Simple, Transparent Pricing
Pay only for what you use. No hidden fees. No egress charges. Cancel anytime.
Up to 60% cheaper
than RunPod on the same GPU models
Multi-Region Infrastructure
GPUs in APAC and beyond. Low-latency for developers in India, SEA, LatAm, and Europe.
60-second deploy
Browser-based. No SSH, no setup.
| GPU Model | Stock | CloudGPU | Vast.ai | RunPod |
|---|---|---|---|---|
| RTX 3090 24GSupplied via P2P agents | Low stock | $0.20/hr | $0.22/hr | $0.22/hr |
| RTX 4090 24G | Low stock | $0.32/hr | $0.32/hr | $0.34/hr |
| RTX 4090D 48GChina-spec 4090, 48GB VRAM | Low stock | $0.59/hr | N/A | N/A |
| 5090 32G | Low stock | $0.49/hr | N/A | N/A |
| A100 40G | Low stock | $0.89/hr | $0.80/hr | $1.89/hr |
| L20 48GChina-spec Ada, big VRAM | Low stock | $0.89/hr | N/A | $1.20/hr |
| L40 48G | Low stock | $0.79/hr | N/A | $1.19/hr |
| L40S 48G | Low stock | $1.19/hr | $1.65/hr | $1.89/hr |
| H20 96GChina-spec Hopper, huge HBM3 | Low stock | $1.59/hr | N/A | N/A |
| Ascend 910B 64GNPU — CANN / MindSpore / PyTorch-NPU only | Low stock | $0.85/hr | N/A | N/A |
| A800 80GChina-spec A100 | Low stock | $1.39/hr | N/A | $2.17/hr |
| H800 80GChina-spec H100, 400Gb/s NVLink capped | Low stock | $3.49/hr | N/A | $3.99/hr |
Vast.ai / RunPod rates observed April 2026 for equivalent SKU. "N/A" = competitor does not offer that GPU in their catalog.
Performance & Workload Fit
Price alone doesn't tell you which GPU is right. L20, H20, and Ascend 910B are China-spec cards that US clouds don't sell — they have unique VRAM / bandwidth profiles worth comparing.
| GPU | Arch | VRAM | FP16 TFLOPS | Mem BW | Best For |
|---|---|---|---|---|---|
RTX 3090 24G Supplied via P2P agents | Ampere | 24 GB | 36 | 936 GB/s | LLM inference ≤ 7B, Stable Diffusion, cheap dev box |
RTX 4090 24G | Ada | 24 GB | 73 | 1008 GB/s | Fine-tuning 7B-13B, Flux / SDXL, fast inference |
RTX 4090D 48G China-spec 4090, 48GB VRAM | Ada | 48 GB | 73 | 1008 GB/s | Fine-tuning 13B-30B at higher batch sizes, 70B quantized inference |
5090 32G | Blackwell | 32 GB | 104 | 1792 GB/s | Latest consumer flagship — Blackwell features, SD/Flux speed |
A100 40G | Ampere | 40 GB | 312 | 1555 GB/s | LLM training / fine-tune 13B-30B, research |
L20 48G China-spec Ada, big VRAM | Ada | 48 GB | 119 | 864 GB/s | 32B inference, multi-model serving, VRAM-heavy workloads |
L40 48G | Ada | 48 GB | 90 | 864 GB/s | 32B inference, visualization + compute combined |
L40S 48G | Ada | 48 GB | 183 | 864 GB/s | High-end inference + light training, better than L40 |
H20 96G China-spec Hopper, huge HBM3 | Hopper | 96 GB | 148 | 4000 GB/s | 70B-120B inference, long-context, MoE models |
Ascend 910B 64G NPU — CANN / MindSpore / PyTorch-NPU only | Ascend | 64 GB | 320 | 1200 GB/s | Domestic-compliant training, MindSpore/PyTorch-NPU workloads |
A800 80G China-spec A100 | Ampere | 80 GB | 312 | 2039 GB/s | FP32/FP16 training 30B-70B, large-memory research |
H800 80G China-spec H100, 400Gb/s NVLink capped | Hopper | 80 GB | 989 | 3350 GB/s | Frontier training, 70B+ fine-tuning, FP8 inference |
VRAM determines model size
24 GB fits 7B at FP16 / 30B quantized. 48 GB fits 13B at FP16 / 70B quantized. 96 GB comfortably runs 70B FP16.
Mem bandwidth = tokens/sec
For LLM inference, throughput scales with memory bandwidth, not TFLOPS. H20's 4 TB/s beats A100 for serving large models.
FP16 TFLOPS = training speed
A100's 312 TFLOPS shines for fine-tuning. 910B's 320 is on paper — actual speed depends on CANN/MindSpore ecosystem maturity.
Cost Calculator
Pricing Plans
On-Demand
Pay as you go. No commitment.
Market Rate
- check Per-hour billing
- check Cancel anytime
- check No minimum spend
- check All GPU models
Monthly
Commit monthly, save 30%.
30% Off
on hourly rates
- check Everything in On-Demand
- check 30% discount on all GPUs
- check Priority provisioning
- check Email support
Prepaid Credits
Buy credits in bulk, get bonus.
$100 → $110
10% bonus credits
- check 10% bonus on all top-ups
- check Credits never expire
- check Use on any GPU model
- check Transferable balance
No hidden fees · No egress charges · Per-hour billing · Cancel anytime