Where to rent an MI300X at the cheapest price

ROCm is no longer the obstacle it was in 2024, and the version numbers show how fast it moves. AMD released ROCm 10.1.0 on 5 October 2026, built with its new TheRock system, and its release notes list the MI300X (gfx942) as supported with PyTorch 2.14.0, JAX 0.11.1, vLLM 0.29.0 and SGLang 0.5.18²⁴²⁵. PyTorch publishes official 2.14.1 wheels for ROCm 7.2 and 7.14, vLLM documents MI300 support from ROCm 6.3 and now ships its own ROCm images in place of AMD's deprecated rocm/vllm ones, Triton lists AMD GPUs on ROCm 6.2 and later as supported hardware, and bitsandbytes says all its features work on CDNA, with gfx942 in its PyPI wheels^26172728. A team on stock PyTorch, vLLM or SGLang with a Hugging Face model usually has an endpoint running the same day.

The gaps are in everything tied to NVIDIA. TensorRT-LLM is an NVIDIA-only library and has no ROCm build; FlashAttention 2 runs on the MI300X through its Composable Kernel and Triton backends, but the Hopper-specific FlashAttention 3 exists on ROCm only as an FP8 path in the Triton backend; and SGLang's own AMD page says that quantisation methods depending on Marlin or NVIDIA kernels, including awq_marlin, gptq_marlin, gguf, modelopt_fp8 and modelopt_fp4, do not work²⁹³⁰³¹. Hugging Face TGI is in maintenance mode and its README names only the MI210 and MI250, DeepSpeed says it is developed and tested on the MI100 and MI200, and xFormers labels its ROCm wheel experimental³²³³³⁴. The kernel backend also decides speed: vLLM's February 2026 post measured its AITER attention up to 3.82 times faster than the older ROCm path on the MI300X³⁵.

In engineering time this splits teams cleanly. If your stack is PyTorch with standard layers, vLLM or SGLang serving a popular open model in BF16 or FP8, or llama.cpp through its HIP backend, budget an afternoon to pull the right ROCm image, pin versions and benchmark the attention backend³⁶. If you depend on TensorRT-LLM engines, Marlin-quantised checkpoints, custom CUDA or PTX kernels, or a training framework tested only on NVIDIA, budget days to weeks: kernels have to be ported through HIP or rewritten in Triton, quantised weights re-exported in a format ROCm supports, and every upgrade retested, because ROCm, PyTorch and vLLM versions must match. At $1.65 against $1.85 for an H100, the MI300X saves $0.20 an hour, about $146 a month per card at full use, so the port pays for itself only when the memory advantage, not the hourly rate, is what you are buying.

The AMD Instinct MI300X is a CDNA 3 accelerator launched on 6 December 2023 with 304 compute units, 192 GB of HBM3 on an 8,192-bit interface at 5.3 TB/s, and peak dense throughput of 1.3 PFLOPS in FP16 and 2.61 PFLOPS in FP8, at a typical board power of 750 W⁹. It ships only as an OAM module, mounted eight to a baseboard and joined by eight Infinity Fabric links of 128 GB/s each, so there is no PCIe version for a workstation and every MI300X you rent sits in an eight-GPU server, whether or not the provider sells you all eight⁹. Its successors are the MI325X with 256 GB, which DigitalOcean already rents at $3.80, and the CDNA 4 MI350X and MI355X with 288 GB³. The memory is the whole argument: 192 GB holds a 70B model in 16-bit weights on one card, which no 80 GB H100 can do, and that alone explains most of the demand for it.

The market splits into two camps. One sells single GPUs: GPUBank in machines of 1, 2, 4 or 8, DigitalOcean in 1 or 8, Seeweb in 1, 2, 4 or 8, Hot Aisle in virtual machines of 1, 2 or 4 and Spheron in 1 or 2²³⁷⁶¹⁰. The other sells only the eight-GPU server or a contract for one: Microsoft Azure with its ND MI300X v5, Oracle with BM.GPU.MI300X.8, Vultr, Crusoe, TensorWave, Cirrascale, Cyfuture AI, Denvr and Hot Aisle's bare-metal tier¹¹¹²¹³. The buyers are mostly inference teams serving models between 70B and 400B parameters, and price-sensitive teams who will trade CUDA's convenience for a lower hourly rate, and between those two camps of sellers the price on providers' own pages runs from $0.99 to $6.00 per GPU-hour for the same silicon.

Three things must match before two MI300X prices can be compared. The first is the mode: on-demand capacity is yours until you stop it, spot or interruptible capacity can be taken back, and reserved capacity is prepaid or contracted for a week to three years. The second is the unit, because a per-GPU price on Azure, Oracle, Vultr, Crusoe or Cirrascale is the price of an eight-GPU server divided by eight, so the smallest bill you can run up there is eight times the headline⁴⁵¹². The third is what the hour includes, from vCPU, RAM and local disk to the public IP, outbound traffic and the disk that keeps billing while the machine is off. A marketplace headline is a host's asking price for a GPU in whatever box it sits in; a managed price is a defined machine with defined inclusions. They are not the same product, so the table below sorts offers by type first.

For this card the five kinds of offer look like this. Managed single-GPU on demand runs from $1.65 on GPUBank through $2.59 on DigitalOcean, about $2.63 on Seeweb and $2.99 on Hot Aisle to $3.59 on Spheron, with Crusoe at $3.45 and Cyfuture AI at $2.79 selling only whole nodes^2376101311. Marketplaces are thin: the only live listings we found were Runpod Community at $0.50 and OpenRelay at $1.99, both through trackers¹¹⁴. Spot capacity costs $0.99 on GPUBank for one GPU, $1.11 per GPU on an eight-GPU Azure VM and $1.85 on Vultr, which GetDeploying showed sold out²⁴¹¹. Reservations start at $1.32 on GPUBank for a week or a month, followed by $1.62 on a 12-month Cyfuture node and $1.91 on a 12-month DigitalOcean contract. We found no serverless MI300X billed by the second of work.

The captures below were taken on 16 Oct 2026 from each provider's own MI300X or pricing page, so every figure can be read next to the conditions printed beside it.

GPUBank wins on the MI300X wherever the buyer wants one card, by the hour, with the machine priced in. Its $1.65 on-demand rate is the cheapest single-GPU MI300X we could confirm on a provider's own page, $0.94 an hour below DigitalOcean and about a dollar below Seeweb and Hot Aisle, and it already covers 24 vCPU, 240 GB of RAM, 2 TB of NVMe, an IPv4 address and all traffic²³⁷⁶. The $0.99 interruptible tier is the cheapest MI300X hour on any provider's own page, and the $1.32 reservation is the only one we found that runs for a week or a month rather than a year. Per-second billing with no minimum, no charge for a stopped machine beyond its disk, and a full credit when a host fails mid-job matter more on this card than on most, because its rivals bill a powered-off Droplet in full or refuse to stop at all²³¹⁰.

It loses in several places. The card is listed only in Ashburn, so a European data-residency requirement points to Seeweb instead¹⁶⁷. A marketplace listing such as Runpod Community at $0.50 is half the price of the interruptible tier when one exists¹. Payment is a prepaid balance topped up in crypto from $25, with no card payments for now, which some finance departments cannot accept; the top-up bonus brings the effective on-demand hour to about $1.22 with $100 and $1.03 with $1,000². Finally, GPUBank's ready-made templates are listed with CUDA builds such as PyTorch 2.8 with CUDA 12.8, so on this card plan to launch your own Docker image, which GPUBank accepts, starting from the official vLLM ROCm image¹⁶¹⁷.

Beyond the providers shown as cards, the MI300X is sold by a handful of node sellers and two small marketplaces. Cyfuture AI lists an eight-GPU node with 128 vCPU and 1.5 TB of RAM at $2.79 per GPU-hour on demand and $1.62 on a 12-month term, and Cirrascale's own pricing page asks $17,999 a month for an eight-GPU server on an annual term or $22,499 on a monthly term, which is $3.08 and $3.85 per GPU-hour over a 730-hour month¹¹¹². Vultr's MI300X appeared on ComputePrices at $1.75 per GPU on an eight-GPU server on 15 October and on GetDeploying as a $1.85 spot cluster that was sold out; Vultr's own page refused our requests, and Spheron describes the $1.75 as a 24-month prepay, so treat it as a long-commitment rate¹⁴¹¹¹⁰.

TensorWave's often-quoted $1.71 is a starting point for a sales conversation: GetDeploying files it under custom terms with data from 31 August, and TensorWave's site publishes no rate card¹¹. Denvr lists eight-GPU nodes on a waitlist with no price, and GetDeploying shows Runpod's 1x MI300X Secure Cloud at $2.39, out of stock, while Runpod's own pricing page no longer lists the card¹¹. On the marketplace side, ComputePrices shows OpenRelay at $1.99 for a single GPU in community and verified tiers, Aquanode at $2.63 and AtmosCompute at $3.35, none of which we could confirm on the providers' own pages¹⁴. The Vast.ai MI300X at $0.49 that a tracker reported in September no longer appeared in the trackers we read on 16 October, which is what a single cheap marketplace listing usually does.

Azure and Oracle are the only large clouds that rent the MI300X, and both sell only the eight-GPU machine. Azure's Standard_ND96isr_MI300X_v5, with 96 vCPU and 1,850 GiB of RAM, costs $48.00 an hour on Linux in East US 2 according to Azure's own retail price API, which is $6.00 per GPU⁴¹⁵. The same API gives $8.8704 an hour for spot, or $1.11 per GPU; $269,107 for a one-year reservation, or $3.84 per GPU-hour; and $553,772 for three years, or $2.63 per GPU-hour⁴. CloudPrice lists the VM in 17 regions and GetDeploying shows up to $7.86 per GPU outside East US 2, and managed disks and internet egress bill on top at Azure's usual rates¹⁵¹¹.

Oracle prices the MI300X at $6 per GPU-hour, or $48.00 an hour for the BM.GPU.MI300X.8 bare-metal node with 224 vCPU and 2 TB of RAM that GetDeploying lists⁵¹¹. AWS and Google Cloud have no MI300X instance in any catalogue or tracker we read. At list price the two large clouds cost 3.6 times GPUBank's on-demand rate ($6.00 against $1.65), so their case rests on things other than price: an existing enterprise agreement, compliance work already done on that cloud, private networking to the rest of your estate, or Azure's spot price, which at $1.11 per GPU is cheap but comes in eight-GPU units with quota to request and short eviction notice.

The MI300X is the right rental in three cases: serving a model too large for one 80 GB card but small enough for 192 GB, such as a 70B model in 16-bit or a 100B to 140B model in FP8, where one MI300X replaces two H100s and the tensor-parallel traffic between them; memory-bound serving of very large models across a node, such as Llama 3.1 405B or DeepSeek-class mixtures of experts; and long-context serving where the KV cache, not compute, sets the batch size. It stops being the right tool for latency-critical serving at low batch, for prefill-heavy work, for anything built on TensorRT-LLM or custom CUDA kernels, and for small-model fine-tuning, where measured wall-clock times favour the H100. The figures below were published by others with the settings shown; we have not rerun them, and they disagree with each other widely, mostly because AMD's own tests pair its best software with a well-tuned setup while independent testers use stock builds.

The most reproducible comparison is MLPerf Inference v4.1 on Llama 2 70B in FP8: AMD's table of submissions 4.1-0070 and 4.1-0043 puts eight MI300X at 22,021 tokens per second in the server scenario against 21,605 for eight H100, and EE Times reports AMD's single-MI300X entry at 3,062.72 tokens per second offline and 2,520.27 in the server scenario, with the eight-GPU entry at 21,028.2¹⁸¹⁹. On Llama 3.1 in FP8, AMD's January 2025 tests on vLLM 0.6.7 and ROCm 6.3 against TensorRT-LLM 0.16 put the MI300X at 0.94 times the H100 on 70B and 1.31 times on 405B¹⁸. Independent work is harsher: SemiAnalysis found that for most scenarios the MI300X is not competitive with the H200 and beats the H100 only on Llama 3 405B and DeepSeek-V3, and arXiv paper 2510.27583 measured 37 to 66 percent of H100 and H200 performance on Llama 70B²⁰²¹.

Per job at GPUBank's prices the answer depends on whose numbers you trust. On the MLPerf server figures, one MI300X produces about 9.91 million tokens an hour, so at $1.65 a million tokens cost $0.167, against $0.190 on an H100 at $1.85. On AMD's 0.94 ratio the MI300X costs 0.95 times the H100 per token, on the 405B ratio 0.68 times, and on the arXiv range between 1.35 and 2.41 times as much. For fine-tuning, Spheron timed a 375-step LoRA run at about 34 minutes on the MI300X against 18 on the H100, which is $0.94 against $0.56 per run at our prices¹⁰. The pattern is consistent: the bigger the model relative to 80 GB, the better the MI300X looks per job.

Against the H100 SXM at $1.85, the MI300X wins outright whenever the model needs more than 80 GB: a 70B model in 16-bit runs on one MI300X at $1.65 an hour, while the H100 needs two cards and tensor parallelism at $3.70. When the model fits in 80 GB, for example a 70B model in FP8 or anything up to about 30B in 16-bit, the H100 usually finishes the job for less money despite its higher rate, because its software is more mature and independent benchmarks put it ahead on 70B-class serving²¹²⁰. Against the H200 SXM at $2.49, with 141 GB at 4.8 TB/s, the MI300X is $0.84 cheaper an hour and holds 51 GB more, but SemiAnalysis found the H200 faster in most scenarios, so the MI300X wins per job only on models or contexts that overflow 141 GB.

The B200 SXM at $4.15, with 180 GB at 8 TB/s and FP4 support, is in a different class for throughput, and for a team serving a large model at high volume its cost per token can undercut the MI300X despite an hourly rate 2.5 times higher; we have no like-for-like measurement to put a number on it. The practical rule is this: pick the MI300X when memory per card is the constraint and your stack is PyTorch, vLLM or SGLang; pick the H100 when the model fits in 80 GB and you depend on CUDA-only tools; pick the H200 when you need more than 80 GB and want CUDA; and pick the B200 when throughput per card matters more than the hourly rate.