*[DediGPU](https://dedigpu.com/) — Markdown mirror of [https://dedigpu.com/gpu](https://dedigpu.com/gpu) · updated 2026-09-02 · index for LLMs: [llms.txt](https://dedigpu.com/llms.txt) · everything: [llms-full.txt](https://dedigpu.com/llms-full.txt)* GPU servers # Every GPU server, side by side. Priced at cost. 13 dedicated servers from the B300 to the RTX 4090, one tenant each, billed monthly at their published cost sheet. Sort the table, open a card for its cost sheet, configurations, memory fit and templates. [Start with an account](https://dedigpu.com/signup) [How we price](https://dedigpu.com/pricing) - Same price in 4 regions - 1× to 8× per server - No setup fee, $0 egress - Crypto only, no KYC Per card, per month, one-month term. Click a heading to sort. | Card | Memory | Bandwidth | Interconnect | vCPU | RAM | NVMe | Port | Per month | 12-mo rate | Cheapest found | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | [**B300**](https://dedigpu.com/gpu/b300) Blackwell Ultra · SXM | **288 GB** HBM3e | 8 TB/s | NVLink 5 · 1.8 TB/s | 32 | 288 GB | 3 TB | 100 Gbit/s | $1,762 | $1,585.80 | $5,840 Latitude.sh (8× node) | | [**B200**](https://dedigpu.com/gpu/b200) Blackwell · SXM | **180 GB** HBM3e | 8 TB/s | NVLink 5 · 1.8 TB/s | 28 | 224 GB | 2 TB | 100 Gbit/s | $1,408 | $1,267.20 | $3,723 Hyperstack (reserved) | | [**H200**](https://dedigpu.com/gpu/h200) Hopper · SXM | **141 GB** HBM3e | 4.8 TB/s | NVLink 4 · 900 GB/s | 24 | 224 GB | 2 TB | 100 Gbit/s | $1,252 | $1,126.80 | $2,037 Hyperstack (reserved) | | [**H100 SXM**](https://dedigpu.com/gpu/h100-sxm) Hopper · SXM | **80 GB** HBM3 | 3.35 TB/s | NVLink 4 · 900 GB/s | 20 | 192 GB | 1.5 TB | 100 Gbit/s | $1,134 | $1,020.60 | $1,230 Latitude.sh | | [**H100 PCIe**](https://dedigpu.com/gpu/h100-pcie) Hopper · PCIe | **80 GB** HBM2e | 2 TB/s | PCIe 5.0 · NVLink bridge (pairs) | 16 | 128 GB | 1 TB | 25 Gbit/s | $810 | $729.00 | $1,278 Hyperstack (reserved) | | [**MI300X**](https://dedigpu.com/gpu/mi300x) CDNA 3 · OAM | **192 GB** HBM3 | 5.3 TB/s | Infinity Fabric · 896 GB/s | 24 | 240 GB | 2 TB | 100 Gbit/s | $782 | $703.80 | $1,891 DigitalOcean (on-demand ×730 h) | | [**A100 80 GB**](https://dedigpu.com/gpu/a100-80) Ampere · SXM | **80 GB** HBM2e | 2.04 TB/s | NVLink 3 · 600 GB/s | 16 | 128 GB | 1 TB | 100 Gbit/s | $654 | $588.60 | $694 Hyperstack (reserved) | | [**RTX PRO 6000**](https://dedigpu.com/gpu/rtx-pro-6000) Blackwell · PCIe | **96 GB** GDDR7 ECC | 1.79 TB/s | PCIe 5.0 | 16 | 128 GB | 1 TB | 25 Gbit/s | $493 | $443.70 | $949 Hyperstack (reserved) | | [**L40S**](https://dedigpu.com/gpu/l40s) Ada Lovelace · PCIe | **48 GB** GDDR6 ECC | 864 GB/s | PCIe 4.0 | 12 | 96 GB | 750 GB | 25 Gbit/s | $356 | $320.40 | $511 Hyperstack L40 (reserved) | | [**RTX 6000 Ada**](https://dedigpu.com/gpu/rtx-6000-ada) Ada Lovelace · PCIe | **48 GB** GDDR6 ECC | 960 GB/s | PCIe 4.0 | 12 | 96 GB | 750 GB | 25 Gbit/s | $331 | $297.90 | $613 RunPod (on-demand ×730 h) | | [**RTX A6000**](https://dedigpu.com/gpu/a6000) Ampere · workstation | **48 GB** GDDR6 ECC | 768 GB/s | PCIe 4.0 · NVLink bridge (pairs) | 10 | 64 GB | 500 GB | 10 Gbit/s | $208 | $187.20 | $256 Hyperstack (reserved) | | [**RTX 5090**](https://dedigpu.com/gpu/rtx-5090) Blackwell · workstation | **32 GB** GDDR7 | 1.79 TB/s | PCIe 5.0 | 8 | 48 GB | 400 GB | 10 Gbit/s | $163 | $146.70 | $402 CloudRift (3-month reserved) | | [**RTX 4090**](https://dedigpu.com/gpu/rtx-4090) Ada Lovelace · workstation | **24 GB** GDDR6X | 1.01 TB/s | PCIe 4.0 | 8 | 40 GB | 300 GB | 10 Gbit/s | $139 | $125.10 | $274 GPU-Mart | “Cheapest found” is the lowest published monthly price for the same card at a provider with public pricing on 2 September 2026; hourly rates converted at 730 hours. Every price on this page is the sum of a six-line cost sheet, rounded up to the dollar, with zero margin on the server: [the method](https://dedigpu.com/pricing). By workload ## Four jobs, one recommendation each. - ### LLM training & pre-training 8-GPU NVLink nodes with 3.2 Tbps InfiniBand between them. Checkpoints land on local NVMe at 14 GB/s. [We recommend B200 180 GB $1,408/mo](https://dedigpu.com/gpu/b200) - [B300 $1,762](https://dedigpu.com/gpu/b300) - [H200 $1,252](https://dedigpu.com/gpu/h200) - [H100 SXM $1,134](https://dedigpu.com/gpu/h100-sxm) - [A100 80 GB $654](https://dedigpu.com/gpu/a100-80) - ### Fine-tuning & inference LoRA, QLoRA and full fine-tunes on 7B to 70B. vLLM and SGLang templates serve the result from the same server. [We recommend H100 SXM 80 GB $1,134/mo](https://dedigpu.com/gpu/h100-sxm) - [H100 PCIe $810](https://dedigpu.com/gpu/h100-pcie) - [A100 80 GB $654](https://dedigpu.com/gpu/a100-80) - [L40S $356](https://dedigpu.com/gpu/l40s) - [MI300X $782](https://dedigpu.com/gpu/mi300x) - ### Image & video generation Flux, SDXL, Wan and HunyuanVideo on GDDR7 cards that were built for exactly this. ComfyUI is one click away. [We recommend RTX 5090 32 GB $163/mo](https://dedigpu.com/gpu/rtx-5090) - [RTX 4090 $139](https://dedigpu.com/gpu/rtx-4090) - [L40S $356](https://dedigpu.com/gpu/l40s) - [RTX PRO 6000 $493](https://dedigpu.com/gpu/rtx-pro-6000) - ### Rendering & simulation 96 GB of ECC memory for scenes that do not fit anywhere else. Blender, Octane, Redshift and Houdini out of the box. [We recommend RTX PRO 6000 96 GB $493/mo](https://dedigpu.com/gpu/rtx-pro-6000) - [RTX 6000 Ada $331](https://dedigpu.com/gpu/rtx-6000-ada) - [RTX A6000 $208](https://dedigpu.com/gpu/a6000) - [RTX 5090 $163](https://dedigpu.com/gpu/rtx-5090) By class ## Three kinds of machine. ### SXM and OAM nodes HBM memory, NVLink or Infinity Fabric all-to-all, 100 Gbit/s ports and InfiniBand for clusters. Training, large-model serving, long context. - [B300 $1,762](https://dedigpu.com/gpu/b300) - [B200 $1,408](https://dedigpu.com/gpu/b200) - [H200 $1,252](https://dedigpu.com/gpu/h200) - [H100 SXM $1,134](https://dedigpu.com/gpu/h100-sxm) - [MI300X $782](https://dedigpu.com/gpu/mi300x) - [A100 80 GB $654](https://dedigpu.com/gpu/a100-80) ### PCIe data-centre cards Passive, ECC, 25 Gbit/s ports. Single-card inference and fine-tuning, render farms, anything that runs unattended for months. - [H100 PCIe $810](https://dedigpu.com/gpu/h100-pcie) - [RTX PRO 6000 $493](https://dedigpu.com/gpu/rtx-pro-6000) - [L40S $356](https://dedigpu.com/gpu/l40s) - [RTX 6000 Ada $331](https://dedigpu.com/gpu/rtx-6000-ada) ### Workstation cards GDDR memory, the best price per image. Diffusion, video generation, small-model serving, LoRA fine-tuning. - [RTX A6000 $208](https://dedigpu.com/gpu/a6000) - [RTX 5090 $163](https://dedigpu.com/gpu/rtx-5090) - [RTX 4090 $139](https://dedigpu.com/gpu/rtx-4090) Questions ## Choosing a card. **Which GPU server should I rent for LLM inference?** For models up to 70B at FP8, one H200 or one MI300X; for 8B to 32B models, an RTX 5090, L40S or RTX PRO 6000. The memory table on each card's page shows exactly which popular models fit, and the order form checks the model you name before you pay. **What is the cheapest dedicated GPU server?** The RTX 4090 at $139 per month, with 24 GB of GDDR6X, 8 vCPU, 40 GB of RAM and 300 GB of NVMe. Like every card it is billed at its cost sheet, with no setup fee and free egress. **Are the prices the same in every region?** Yes. Reykjavík, Helsinki, Amsterdam and Singapore carry the same per-card price for every model. Pick by latency to your users or your data. **Can I mix cards or change the count later?** A server holds 1, 2, 4 or 8 cards of one model. To change the count or the model, order a new server; the price is linear, so two four-card servers cost the same as one eight-card node, but the eight-card node has the whole NVLink fabric. ## Pick a card, top up, order. An account is an email address. The first term is taken from a crypto balance; root is live in under 5 minutes. [Create an account](https://dedigpu.com/signup) [Read the cost sheet](https://dedigpu.com/pricing)