Solutions · Model training
GPU servers for model training, billed by the month
A training run is the definition of a machine that stays busy: weeks of wall-clock time at full utilisation, checkpoints every few hours, and a bill that hourly clouds price for the idle capacity they keep on standby. A monthly server or node removes the meter — and the idle.
- Recommended GPU$1,249NVIDIA H100 SXM, 80 GB HBM3, per month
- Same GPU, hourly on-demand$2,446market median × 730 h
- You keep−49%8× node from $7,990 a month
Hardware
Which GPU for model training
Four picks from the catalogue, each a dedicated server with host, NVMe, network and support included. Click through for specs, price by term and the market comparison.
| GPU | Why for this workload | Per GPU, per month | Availability |
|---|---|---|---|
| NVIDIA H100 SXM80 GB HBM3 · Hopper | The default for anything up to 70B on an 8-GPU node: FP8 Transformer Engine, NVLink 4, 80 GB HBM3. | $1,249 | 14 left |
| NVIDIA H200141 GB HBM3e · Hopper | Same compute, 141 GB: larger micro-batches and longer sequences without more GPUs. | $1,690 | 9 left |
| NVIDIA B200180 GB HBM3e · Blackwell | Blackwell for the biggest runs: 180 GB and 8 TB/s per GPU, NVLink 5, FP4/FP8. | $2,690 | 8 left |
| NVIDIA A10080 GB HBM2e · Ampere | Fine-tunes and classic deep learning, where FP8 does not help and budget does. | $649 | 21 left |
Highlighted row: our default recommendation for model training. Prices per GPU per month in USD, excl. VAT; 3-, 6- and 12-month terms take 5–15% off. Stock counters updated 18 September 2026.
Sizing
How much GPU memory does training need?
Rules of thumb we use when a customer asks which GPU to order. Memory decides; everything else is speed. For a figure for your own model, use the GPU memory calculator.
GPU memory by model
The open models customers ask about most, worked out with the same formulas as the calculator: a 2k-token sequence, a micro-batch of 4 and gradient checkpointing. Each model has its own page with every precision and context length.
| Model | LoRA, 16-bit base | Full fine-tune | Suggested here, full fine-tune |
|---|---|---|---|
| Llama 3.1 8B | 23 GB | 135 GB | 2× A100 · $1,298/mo |
| Qwen3 14B | 40 GB | 246 GB | 4× A100 · $2,596/mo |
| Mistral Small 3.2 | 59 GB | 394 GB | 2× B300 · $7,580/mo |
| Qwen3 32B | 82 GB | 540 GB | 8× A100 · $5,192/mo |
| Llama 3.3 70B | 173 GB | 1,158 GB | 8× B200 · $18,900/mo |
GB are 109 bytes, runtime overhead included. Suggestions favour one large card over several small ones and never exceed the stock on display. More models in the GPU memory calculator.
From order to first run
- Pick the GPU and the termOne GPU for a fine-tune, an 8-GPU node for a pre-train; a 3-, 6- or 12-month term takes 5–15% off a run you know will last.
- Boot a ready imagePyTorch 2.7 with CUDA 12.6, the Hugging Face stack, JAX, or a bare Ubuntu with the driver — chosen at checkout, live in under 10 minutes.
- Stage data on local NVMe1–4 TB per server and 30 TB per node, as fast as the GPUs can read; attach block volumes for larger corpora.
- Checkpoint, monitor, renewCheckpoints stay on your NVMe, nvidia-smi and DCGM are yours, and the renewal settles from your balance so the run is never interrupted by a card.
In every server
- Single-tenant bare metal. Your GPU, CPU cores, RAM and NVMe are yours alone. No noisy neighbours, no oversubscription.
- 10–25 Gbps uplink. Dedicated bandwidth per server with 20 TB of outbound transfer included every month. No egress fees.
- Local NVMe storage. Fast scratch space for datasets and checkpoints, sized to the GPU. Expand with attached block volumes.
- Ready-to-train images. Ubuntu with NVIDIA drivers, CUDA 12, cuDNN and Docker — or PyTorch, TensorFlow, vLLM, Kubernetes and Windows Server images at checkout.
- Public IPv4 & IPv6. A dedicated IPv4 address and a /64 IPv6 block on every server, with reverse DNS on request.
- DDoS protection. Always-on volumetric mitigation at the network edge, included at no charge.
Model training on a monthly GPU: questions answered
Is a monthly server cheaper than on-demand for training?
For a run that keeps the GPU busy, by a wide margin: an H100 SXM costs $1,249 a month here against roughly $2,446 for a full month at the market-median on-demand rate. The crossover is around 373 hours a month — below that, per-second billing wins, and we say so.
Can I train across more than one node?
Yes. Each node carries 3.2 Tb/s of InfiniBand for multi-node training with NCCL; for more than one node, open a sales ticket and we place them on the same fabric.
Do you sell spot or preemptible capacity?
No. Every server is dedicated and stays up for the whole term — the checkpoint you save at 03:00 is the one you resume from, not the one you lose.
Start model training tonight
Every GPU above is in stock. Fund your balance, pick the server, and be at a root prompt in under 10 minutes.