AI Workload Cost Calculator
Pick a workload. See which GPUs have the VRAM to run it, and the real $/hour and $/1M-token cost at today's cheapest live rental prices — not the hardware sticker price.
📊 Updated from the live gpuprice.in feed (source: parallel.ai).
1. Pick your workload
Run value
2. Live rental rates for the selected GPU
| PROVIDER | ON-DEMAND $/HR | SPOT $/HR | ₹/HR (OM-DEMAND) |
|---|---|---|---|
| Loading live rates… |
3. Full cost comparison — every GPU that fits
| GPU | CHEAPEST $/HR | $/DAY | $/MO (730H) | $/1M TOKENS | ₹/1M TOKENS |
|---|---|---|---|---|---|
| Computing… |
Read this table: the GPU that fits your workload and is cheapest at the bottom line wins. For an 8B chat app the RTX 5090 is the $/token king; an 80GB Llama 70B needs H100/A100; a 141GB+ or 192GB model needs H200/B200.
How to read the cost numbers
- $/1M tokens = cheapest live on-demand $/hr ÷ (tokens/hr on that GPU) × 1,000,000. It answers "am I better off renting a GPU or paying an API?"
- $/day & $/month use the same 24h / 730h methodology as the rest of gpuprice.in.
- Token rates are single-GPU methodology defaults (real-world ranges for each model class on that GPU). Throughput scales with batch size, quantization (FP8 vs FP16) and the exact provider silicon — treat $/1M-token as a directional planning number, not a vendor benchmark. Raise the default to see the ceiling.
- Spot vs on-demand: the current feed has H100/H200 spot 28–60% below on-demand. If your batch or inference job can checkpoint-and-resume, spot makes every number above roughly half — at the cost of interruption risk.
- VRAM is the gate. A GPU with less VRAM than your model needs simply cannot host it, no matter how cheap the $/hr — that's why the table only lists GPUs that fit.
Frequently asked questions
Which GPU can run my model — how do I know?
Match VRAM. An H100 (80GB) and A100 (80GB) fit Llama-70B-class models; an H200 (141GB) or B200 (192GB) fits 140B+ and 400B-class MoE models; the RTX 5090 (32GB) handles small 8–13B models. The calculator lists which GPUs fit the workload you picked, with each one's cheapest live rate.
Is renting better than paying an API for inference?
It depends entirely on utilization. The $/1M-token number on this page is your break-even: if an API provider charges more per token than your self-hosted GPU cost, renting wins once you keep the GPU busy. If you run quietly for a few hours a day, an API is simpler and cheaper.
Why is the RTX 5090 so cheap per token?
Because it's a consumer Blackwell GPU renting for a fraction of an enterprise datacenter card (the live feed shows it well under $1/hr). For small models it's the cheapest $/token on the site — but its 32GB VRAM caps it to small workloads. Big workloads need the bigger VRAM cards.