Open-source LLM hosting
Run Llama, Mistral, Gemma, Qwen and similar models on hardware you control.
Models
Llama family, Mistral, Gemma, fine-tuned variants
GPU / AI Compute
GPUs, live.
Real ones, priced live.
GPU compute for open-source LLMs, computer vision, fine-tuning, and embeddings. Live pricing below. GPU stock is allocated per order and provisioned by hand, so setup is not instant.
Order today, allocated per order Live pricing in your currency Root access and full control
kpanel.kapsulecloud.com/compute
GPU Compute
Live pricing, order today
Hardware specConfirmed
Pre-installed AI stacksConfirmed
API + KPanel surfaceConfirmed
Pricing modelPer-hour confirmed
Live pricingConfirmed
GDPR-compliant data handlingConfirmed
Honest signal: stock is allocated per order and provisioned by hand.
NVIDIA
RTX-class GPUs
4 tiers
Spark to Reactor
Per-order
Provisioned by hand
API
Provisioning + monitoring
Target workloadsOpen-source LLM inference·Computer vision·Fine-tuning·Embedding pipelines·Diffusion models·Audio / speech models
The full ladder
Every tier, unmanaged and managed
Forge
$2,980.00Unmanaged /mo
NVIDIA RTX PRO 6000 · 96 GB VRAM · 256 GB RAM
Managed /mo: $6,820.00
Foundry
$4,470.00Unmanaged /mo
NVIDIA RTX PRO 6000 · 96 GB VRAM · 512 GB RAM
Managed /mo: $10,220.00
Reactor
$5,740.00Unmanaged /mo
NVIDIA RTX PRO 6000 · 96 GB VRAM · 768 GB RAM
Managed /mo: $13,140.00
Limited, reserve when short
What you'll run on it
Workloads we're targeting.
These are the workloads the GPU tiers are built for. If yours matches, pick the tier that fits and order it.
Computer vision
Image classification, object detection, segmentation. Training and inference on the same platform without re-deploying.
Models
YOLO, SAM, custom PyTorch / TensorFlow models
Fine-tuning
LoRA / QLoRA fine-tunes on open-base models. Per-hour billing means you only pay for actual training time.
Models
LoRA, QLoRA, full fine-tunes on smaller models
Embeddings + RAG
Generate embeddings at scale for retrieval-augmented generation pipelines. Connect to your vector DB.
Models
sentence-transformers, BGE, e5, custom embedders
Diffusion / image-gen
Stable Diffusion, FLUX, Kandinsky. Batch-generate or expose as an API endpoint.
Models
SD, SDXL, FLUX, ControlNet, custom checkpoints
Audio / speech
Speech-to-text, text-to-speech, voice cloning. Whisper transcription pipelines.
Models
Whisper, Coqui, Bark, custom TTS
Available right now
Need compute today? Use Cloud VPS + JupyterHub.
Our Cloud VPS lineup runs JupyterHub as a one-click installer. CPU-bound inference, smaller models, embedding generation, batch pipelines: all run today without waiting for GPU.
- One-click JupyterHub install on any Cloud VPS tier
- From $62.00/mo Comet (2 vCPU, 4 GB) to $779.00/mo Nova (16 dedicated vCPU, 64 GB)
- Per-hour billing, NVMe Gen4 storage, root SSH
- Migrate to GPU compute at launch, your code already runs
FAQ
Frequently asked questions
How fast can I get a GPU box?+
Order any tier today at the price shown. GPU stock is allocated per order and provisioned by hand (Robot-class hardware), so there is a short lead time between order and handover rather than instant self-provisioning. We confirm timing at checkout.
What hardware will you use?+
NVIDIA GPUs across four tiers: Spark on an RTX 4000, and Forge, Foundry, and Reactor on the RTX PRO 6000 with increasing system memory. Each tier lists its GPU, VRAM, and RAM on the pricing grid above.
Where are the GPUs located?+
We confirm the exact data-centre location with you before handover. GPU stock is allocated per order and provisioned by hand, so we are transparent about the region and specifics at launch.
What about pricing?+
Live monthly pricing is on this page, per tier, in your currency. Both unmanaged and managed prices are shown.
Can I do anything useful today while I wait?+
Yes. Spin up a Cloud VPS with JupyterHub one-click installed. CPU-bound LLM inference (smaller models), embedding generation, batch pipelines, all run there today. When GPU lands you migrate the code, no rewrite.
What workloads are not a fit?+
Training large foundation models from scratch (we are targeting fine-tuning and inference, not the >$10M training-run scale). Closed-source proprietary models requiring vendor agreements. Anything requiring InfiniBand-scale multi-GPU clusters at launch (multi-GPU is roadmap, not GA).
How is GPU stock allocated?+
Stock is allocated per order and provisioned by hand, so there can be a short lead time between order and handover. This is the same manual-fulfilment path as our dedicated and storage server lines.
Will Kora know how to help me with GPU workloads?+
Kora knows the GPU compute product: the hardware, the pricing, the supported frameworks, and the deploy patterns.
GPU / AI Compute
Spin up a GPU box today.
Live monthly pricing per tier, root access. Stock is allocated per order and provisioned by hand, so setup is not instant.