AMD Instinct MI300X cloud prices

The cheapest MI300X hour we could confirm on a provider's own page on 16 October 2026 is GPUBank's interruptible tier at $0.99 per GPU-hour, and the cheapest non-interruptible one is GPUBank on demand at $1.65 with CPU, RAM, NVMe, an IPv4 address and traffic included 1. A Runpod Community marketplace listing at $0.50 appeared on the MadeByAgents tracker on 15 October, but we could not open it on Runpod's own pages and a single cheap listing can be gone within the hour 15. Per GPU, Azure spot at $1.11 also looks cheap, but it rents only an eight-GPU VM and can be reclaimed 11. Among single-GPU on-demand clouds GPUBank is followed by DigitalOcean at $2.59, Seeweb at about $2.63 and Hot Aisle at $2.99 4 5 7.

An MI300X costs between about $0.50 and $6.00 per GPU-hour in October 2026, and the mode explains most of that range. Marketplace listings sit around $0.50 to $1.99, spot capacity at $0.99 on GPUBank, $1.11 per GPU on Azure and $1.85 on Vultr, and single-GPU on demand from $1.65 on GPUBank to $3.59 on Spheron 15 14 1 11 13 9. Reserved capacity runs from $1.32 on a GPUBank month to $3.85 on a Cirrascale monthly term, and Azure and Oracle on demand cost $6.00 per GPU, sold as $48.00-an-hour eight-GPU machines 16 10. Aggregator medians differ because each samples different offers, so compare the provider rows rather than an average.

You can rent a single MI300X from GPUBank, DigitalOcean, Seeweb, Hot Aisle and Spheron, and marketplace hosts also rent single cards out of larger servers 1 4 5 7 9. Azure, Oracle, Vultr, Crusoe, TensorWave, Cirrascale, Cyfuture AI and Denvr sell only the eight-GPU server or a contract for one, so their per-GPU price is one eighth of a bill that is eight times larger 13. For a single 70B model in 16-bit, one card is enough, since about 141 GB of weights fits in 192 GB with room for the cache. For models above roughly 170B parameters at 8-bit, or for tensor-parallel training, you need several cards in one machine: GPUBank and Seeweb sell 2, 4 or 8, and DigitalOcean sells 8.

Yes, a 70B model fits on one MI300X in 16-bit precision, which is the main reason to rent this card. Llama 3.3 70B has 70.55 billion parameters, which take about 141 GB at two bytes each, leaving roughly 50 GB of the 192 GB for the KV cache and activations, enough for moderate concurrency at typical context lengths 37 3. The same model in 16-bit does not fit on an 80 GB H100 and leaves almost nothing free on a 141 GB H200. In FP8 it takes about 71 GB and fits an H100 too, so the MI300X's advantage shrinks to cache headroom, which still matters for long contexts and many concurrent users. Larger models such as Mixtral 8x22B or Mistral Large 2 fit on one MI300X in FP8 41 40.

The MI300X is about as fast as the H100 in AMD's own 70B benchmarks, slower in several independent tests and faster on very large models. In MLPerf Inference v4.1 on Llama 2 70B, AMD's table puts eight MI300X at 22,021 tokens per second in the server scenario against 21,605 for eight H100 31. AMD's January 2025 tests put it at 0.94 times the H100 on Llama 3.1 70B in FP8 and 1.31 times on 405B 31. SemiAnalysis found it 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 70B 33 34. For a small LoRA run Spheron timed 34 minutes against the H100's 18 9. Benchmark your own workload before committing.

No. As of 16 October 2026 we found no MI300X instance on AWS or Google Cloud in any catalogue or tracker we read, and the only large clouds renting the card are Microsoft Azure and Oracle 11 10 13. Azure sells it as the Standard_ND96isr_MI300X_v5 virtual machine with eight GPUs, 96 vCPU and 1,850 GiB of RAM at $48.00 an hour in East US 2, $6.00 per GPU, with spot at $8.8704 for the whole VM and reservations of one or three years 11 12. Oracle prices the BM.GPU.MI300X.8 bare-metal shape at $6 per GPU-hour, again only as a whole eight-GPU node 10. A team that needs one or two cards rather than eight will find them only on specialised clouds: GPUBank at $1.65, DigitalOcean at $2.59, Seeweb at about $2.63 and Hot Aisle at $2.99, each billed for the card alone 1 4 5 7.

ROCm is ready for production if your stack is PyTorch, vLLM or SGLang, and not yet if you depend on NVIDIA-only libraries. ROCm 10.1.0 shipped on 5 October 2026 with the MI300X listed as supported alongside PyTorch 2.14, vLLM 0.29 and SGLang 0.5.18 builds, and PyTorch publishes official ROCm wheels 17 19. TensorRT-LLM has no ROCm build, FlashAttention 3 exists on ROCm only as an FP8 Triton path, and SGLang lists Marlin, GGUF and ModelOpt quantisation as not working on AMD 26 23 22. Backend choice matters too: vLLM measured its AITER attention up to 3.82 times faster than the older ROCm path on the MI300X 21. On GPUBank, which launches any Docker image, start from a ROCm image such as vLLM's official one rather than the CUDA templates 2 20.

The MI300X spot price is worth the risk for checkpointed batch work and not for endpoints people depend on. GPUBank's interruptible tier saves 40 percent ($0.99 against $1.65), gives five minutes of notice and a checkpoint hook, stops the machine rather than deleting it and credits 30 minutes if the notice is missed 1. Azure spot at $1.11 per GPU saves 82 percent against its own $6.00 on-demand price, but comes as an eight-GPU VM with quota to request 11. Seeweb's spot terminates the instance after a two-minute notice, and its page shows no spot price below the standard one 5. With checkpoints every 30 minutes and a 5 percent rework allowance, 100 eight-GPU hours with 2 TB of disk cost about $892 on GPUBank spot against $1,380 on demand.

The MI300X is cheaper than the H100 on every provider we checked that sells both, because fewer buyers compete for it and its software is less mature, so sellers price it to win customers away from CUDA. On GPUBank it costs $1.65 against $1.85 for the H100 SXM 1. On DigitalOcean the gap is wider, $2.59 against $4.41 per GPU for an H100 Droplet, even though the MI300X has 192 GB of memory against 80 4. SemiAnalysis argued in 2025 that the MI300X needed to cost under $2.10 to $2.40 an hour on one-month contracts to match the H200 on performance per dollar for reasoning workloads, and most specialist clouds now price it below that 33. Whether it is cheaper per job depends on the model size, as the performance section shows.

Rent the MI300X unless your model needs between 192 and 256 GB on a single card. The MI325X uses the same CDNA 3 design with 256 GB of HBM3E, so it runs the same ROCm software and adds memory and bandwidth rather than new features 4. On DigitalOcean, the main cloud that lists both by the single GPU, the MI325X costs $3.80 an hour on demand against $2.59 for the MI300X, 47 percent more, and $2.88 against $1.91 on a 12-month reservation 4. On GPUBank the MI300X costs $1.65 with the whole machine included 1. A 70B model in 16-bit, a 100B to 140B model in FP8 or a 235B mixture of experts at 4-bit already fits in 192 GB with room for the cache, so the extra 64 GB pays only for longer contexts, larger batches or models just beyond that line, and for those a machine with two MI300X cards is the other option to price.

A month of one MI300X running continuously costs $1,204.50 on GPUBank on demand ($1.65 times 730 hours), $963.60 on a GPUBank monthly reservation, $1,890.70 on DigitalOcean on demand, $1,394.30 on DigitalOcean's 12-month reservation, about $1,923 on Seeweb and $2,182.70 on Hot Aisle 1 4 5 7. Add storage where it bills separately: on GPUBank, 200 GB of persistent disk adds $6.00. Funding GPUBank through $1,000 top-ups, which carry a 60 percent bonus, brings the cash cost of an on-demand month to about $753 1. An eight-GPU month on Azure or Oracle at $48.00 an hour costs $35,040 before disks and egress 11 10.

Whether a stopped MI300X bills depends entirely on the provider, and this is where many invoices go wrong. On GPUBank a stopped machine bills only its persistent disk at $0.03 per GB-month, so compute, the IPv4 address and the local NVMe cost nothing while it is stopped 1. On DigitalOcean a powered-off GPU Droplet bills its full rate until it is destroyed, and destroying it discards the scratch disk 4. Spheron's MI300X instances cannot be stopped at all, only destroyed 9. Seeweb and Hot Aisle do not state the stopped-state charge on the pages we read, and on Azure a deallocated VM stops its compute meter while its managed disks keep billing. If you work in bursts, compare the stopped-state charge before comparing hourly rates.

Per-second billing means a 61-minute job costs 61 minutes, while hourly rounding bills the same job as 120 minutes, 97 percent more. On GPUBank a 61-minute session costs $1.65 times 61 divided by 60, or $1.68, and a 90-second test about $0.04, with no minimum 1. DigitalOcean bills per second with a 5-minute minimum, which only affects very short tests 4. Hot Aisle bills by the minute 7. Spheron's 20-minute minimum makes even a 5-minute test cost $1.20 at $3.59 9. Seeweb bills by the hour and does not say how it rounds a partial hour; across 40 development sessions of 61 minutes, minute rounding there would come to about $107 a month and hour rounding to about $211 5.

Buying a new MI300X server is not cheaper than renting from GPUBank even at full use, on our assumptions. A new eight-GPU Supermicro server was listed at 278,193 euros before VAT on 16 October 2026, about $311,800 35 6. With $110,000 of resale after two years, 8.5 kW at an assumed $0.15 per kWh and an assumed $2,000 a month of colocation, owning costs about $10,408 a month plus $1.275 per node-hour. Renting eight GPUs from GPUBank costs $13.20 per node-hour on demand, so break-even would need about 873 hours a month, more than a month has, and a reserved month at $7,708.80 is cheaper still 1. A used server near $200,000 brings break-even against on demand to about 482 hours, but owning still loses to the reservation.

Seeweb is the only provider we found renting single MI300X GPUs from EU data centres, at 2.35 euros an hour from its Italian sites with traffic free and commitments down to 2.00 euros for twelve months 5. GPUBank's European regions, Paris and Frankfurt, do not carry the MI300X; it lists the card only in Ashburn 2. Azure lists the ND MI300X v5 in 17 regions according to CloudPrice, at up to $7.86 per GPU outside East US 2 on GetDeploying, and Oracle prices the shape at $6 per GPU-hour, both only as eight-GPU machines 12 13 10. For EU data residency on one card, Seeweb is the practical choice, and for a full node the large clouds' European regions are the alternative.