GPU cloud pricing in taka
Different GPU, different rate.
AI COMPUTER PRICING
These are demand-study estimates, not final launch prices. Rates may change with capacity, FX and tax treatment.
| AI Computer | Memory | Architecture | Indicative / hour | 40 hours | Per prompt | Per image | |
|---|---|---|---|---|---|---|---|
| NVIDIA L4Fast inference, video AI, efficient fine-tuning | 24 GB | Ada | ৳147 | ৳5,880 | ≈ 2.3 paisa | ≈ 53 paisa | |
| NVIDIA A10Rendering, vision, medium training workloads | 24 GB | Ampere | ৳196 | ৳7,840 | ≈ 2.5 paisa | ≈ 54 paisa | |
| NVIDIA L40SGenerative AI, rendering, larger fine-tunes | 48 GB | Ada | ৳343 | ৳13,720 | ≈ 1.8 paisa | ≈ 43 paisa | |
| NVIDIA A100LLM fine-tuning, research, high-throughput training | 80 GB | Ampere | ৳638 | ৳25,520 | ≈ 2.5 paisa | ≈ 71 paisa | |
| NVIDIA H100Large model training, transformer workloads, HPC | 80 GB | Hopper | ৳1,047 | ৳41,880 | ≈ 2.1 paisa | ≈ 73 paisa | |
| NVIDIA H200Memory-heavy LLMs, retrieval, frontier research | 141 GB | Hopper | ৳1,390 | ৳55,600 | ≈ 2.4 paisa | ≈ 89 paisa | |
| NVIDIA B200Next-gen Blackwell: large-LLM training and the fastest inference | 180 GB | Blackwell | ৳1,652 | ৳66,080 | ≈ 1.6 paisa | ≈ 69 paisa | |
| NVIDIA B300Most powerful: 288 GB for trillion-parameter models and reasoning AI | 288 GB | Blackwell Ultra | ৳1,880 | ৳75,200 | ≈ 1.5 paisa | ≈ 68 paisa |
Per-prompt and per-image figures are rough equivalents: a prompt is one ~500-token answer from an 8B open model (Llama/Qwen class) and an image is one 1024×1024 SDXL-class image, with the AI Computer kept busy. An idle AI Computer still bills by the hour, so light use costs more per prompt.
Indicative pricing: Displayed BDT rates use current infrastructure, service and applicable-tax assumptions. Storage, network transfer and special software may be separate.