*[DediGPU](https://dedigpu.com/) — Markdown mirror of [https://dedigpu.com/gpu/mi300x](https://dedigpu.com/gpu/mi300x) · updated 2026-09-02 · index for LLMs: [llms.txt](https://dedigpu.com/llms.txt) · everything: [llms-full.txt](https://dedigpu.com/llms-full.txt)*

AMD · CDNA 3

# Dedicated AMD Instinct MI300X server, $782 a month.

The whole card — 192 GB of HBM3 at 5.3 TB/s — with 24 dedicated vCPU, 240 GB of RAM and 2 TB of NVMe, on a 100 Gbit/s port with free egress. Root in under 5 minutes. Priced from a cost sheet you can read below, with zero margin on the server.

- 1× to 8× cards, full Infinity Fabric · 896 GB/s
- 4 regions, same price
- No setup fee, no KYC, crypto balance
- 99.9% SLA credited ×10

- **192 GB** HBM3 per card
- **5.3 TB/s** memory bandwidth
- **896 GB/s** Infinity Fabric
- **100 Gbit/s** port, $0 egress
- **750 W** board power, in the cost sheet

The cost sheet

## What an MI300X costs us, per card, per month.

Six lines, added up, rounded up to the next dollar. That is the price. The assumptions are the same for every card and are explained on the [pricing page](https://dedigpu.com/pricing#method).

- Hardware $15,000 card + $8,750 chassis share, over 36 months $659.72
- Power 558 kWh at $0.052 · 950 W, 70% load, PUE 1.15 $29.03
- Rack & cooling HGX baseboard, liquid-assisted $45.00
- Network & IP transit share, one IPv4, port $9.00
- Spares & failures 2% of the card price per year $25.00
- Operations monitoring, on-call, provisioning $14.00
- Total rounded up to the next dollar $781.75 → $782

Cheapest published price we found for the same card: $1,891 per month at DigitalOcean (on-demand ×730 h) (on-demand), checked 2 September 2026 — 59% more than our cost. Where the money comes from, then: options with a normal margin, power bought at scale, full racks, no card network. [Read how](https://dedigpu.com/pricing#margin).

Configurations

## One card to eight. The price is linear.

CPU, RAM and NVMe scale with the cards. Every count sits on the same 8-way baseboard with the full fabric between your cards.

| Server | GPU memory | CPU | RAM | NVMe | Network | Per month | 12-month rate |  |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| **1× MI300X** | 192 GB | 24 vCPU | 240 GB | 2 TB | 100 Gbit/s | $782 | $703.80 | [Order](https://dedigpu.com/signup?gpu=mi300x&n=1) |
| **2× MI300X** | 384 GB | 48 vCPU | 480 GB | 4 TB | 100 Gbit/s | $1,564 | $1,407.60 | [Order](https://dedigpu.com/signup?gpu=mi300x&n=2) |
| **4× MI300X** | 768 GB | 96 vCPU | 960 GB | 8 TB | 100 Gbit/s | $3,128 | $2,815.20 | [Order](https://dedigpu.com/signup?gpu=mi300x&n=4) |
| **8× MI300X** | 1536 GB | 192 vCPU | 1920 GB | 16 TB | 100 Gbit/s + IB | $6,256 | $5,630.40 | [Order](https://dedigpu.com/signup?gpu=mi300x&n=8) |

Per month for a one-month term; the 12-month rate is the same server on a 12-month term (−10 %). Options — extra NVMe, IPv4, 25 Gbit/s uplink, private VLAN — are added on the order form at the prices on the [pricing page](https://dedigpu.com/pricing#options).

Terms

## Prepay longer, pay less per month.

A longer term costs us less to run — one provisioning, no re-image between tenants, a firmer power forecast — and the difference is passed on. The rate is locked for as long as the server renews.

| Term | Discount | Per month | Charged at order | Saved over the term |
| --- | --- | --- | --- | --- |
| **1 month** | cost sheet | $782.00 | $782.00 | — |
| **3 months** | −3 % | $758.54 | $2,275.62 | $70.38 |
| **6 months** | −6 % | $735.08 | $4,410.48 | $281.52 |
| **12 months** | −10 % | $703.80 | $8,445.60 | $938.40 |

For one MI300X; multiply by the card count. Renewal takes the same term at the same rate. [How renewals, grace and suspension work](https://dedigpu.com/docs/terms-and-renewals).

Specifications

## Per card, before you multiply.

- **GPU:** AMD Instinct MI300X
- **Architecture:** CDNA 3
- **GPU memory:** 192 GB HBM3
- **Memory bandwidth:** 5.3 TB/s
- **Interconnect:** Infinity Fabric · 896 GB/s
- **Form factor:** OAM module on an 8-way baseboard
- **Board power:** 750 W
- **CPU per card:** 24 dedicated vCPU
- **System RAM per card:** 240 GB ECC
- **Local NVMe per card:** 2 TB
- **Network:** 100 Gbit/s · one IPv4 · $0 egress
- **Cards per server:** 1×, 2×, 4×, 8×
- **Out-of-band:** IPMI console, power control, virtual media

What fits

## Popular models on 1× MI300X.

Memory needed = parameters × bytes per parameter × 1.2 (KV cache and runtime). Change the card count above and the table follows. [The full guide](https://dedigpu.com/guides/how-much-vram-to-run-an-llm).

| Model | BF16 | FP8 | INT8 | INT4 |
| --- | --- | --- | --- | --- |
| Llama 3.1 8B | 20 GB | 10 GB | 10 GB | 5 GB |
| Mistral Small 3.1 24B | 58 GB | 29 GB | 29 GB | 15 GB |
| Gemma 3 27B | 65 GB | 33 GB | 33 GB | 17 GB |
| Qwen3 32B | 77 GB | 39 GB | 39 GB | 20 GB |
| Llama 3.3 70B | 168 GB | 84 GB | 84 GB | 42 GB |
| Qwen2.5 72B | 173 GB | 87 GB | 87 GB | 44 GB |
| Llama 4 Scout · 109B MoE | 262 GB | 131 GB | 131 GB | 66 GB |
| Mixtral 8x22B · 141B MoE | 339 GB | 170 GB | 170 GB | 85 GB |
| Qwen3 235B-A22B · MoE | 564 GB | 282 GB | 282 GB | 141 GB |
| Llama 3.1 405B | 972 GB | 486 GB | 486 GB | 243 GB |
| DeepSeek-V3 / R1 · 671B MoE | 1611 GB | 806 GB | 806 GB | 403 GB |

Green fits in 192 GB with a working KV cache; red does not. FP8 needs Ada, Hopper, Blackwell or MI300X silicon. MoE models count every expert.

Software

## Pre-installed before you log in.

Pick the system and the template on the order form; serving templates pull the model you name before first boot. [Every template, with its ports](https://dedigpu.com/templates).

### Operating systems

- **Ubuntu 24.04 LTS** Kernel 6.8 · the default for every template
- **Ubuntu 22.04 LTS** Kernel 5.15 HWE · for stacks pinned to CUDA 11.8 / 12.1
- **Debian 12** Bookworm · minimal, driver from the vendor repository
- **Rocky Linux 9** RHEL-compatible · Slurm, enterprise and HPC stacks

### Templates

- Bare OS + driver
- PyTorch 2.7
- vLLM 0.9 + model
- Text Generation Inference 3 + model
- Ollama + Open WebUI + model
- Docker + GPU toolkit

Root over SSH, IPMI console and an API token come with every server. [Access docs](https://dedigpu.com/docs/ssh-and-root-access).

What it is for

## The jobs this card is right for.

- ### Fine-tuning & inference
  LoRA, QLoRA and full fine-tunes on 7B to 70B. vLLM and SGLang templates serve the result from the same server.
  Also a fit. Our first pick here is the [H100 SXM](https://dedigpu.com/gpu/h100-sxm) at $1,134/mo.

Regions

## Deployable in four regions, at the same price.

- **Reykjavík** 18 ms London · 40 ms New York 100% geothermal + hydro
- **Helsinki** 27 ms Frankfurt · 36 ms London 100% wind + hydro
- **Amsterdam** 7 ms Frankfurt · 12 ms London 100% wind
- **Singapore** 38 ms Tokyo · 45 ms Sydney Grid + certified RECs

[Latency tables, power and cooling per region](https://dedigpu.com/regions).

Alternatives

## Cards near this price.

Every one billed the same way. [Compare all 13 side by side](https://dedigpu.com/gpu).

- [H100 PCIe 80 GB HBM2e · Hopper $810 /mo · at cost](https://dedigpu.com/gpu/h100-pcie)
- [A100 80 GB 80 GB HBM2e · Ampere $654 /mo · at cost](https://dedigpu.com/gpu/a100-80)
- [RTX PRO 6000 96 GB GDDR7 ECC · Blackwell $493 /mo · at cost](https://dedigpu.com/gpu/rtx-pro-6000)
- [H100 SXM 80 GB HBM3 · Hopper $1,134 /mo · at cost](https://dedigpu.com/gpu/h100-sxm)

Guides

## Read before you rent.

- [Inference How much VRAM do you need to run an LLM? Every size, every precision A single rule, a table for eleven popular models from 8B to 671B at BF16, FP8, INT8 and INT4, and the cheapest dedicated server that fits each one. 2 September 2026 · 9 min read](https://dedigpu.com/guides/how-much-vram-to-run-an-llm)
- [Inference Serve Llama 3.3 70B with vLLM on a dedicated GPU server, step by step Pick the server that fits, pre-load the model at order time, check the OpenAI-compatible API, tune context and parallelism, and put a token in front of it. 2 September 2026 · 10 min read](https://dedigpu.com/guides/serve-llama-70b-with-vllm)

Questions

## About the MI300X server.

**How much does a dedicated MI300X server cost per month?**

$782 per card per month for a one-month term, which is the sum of the six lines of its cost sheet rounded up to the dollar: hardware over 36 months, power at our contract rate, rack and cooling, network and IP, spares, operations. A 3, 6 or 12-month term is 3, 6 or 10 % cheaper per month ($703.80 on 12 months). No setup fee, no egress charge, no contract beyond the term.

**Is the MI300X dedicated or shared?**

Dedicated. You rent a physical server with the whole card: no vGPU profile, no time-slicing, no other tenant on the machine. nvidia-smi shows the full 192 GB because you have the full 192 GB. The CPU cores, the RAM and the NVMe are yours too, and so is the IPMI console.

**How many MI300X cards can I have in one server?**

1, 2, 4 or 8, at $782 per card: the chassis share is already in each card's cost, so the price is linear. Every count sits on the same 8-way baseboard with the full Infinity Fabric · 896 GB/s fabric between your cards; eight-card nodes add 3.2 Tbps InfiniBand for clusters of 16 to 512 cards.

**Which LLMs fit on an MI300X?**

With the fit rule (parameters × bytes per parameter × 1.2), one card holds up to Mixtral 8x22B · 141B MoE at FP8 and up to Qwen3 235B-A22B · MoE at INT4. An eight-card server (1536 GB) holds up to DeepSeek-V3 / R1 · 671B MoE at FP8. The order form checks the model you name against the server you pick before you pay.

**Which operating systems and templates run on the MI300X?**

Ubuntu 24.04 LTS, Ubuntu 22.04 LTS, Debian 12, Rocky Linux 9. Templates: Bare OS + driver, PyTorch 2.7, vLLM 0.9, Text Generation Inference 3, Ollama + Open WebUI, Docker + GPU toolkit. Serving templates pull the model you name before first boot.

**How long does provisioning take, and can I cancel?**

A single-card server is online in under 5 minutes, an eight-card node in under fifteen minutes, in any of the four regions. The term renews automatically from your balance at the locked price; turn renewal off and the server stops at the end of the paid term, or cancel at any time. No notice period.

**What is the cheapest MI300X rental elsewhere?**

The lowest published monthly price we found for the same card on 2 September 2026 was $1,891 at DigitalOcean (on-demand ×730 h) (on-demand; hourly rates converted at 730 hours). That is 59 % more than our cost sheet. We re-check every month and print the date.

## Order an MI300X now.

$782 a month at cost, online in under 5 minutes. Create an account, top up in crypto, pick a region.

[Order this server](https://dedigpu.com/signup?gpu=mi300x) See the cost sheet