If you need to train ONNX models for MetaTrader 5, or run inference on a model too big for CPU, you need GPU compute. Buying a card is one option. For most retail traders — especially those with laptops, no Windows-friendly local PC, or sporadic training workloads — renting by the hour is cheaper.
Four providers cover this market well. Each one has trade-offs that aren't always obvious from their landing pages. We've used all four. Here's the honest comparison.
What's in this article
TL;DR — which one for which use case
- Training a model occasionally, cheapest possible: Vast.ai. Bid pricing for community providers, ~$0.20/hour for an RTX 4090.
- Training a model, want a reliable interface: RunPod. Slightly more expensive than Vast.ai, much smoother UX.
- Production inference 24/7 with predictable billing: DigitalOcean GPU Droplets. Flat monthly pricing, business-grade SLA.
- Need Windows on the GPU instance (running MT5 there): Vultr. The only one of the four with first-class Windows GPU images.
The full comparison table
| Provider | Pricing model | Cheapest GPU (~$/hr) | Windows images? | Best for |
|---|---|---|---|---|
| RunPod | Per-hour or per-second | RTX 4090 ~$0.34 | Limited (containers) | Training workflows |
| Vast.ai | Bid / market | RTX 4090 ~$0.20 | Some hosts | Cheapest training |
| DigitalOcean | Per-hour or monthly | NVIDIA L40S ~$1.57 | No (Linux only) | Predictable production |
| Vultr | Per-hour or monthly | A16 / L40S from ~$0.50 | Yes (first-class) | Running MT5 directly on GPU |
RunPod — flexibility, community pods, decent UI
RunPod splits its catalogue into two tiers: Secure Cloud (their own datacenter hardware, more expensive, more reliable) and Community Cloud (independent providers, cheaper, varying reliability). For training a model you don't need to retry constantly, the community tier is great. For anything sensitive, secure cloud.
- Pricing: from ~$0.20/hr for a Tesla T4 (community), ~$0.34/hr for an RTX 4090 (community), ~$1.99/hr for an H100 (secure).
- UI: the cleanest of the four. Pre-built templates for PyTorch, TensorFlow. SSH/Jupyter access included.
- Windows: not native — everything runs in Linux containers. You can run Wine/Bottles for MT5, but it's awkward.
- Best for: training PyTorch/TensorFlow models, fast iteration. Export the
.onnxthere, download it, deploy elsewhere.
Vast.ai — cheapest, harder UX
Vast.ai is a marketplace — independent operators list their GPUs, you bid by the hour. Prices are aggressive (often half of RunPod community), but reliability is genuinely variable. A "host" can take their machine offline; your job dies.
- Pricing: wildly variable. RTX 4090 from $0.20/hr "interruptible," ~$0.40/hr "on-demand."
- UI: dense. Filter by GPU model, by host reliability, by network speed. Once you know what to look for, it's powerful. First-time users get lost.
- Windows: some hosts offer it, most don't. Filter by OS.
- Best for: cheapest possible training when you can re-run if a job dies. Long-running batch tasks. Not production.
DigitalOcean GPU Droplets — predictable pricing
DigitalOcean launched GPU Droplets in 2024 with H100 instances; in 2026 they've added L40S as the cheaper option. Pricing is flat — no bidding, no market — which is what businesses want.
- Pricing: NVIDIA L40S from ~$1.57/hr, H100 from ~$3.39/hr. Monthly cap pricing available.
- UI: the same dashboard as the rest of DigitalOcean. Spin up in 60 seconds.
- Windows: not on GPU Droplets — Linux only. Important.
- Best for: companies already on DigitalOcean for other services, anyone who needs predictable monthly billing.
Vultr Cloud GPU — hourly Tesla cards with Windows
Vultr's Cloud GPU offering covers Tesla A16, A40, A100, L40S, H100. The differentiator: first-class Windows Server images on GPU instances. None of the other three do this well. If you want to run MetaTrader 5 itself on the GPU server (rather than on a separate machine), Vultr is the only practical option.
- Pricing: A16 fractional GPU from ~$0.50/hr, A100 from ~$2.30/hr.
- UI: straightforward. Pick region, image, GPU type.
- Windows: yes — Windows Server 2022 images available on GPU instances. Click-through MT5 install.
- Best for: "I want a Windows machine with an NVIDIA GPU and MT5 on it, available 24/7, billed monthly."
Windows + MT5: the catch
The honest caveat: MetaTrader 5 requires Windows. The most affordable training providers (RunPod community, Vast.ai community) are Linux-only. If you want MT5 itself running on the GPU instance, your only realistic option is Vultr.
For most workflows, this is fine: train on Linux GPU (cheap), download the .onnx, deploy to a separate Windows machine (could be your local PC, a forex VPS, or a Vultr Windows instance) for running MT5. See the forex VPS vs GPU cloud article for the full split-infrastructure pattern.
If you genuinely need GPU + MT5 on the same machine 24/7 (large model, can't fall back to CPU), Vultr is the answer. The monthly cost is real — an A100 instance running 24/7 is over $1500/month — but for some workloads it's the only setup that works.