
Every rack holds one GPU server with 8 to 16 of them. Each copy of your model runs on a single server.
- GPUs per server
- 8–16
- Memory per GPU
- 96 GB
- GPU memory per server
- 768 GB – 1.5 TB
- Interconnect
- PCIe
Largest model per copy: what fits in one server — up to 1.5 TB of GPU memory (more with FP4). Models that span several servers aren't offered yet.
What teams run on it
LLM inference
Chat and completion APIs with vLLM, SGLang, TGI or your own server.
Image & video generation
Diffusion and video models behind an HTTP API.
Embeddings & reranking
The retrieval side of RAG, scaled out as copies.
Speech
Transcription and text-to-speech models.
Vision
Detection, segmentation and multimodal models.
Anything in a container
If it serves HTTP on a port, it runs. Private registries supported.
Give a rack a home
Our racks live in backyards, garages and outbuildings, spread out so no two sites are near each other. If you have the space, we'd like to hear from you.
We bring the rack
A compact rack — one control-plane server and one GPU server — delivered and installed by our team.
We bring the internet
Kernel2Cloud provides each rack's own connection, separate from your home network.
We run it remotely
Health, temperature and faults are watched around the clock. You never touch the software.
Have space for a rack? Tell us where you are and what space and power you have, and we'll follow up.
Apply to host →Built to run GPUs in places that aren't data centers
Racks sit on residential power and internet, in places we don't physically control. The platform is designed around that.
One pool, many sites
The scheduler treats every rack as part of a single on-demand pool.
How it places work- Each copy lands on one server with enough free GPUs
- Copies of a deployment are spread across racks
- Rollouts replace one copy per pass and pause when a new version can't start
- Each buyer gets whole servers; nobody shares a machine
GPU health from NVIDIA DCGM
Every GPU reports its health continuously, and problems become alerts for our operators.
What we do with it- GPU health alerts straight from DCGM telemetry
- Servers with problems are cordoned and drained, and their copies move elsewhere
- When a site loses power or cooling, the pool shrinks instead of failing
- All rack metrics are kept centrally for every site
Security for untrusted places
A rack in a backyard has to protect itself, so security starts at boot.
Layers- Immutable operating-system image on every server, with secure boot and TPM
- New racks register themselves and wait for an operator's approval
- Whole servers per buyer, on hardened Kubernetes
- Endpoints only answer requests carrying the buyer's API key
- Buyers can't override GPU driver settings from their containers
An operations console for the fleet
Our team runs every site from one place.
Tools- Fleet map with each rack at its site
- Hardware inventory read from the servers themselves
- Remote terminals on racks, without opening ports at the site
- An audit log of approvals, cordons and deployments
From your request to the silicon
We build and run every layer, from the server's operating system to the portal you deploy from.
Compute that lives where people do
Big data centers wait years for power. We put GPUs into many small places that already have it.
Our mission
Make AI compute easier to get by building a GPU cloud out of many small sites instead of a few giant ones — and sharing the benefit with the people who host it.
How we work
We own and operate every rack ourselves. Buyers deal with one cloud; hosts deal with one team.
Our values
- Reliability — no single site is a single point of failure
- Security — designed for untrusted places
- Honesty — clear about what's built and what's next
- Neighbourliness — good guests on every property
Need GPUs, or have space for a rack?
Tell us about your workload or your site. Our team answers both.
Kernel2Cloud Inc
372 Fairview WayMilpitas, CA 95035
United States
Kernel2Cloud Limited
480/17, 45th CrossJayanagar 8th Block
Bangalore 560082, India
From image to endpoint
Talk to us
Share your model, GPUs per copy and how many copies you need.
Get an account
Sign in to the portal, create API keys and add your registry login if your image is private.
Deploy
Create a deployment from the portal or the API.
Call your endpoint
Send requests to your HTTPS hostname with your API key.
What GPUs do you run?
NVIDIA RTX PRO 6000 Blackwell with 96 GB each. Every rack has one server with 8 to 16 of them — 768 GB to 1.5 TB of GPU memory — connected over PCIe.
How large a model can I run?
One copy runs on one server, so up to 16 GPUs and 1.5 TB of GPU memory per copy (more with FP4 weights). Models that span several servers aren't offered yet. Run more copies to serve more traffic.
What happens if a site goes offline?
Copies of your deployment run on different racks, and the gateway sends traffic only to copies that are ready. If a rack drops off, the others keep answering while the scheduler replaces the missing copy.
Do I share servers with other customers?
No. Each buyer gets whole servers, and you only see your own deployments and logs.
How is compute priced?
Pricing depends on your workload. Get in touch and tell us what you need.
What do I need to host a rack?
A dry, covered space with a suitable electrical supply. We provide the hardware and the internet connection, and assess every site before installing.