NebiusEU-NorthOn-demand
$1.99/ GPU / hr
Open provider
Cloud GPU · verified 2026-09-05
AMD MI300X
Instinct192 GB
192 GB HBM3 at 5.3 TB/s on CDNA 3. The CUDA-free way to fit a 70B on one GPU. The software is the product. If you are not ready for ROCm, this is an expensive science project.
VRAM192 GB
From /GPU/hr$1.99
ArchitectureInstinct
VendorAMD
Sponsored
Featured MI300X offers
The four cheapest in-stock on-demand MI300X hours in this sample. Each card is a sponsored placement.
CrusoeUS-WestOn-demand
$2.29/ GPU / hr
Open provider
CoreWeaveUS-EastOn-demand
$2.49/ GPU / hr
Open providerSponsored does not change ranking. Commissions never move a row, a GPU, or a displayed price.
In brief
What to know before you rent
- 192 GB is why you are here, more than H100, more than H200.
- FP8 ~2,615 TFLOPS dense. Competitive, not a CUDA clone.
- Infinity Fabric 8-GPU boards, 750 W, OAM, not a PCIe GeForce.
Catalog listings
MI300X Pricing and Availability
Published on-demand rates first. Spot and waitlists are labeled. Confirm the live rate before you provision.
ProvidersRegionBilling typeInterconnect/GPU/hr
Our reading
Who this GPU is for
I send people to MI300X when the bytes do not fit NVIDIA 80/141 GB and the team can live in ROCm, or already does. I do not send a PyTorch-CUDA-only shop here to “save on VRAM” and then spend two weeks on kernels. Memory is useless if the step never starts.
Rent it if ROCm-fluent labs, 70B+ that wants one GPU of HBM, HPC+AI mixes that like CDNA FP64.
Skip it if CUDA-only stacks, people who needed a Vast.ai 4090, anyone who thought AMD was a drop-in H100.
How we compare hoursDatasheet
Technical specifications
AMD Instinct MI300X · Per GPU · OAM
Compute · dense
- FP8
- 2,615 TFLOPS~5,230 with sparsity
- FP16 / BF16
- 1,307 TFLOPS~2,615 with sparsity
- TF32
- 654 TFLOPS~1,307 with sparsity
- INT8
- 2,615 TOPS~5,230 with sparsity
- FP32
- 163 TFLOPSmatrix and vector class
- FP64
- 82 TFLOPS163 TFLOPS matrix
- Precision support
- FP8, FP16, BF16, TF32, FP32, FP64, INT8
Memory
- Capacity
- 192 GB HBM3
- Bandwidth
- 5,300 GB/s
- Bus width
- 8,192-bit
- Infinity Cache
- 256 MB
- ECC
- Yes
Silicon
- Architecture
- CDNA 3
- Process
- 5 nm / 6 nm FinFET chiplets
- Compute units
- 304
- Stream processors
- 19,456
- Matrix cores
- 1,216
Fabric and host
- GPU interconnect
- Infinity Fabric, 7 links~896 GB/s aggregate in 8-GPU UBB
- Host interface
- PCIe 5.0 x16
Power
- Board power
- 750 W
Platform
- Partitioning
- SR-IOV, up to 8 partitions
MI300X is an OAM on a UBB, not a dual-slot gaming card. Cloud hours are whole accelerators in AMD-shaped servers. If the provider’s image is CUDA-only, you rented the wrong cloud.
Trade-offs
Pros and cons of MI300X
What this GPU does well, and when the hour is the wrong buy.
Strengths
- 192 GB HBM3 at 5.3 TB/s, the fit argument is real
- Strong FP64/matrix vs Hopper’s AI-first mix
- 8-GPU UBB with Infinity Fabric is a serious node
- When NVIDIA SKUs are waitlisted, AMD is sometimes just… available
Limits
- ROCm is the tax. Pay it or leave
- Fewer on-demand rows on this index than Hopper
- Kernel and framework gaps still ambush greenfield teams
- 750 W OAM, not a consumer slot
ROCm is the interconnect
People compare TFLOPS to H100 and skip the framework. That is how AMD reviews become fan fiction. If vLLM/ROCm, PyTorch, and your flash-attn story are ready, 192 GB is a weapon. If not, H200 is the boring adult choice.
192 GB vs NVIDIA memory SKUs
H100 80, H200 141, B200 ~192. MI300X sits in the B200-bytes club at Hopper-ish generation software risk. I use it when NVIDIA 141 GB still spills and CUDA is not a religion.
How I actually rent it
- 01
Prove ROCm
Run the image on a small job first. Do not discover ROCm on epoch 1 of a 70B.
- 02
Rent the node shape
Single GPU vs 8× UBB. Fabric matters for the same reasons NVLink does.
- 03
Keep an NVIDIA backup
If the kernel path dies, H200 should already be in the tab.
FAQ
MI300X questions
Who has the cheapest MI300X?+
As of September 5, 2026, the cheapest verified in-stock on-demand MI300X tracked by GPUBeacon is $1.99 per GPU per hour from Nebius.
How much does the MI300X cost per month?+
As of September 5, 2026, at 720 hours per month, one MI300X costs an estimated $1,649 at the median on-demand price in this sample. Cheapest verified in stock: $1,433 per month on-demand.
Where can I rent an MI300X?+
GPUBeacon currently tracks MI300X configs from 3 providers. The cheapest verified in-stock on-demand configs come from Nebius. See the full price comparison above for every provider and config.
Are MI300X prices going up or down?+
GPUBeacon publishes a catalog sample, not a 90-day price index. As of September 5, 2026, the median on-demand MI300X in this sample is $2.29 per GPU per hour. The cheapest verified in-stock hour is $1.99.
Why choose the MI300X?+
ROCm-fluent labs, 70B+ that wants one GPU of HBM, HPC+AI mixes that like CDNA FP64.
When is the MI300X not a good fit?+
CUDA-only stacks, people who needed a Vast.ai 4090, anyone who thought AMD was a drop-in H100.
Is MI300X faster than H100?+
On raw 192 GB fit, it can win by not splitting. On CUDA-optimized FP8 training, H100/H200 often still win wall-clock. Measure the job.
Do I need to rewrite everything?+
Not always. Not never. Treat ROCm as a port, not a checkbox. Budget time.
Keep comparing
Alternatives to AMD MI300X
Same job, different envelope. Compare VRAM, bandwidth and the hour before you lock a badge.
NVIDIA H200
141 GB HBM3e · 9% less bandwidth than the MI300X
From$3.00/ GPU / hr
CompareNVIDIA B200
180 GB HBM3e · 51% more bandwidth than the MI300X
From$3.75/ GPU / hr
CompareNVIDIA H100
80 GB HBM3 · 37% less bandwidth than the MI300X
From$1.73/ GPU / hr
CompareNVIDIA A100
80 GB HBM2e · 62% less bandwidth than the MI300X
From$1.19/ GPU / hr
CompareSources
Where these numbers come from
TFLOPS and TDP are from the vendor datasheet. Hourly prices are from the GPUBeacon catalog sample. Open the originals, then confirm the live SKU with the provider before you rent.