Deploy AI models on any GPU. Pay by the second, not the token.
Choose your model, pick your GPU, configure scaling and deploy. Time-based pricing, no token math, no surprises.
Billed per second · Scale-to-zero · No commitment
Three reasons to switch to serverless GPU.
What we do differently from other GPU inference providers.
What is serverless GPU? Serverless GPU runs your model on a dedicated GPU endpoint that scales with traffic and can scale down to zero when idle. You pay for GPU time, by the second, instead of per token.
Pricing
Billed by time, not tokens
Fixed GPU price per hour, billed per second. The longer your prompts and outputs, the more you save vs. token-based providers. No input/output cost split to decode.
Control
You choose the GPU
L40S, RTX PRO 6000 Blackwell, H200, B200 or B300: pick the GPU that fits your model and budget, even in serverless mode. Most providers abstract hardware away.
Flexibility
Serverless to dedicated in one click
Start serverless with scale-to-zero. When traffic grows, switch to a dedicated GPU: same console, same API key, same workflow. No migration.
Sovereignty
Sovereign AI Cloud
Your endpoint runs on our own NVIDIA GPU servers in Paris and Lyon. Your prompts and outputs are never used to train models.
Running quick tests or low volume? The Models API may cost less: you pay per token, with nothing to deploy.
Explore the Models API →From zero to production in four steps.
You are fully guided in our console: Model selector → GPU selector → Scaling config → Deploy.
01
Choose a model
Browse the catalog or paste any Hugging Face model ID.
02
Pick your GPU
Select the GPU type and quantity per replica.
03
Configure scaling
Light, Heavy or Custom: set min/max replicas and queue.
04
Deploy & monitor
Hit deploy, get your API endpoint, track performance live.
Auto-scaling
How auto-scaling works
Your endpoint runs between a minimum and a maximum number of replicas that you set. Replicas start when requests queue up and stop when traffic falls, so you only pay for the GPUs actually running.
Always-on baseline. Your minimum replicas stay warm and billable. Set it to 0 to scale to zero.
Spike, more replicas. Extra replicas start automatically, up to the maximum you choose.
Quiet, back to baseline. Replicas stop when demand falls, and billing stops with them.
Simple GPU pricing. No token math.
Pay per GPU-hour, billed to the second. Auto-scaling adjusts the number of active GPUs to match your traffic.
L40S
$1.99 /hr
Best for: small and quantized models, image generation, dev/test
Deploy now →RTX PRO 6000 Blackwell
$2.99 /hr
Best for: cost-optimized inference, batch processing, dev/test
Deploy now →H200
$6.99 /hr
Best for: large models (70B+), production workloads, low latency
Deploy now →B200
$8.99 /hr
Best for: 200B+ models, maximum throughput, enterprise scale
Deploy now →B300
$9.99 /hr
Best for: 400B+ models, long context, multi-model serving
Deploy now →Same workload. Different bill.
Same model. Same documents. Only the billing changes.
Arkane Cloud
1 hour of RTX PRO 6000 Blackwell running gpt-oss-120B. Process entire documents: all input tokens included at zero extra cost.
Token-based provider
Same hour, same model on Together AI. At $0.15/M input + $0.60/M output, document-heavy workloads cost 2.8× more.
Processing long documents (16:1 input-to-output ratio) on 1× RTX PRO 6000 Blackwell at 787 tok/s. Input tokens are free on GPU, while token providers charge for every one. Another example, for RAG workloads (8:1 ratio): $2.99 vs $5.10, a 41% saving.
Based on Together AI public list prices and throughput measured by Arkane Cloud.
Your own model or ready-made models?
Both give you an OpenAI-compatible API hosted in Paris and Lyon. The difference is what you pay for and how much control you get.
Models API
Serverless GPU
You get
A shared API to our model catalog
Your own endpoint, running the model you choose
Models
30+ ready-to-call open-source models
Any catalog model, or any vLLM-compatible Hugging Face model, including fine-tuned ones
You pay for
Tokens (input, output, cached) or images
GPU time, billed per second
GPU choice
Managed by Arkane Cloud
You choose the GPU
When idle
$0, you only pay per request
$0 with scale-to-zero
Best for
Prototyping, variable traffic, testing many models
Custom or fine-tuned models, long prompts, steady traffic
Questions about Serverless GPU
How does time-based pricing work exactly?
You pay per GPU-hour, billed to the second. While your endpoint is active and processing requests, you're charged the hourly rate for the GPU(s) you selected. With scale-to-zero enabled, you pay $0 when there's no traffic. No token counting, no input/output rate split.
How does autoscaling work?
Your endpoint runs between a minimum and a maximum number of replicas that you set. As requests queue up, replicas are added up to your maximum, then removed when demand falls. Set the minimum to 0 to scale to zero when idle. The Light and Heavy presets keep replicas running even without traffic.
What about cold starts?
Cold start times depend on the model size and GPU. For most models, expect 30 to 60 seconds on the first request after scale-to-zero. Set min replicas to 1 to eliminate cold starts entirely: you'll pay for one GPU continuously but get instant responses.
What can I deploy?
A model from our catalog or from Hugging Face, served with a managed vLLM runtime, or your own custom container: a public image that runs your own model server or HTTP application on the GPUs you choose, with the same scaling and per-second billing.
Can I deploy my own fine-tuned model?
Yes. Paste any Hugging Face model ID (public, or private with an HF token) and we'll deploy it on your chosen GPU. Not limited to our catalog: any vLLM-compatible model works. Still training it? See LLM fine-tuning on Arkane Cloud.
What are quiet hours?
Quiet hours are periods when your endpoint runs at minimum replicas instead of scaling up. For example, an RTX PRO 6000 Blackwell endpoint set to min 1 / max 3 replicas (the maximum is up to you) costs $2.99/hr at quiet (1 replica) and $8.97/hr at peak (3 replicas). You only pay for the GPUs actually running. No fixed discount schedule, just automatic scaling based on your traffic.
How do I switch from token-based inference to GPU-hour pricing?
Choose your model, pick a GPU and deploy. You get an OpenAI-compatible API endpoint, so most codebases only need a URL and key swap. No architecture changes, no vendor lock-in. More on running AI inference with Arkane Cloud.
Where are your data centers?
Our GPU infrastructure runs on our own servers in colocation datacenters in Paris and Lyon, France. Your prompts and outputs are never used to train models.
Ready to deploy?
Start with serverless, scale to dedicated. The playground is free, no credit card required.
No commitment · Scale-to-zero = $0 when idle
