AI Fabric
GPU power, right from your code. Pick the GPU type and count in a single line, then fine-tune models, serve them over HTTPS or run large batch jobs, with no clusters to build or maintain. Your containers scale up with demand, scale down when idle, and you only pay by the minute while they run.
Built and operated by Blade, Shadow's AI infrastructure brand, under its own terms.
- Status
- Alpha
- Categories
- AI & Data, Compute & Storage
- Access
- Early access
- Release date
Key facts
- For
- Python developers building with AI
- Alpha
- December 2026
- Where
- GPUs in France, images and volumes on OVHcloud
- Price
- by the minute while your containers run
From your code to a GPU
You write and deploy from your machine with a dedicated SDK or CLI; Blade runs the GPUs in France and bills you by the minute; OVHcloud stores the container images and volumes your jobs start from.
Your GPU is one line of Python
There is no cluster to size and no node to patch. With the bladekit SDK, the GPU is an argument on the function that needs it: the card model, then the number of cards after a colon. The blade CLI runs it remotely, streams its logs back to your terminal, and stops billing when the container stops.
import bladekit
app = bladekit.app("fine-tune")
@app.function(gpu="RTX-A4500:4") # four cards, one container
def train(dataset: str):
...
Run it once with blade run, iterate with live reload on the GPU with blade serve, and ship endpoints and schedules as a versioned release with blade deploy, which you can roll back.
From fine-tuning to inference
Fine-tune and train
Bring Unsloth, TRL, PEFT or diffusers. Checkpoints land on a volume, and a run on a reclaimed card resumes where it stopped.
Serve over HTTPS
Put vLLM, SGLang or your own FastAPI server behind an HTTPS URL, with streaming and WebSockets. OpenAI-compatible clients work as they are.
Batch and scheduled jobs
Fan a function out over thousands of inputs with .map(), schedule it with a cron, and check progress from any machine.
Sandboxes for agent code
Run the code your AI agents write in isolated containers, with outbound network access cut off when you need it.
Scales with demand
Replicas are added as requests queue up and removed once idle, down to zero when nobody calls.
Pay by the minute
Billing runs from container start to stop. Interruptible cards cost less; reserved ones are never reclaimed.
Questions, answered.
Blade. AI Fabric is built, operated and billed by Blade, Shadow's AI infrastructure brand, under Blade's terms. OVHcloud presents it on Labs and stores its container images and volumes.
In December 2026. Access opens gradually: describe your workload on the AI Fabric page to request it.
NVIDIA GPUs, up to eight per container on a single machine. The available models, their memory and their prices are listed on the AI Fabric page.
By the minute, from the moment a container starts until it stops, model loading and the idle time before scale-down included. Image pulls are not billed, and the idle timeout is yours to set. Prices are set by Blade and published on blade.sh.
A cheaper GPU that can be reclaimed while it runs. A job on one saves its state when that happens and is queued again. For a service that must stay up, keep replicas on reserved cards.
Blade runs the GPUs in France. Container images and volumes are stored on OVHcloud.
Python 3.10 or later, with the bladekit SDK (pip install bladekit) and the blade command line. Start from a ready-made image for PyTorch, vLLM or SGLang, or bring your own Dockerfile or registry image.
Shape what ships next.
Join the User Experience Program to test upcoming products, get early alpha and beta access, and put your fingerprints on the roadmap.