Azure MI300X Pricing 2026: ND MI300X v5 Cost vs Spheron

Azure's Standard_ND96isr_MI300X_v5 lists at $48.00-$96.00/hr on-demand, or $6.00-$12.00/hr per GPU, per CloudPrice's Azure rate tracker. That's the instance behind most "azure mi300x pricing" searches, and it has a pricing structure that runs backwards from what we found on Azure's NVIDIA side. We covered Azure's ND H200 v5 a few weeks ago and flagged a gap: no published reserved tier and no distinct spot tier either. MI300X has the same missing reserved tier, but it does have a working spot rate, the inverse gap. Spheron started renting the MI300X on 10 Sep 2026, so a same-silicon comparison is now possible: this post puts Azure's ND MI300X v5 next to Spheron MI300X on rate, minimum purchase and provisioning, and keeps Spheron's NVIDIA H100 and H200 as a secondary angle on cost per GPU-hour and cost per token. For the hardware tradeoff behind that NVIDIA angle, memory capacity versus realized throughput, see our AMD MI300X vs NVIDIA H100 guide.

TL;DR: Azure ND MI300X v5 Cost vs Spheron MI300X

Azure's ND96isr MI300X v5 costs $6.00-$12.00 per GPU-hour on-demand, sold only as 8-GPU VMs (15 Jul 2026). Spheron rents the same AMD MI300X on-demand at $3.59/hr per GPU as of 13 Oct 2026, in 1x or 2x instances. Compare live GPU rates.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 13 Oct 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

What Is the Azure ND MI300X v5 (Standard_ND96isr_MI300X_v5)?

The ND MI300X v5 series is Azure's AMD-accelerator flagship, built for large-scale deep learning training and tightly coupled generative AI and HPC workloads. There's a single VM size in the series, Standard_ND96isr_MI300X_v5, and it reached general availability on May 21, 2024, with Hugging Face announced as the launch customer. "The AMD Instinct MI300X and ROCm software stack is powering the Azure OpenAI Chat GPT 3.5 and 4 services, which are some of the world's most demanding AI workloads," said Victor Peng, president of AMD, at the time. Hugging Face's Julien Simon, chief evangelist officer, added that the partnership would let Hugging Face users "run hundreds of thousands of AI models available on the Hugging Face Hub on Azure with AMD Instinct GPUs without code changes." Both quotes are from the May 2024 AMD launch announcement, not a 2026 statement, so read the "powering ChatGPT" line as a two-year-old snapshot of that partnership, not a claim about today's OpenAI infrastructure mix.

Specs: 8x MI300X 192GB, Infinity Fabric, and InfiniBand

Per Microsoft's own ND MI300X v5 series documentation, the instance has:

• 8x AMD Instinct MI300X GPUs with 192GB HBM3 each (1,535GB total accelerator memory)
• 96 vCPUs (Intel Xeon Sapphire Rapids)
• 1,850 GiB RAM
• 4th-gen AMD Infinity Fabric links between GPUs at 128 GB/s per GPU, 896 GB/s aggregate
• A dedicated, topology-agnostic 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connection per GPU (3.2 Tb/s aggregate per VM), with GPUDirect RDMA support for scaling to thousands of GPUs across VMs in the same scale set
• 1x 1,000 GiB local temp disk plus up to 8 NVMe disks at 28 TiB each
• Max network bandwidth of 80,000 Mbps across 8 NICs

Notice the interconnect split: GPU-to-GPU traffic inside the VM runs over AMD's own Infinity Fabric, while node-to-node scaling runs over NVIDIA's InfiniBand hardware. That's not a typo. Azure builds its ND-series network fabric on Quantum-2 InfiniBand regardless of which accelerator vendor sits inside the box, so an AMD compute VM still depends on NVIDIA networking silicon to scale past a single node. For the architecture comparison against NVIDIA's own H200, including memory bandwidth and MLPerf results, see AMD MI300X vs NVIDIA H200.

ND MI300X v5 On-Demand Pricing Breakdown

There's no sticker price for this instance on Azure's standard retail pricing pages. The numbers below come from CloudPrice, which scrapes Azure's rate card region by region, the same source pattern we used for the H100 and H200 SKUs.

Per-GPU Cost Across Regions ($6-$12/hr)

CloudPrice's regional breakdown for Standard_ND96isr_MI300X_v5 on-demand Linux pricing shows:

That's roughly a 2x spread between the cheapest and most expensive published region, on identical hardware, the same pattern we've seen across every ND-series SKU on this blog. Thunder Compute's July 1, 2026 pricing tracker independently confirms the $6.00/GPU-hr floor for Azure's ND MI300X v5, matching CloudPrice's cheapest-region figure exactly. CloudPrice's own regional table is otherwise paywalled behind an "activate subscription to see all regions" gate, so the cheapest and priciest bands above are the two data points we can verify directly; expect intermediate regions to land somewhere between $6.00 and $12.00 per GPU.

Pricing fluctuates based on GPU availability. The prices above are based on 15 Jul 2026 and may have changed. Check current GPU pricing for live rates.

Azure MI300X Spot Pricing: The Discount Path That Actually Exists

Here's the part that flips the script from our H200 coverage: Azure does publish a low-priority (spot) rate for this SKU, and it's a real discount, not a rounding error. Per SpareCores' server pricing database, the ND MI300X v5 spot tier runs $8.8704/hr in the cheapest regions (East US 2, West US 3) up to $17.7408/hr in Brazil South, working out to roughly $1.11-$2.22 per GPU. That's an 81-82% discount off the matching on-demand rate in the same region, consistent across both the cheap and expensive ends of the range.

Spot capacity on Azure is reclaimable at any point with no notice, standard low-priority VM behavior, so this rate only makes sense for checkpointed, interruption-tolerant training or batch inference, not anything that needs to run to completion on a deadline.

Why MI300X Has Spot Pricing When Azure's H200 SKU Doesn't

We don't have Microsoft's internal reasoning here, so treat this as an informed read of the public data rather than a confirmed explanation. MI300X reached general availability over two years ago (May 2024), giving Azure far more operating history and demand data to build a spot pool from than the newer ND H200 v5 SKU has accumulated. AMD accelerator demand on Azure also likely runs thinner than NVIDIA demand overall, since most large training jobs default to CUDA-first tooling, which would leave more idle MI300X capacity available to resell at a discount during off-peak windows. Whatever the mechanism, the practical result is that this is currently the only ND-series pricing gap running in AMD's favor on Azure: a real discount lever exists for MI300X that doesn't exist yet for H200.

The Reserved Instance Gap (and Why There's No Commitment Discount Yet)

Outside of spot, the MI300X SKU falls back into the same pattern we documented for Azure's ND H200 v5: neither CloudPrice nor SpareCores list a 1-year or 3-year reserved rate for Standard_ND96isr_MI300X_v5. Compare that to Azure's ND H100 v5 series, which has a full published discount curve, roughly 35% off on-demand with a 1-year commitment and 55% off with a 3-year commitment. Standard Azure reserved-instance discounts on ND-series GPU VMs generally land in that same 35%-ish range at the 1-year mark, but no such published number exists yet for MI300X.

That doesn't mean Azure will never offer it. Newer or lower-volume SKUs typically lag their reserved-tier rollout by a few quarters after general availability, and MI300X's public reserved gap has now persisted well past that window. Until it changes, every on-demand hour on this instance is billed at the full $6.00-$12.00/GPU-hr rate shown above, with spot as the only discount lever, short of negotiating a custom Enterprise Agreement rate directly with Microsoft that isn't publicly quoted anywhere.

Hidden Costs on Top of the Hourly Rate

The instance rate is the number every pricing tracker quotes, and it's not the whole bill. The extra line items are the same general Azure platform charges we've documented on the H100 and H200 SKUs, since they apply regardless of which accelerator sits inside the VM: outbound data transfer at roughly $0.087/GB for the first 5 TB/month out of East US, Premium SSD managed disks at $135.17/month for a 1024 GiB P30 disk (about $0.132/GB/month, per Azure's retail pricing API; the local NVMe on this SKU doesn't persist across stops), Azure Files for datasets shared across multiple nodes, Azure Monitor and Log Analytics for anything past default metrics, and a support tier if you need a response-time SLA.

Here's what that looks like for a single Standard_ND96isr_MI300X_v5 node run 40 hours a week (172 hours/month) at the cheapest published on-demand rate, with 10 TB of monthly egress and 2 TB of Premium SSD for checkpoints:

The extras add about $1,341/month, or roughly 16% on top of pure compute, a bigger proportional bite than we found on the pricier H200 SKU, where the same extras added closer to 7.5%. That's not because MI300X's ancillary costs are higher, they're identical Azure platform charges, it's because MI300X's lower compute base means fixed costs like egress and storage make up a larger share of the total bill. Cheaper compute doesn't automatically mean a cheaper stack once storage and networking enter the picture.

Pricing fluctuates based on GPU availability. The prices above are based on 15 Jul 2026 and may have changed. Check current GPU pricing for live rates.

Azure MI300X vs Spheron: Cost Per GPU-Hour and Per-Token Reality Check

Azure ND MI300X v5 vs Spheron MI300X: Same Silicon, Different Shape

Spheron added the AMD MI300X to its marketplace on 10 Sep 2026, and its public offers API lists it in 1x and 2x configurations. Larger MI300X configurations, including 8-GPU nodes, are available from Spheron as reserved capacity you request rather than launch. Both clouds run the same 192GB HBM3 MI300X inside a VM. What differs is how you're allowed to buy it:

Rate aside, the bigger difference is what you have to buy. Azure's cheapest way into an MI300X is the whole 8-GPU VM, $48.00/hr in East US 2 before egress and storage, and only after the quota request clears. On Spheron you can launch one MI300X, billed per minute after a 20-minute minimum (60 minutes on a 2x instance), find out whether your ROCm stack runs, and destroy it when the job is done. Plan for that last step: Spheron's MI300X instances can be restarted but not stopped, so they bill until destroyed, and copy checkpoints off before you do.

Azure keeps two real advantages. Its InfiniBand fabric scales a training job across many VMs and you buy it self-serve, where anything past a 2x MI300X on Spheron is reserved capacity you request rather than launch. And its low-priority tier takes 81-82% off Azure's own on-demand rate, while Spheron listed no MI300X spot tier as of 17 Sep 2026. If your AMD workload checkpoints well and tolerates reclaims, or needs more than 384GB of GPU memory in one job, Azure is still the better fit.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 13 Oct 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

When an NVIDIA Card on Spheron Is the Better Buy

Same silicon isn't the only question. If your stack doesn't depend on ROCm, an NVIDIA GPU can win on cost per token. These are Spheron's rates for its H200 SXM5 instances and other NVIDIA GPUs as pulled from its offers API on 15 Jul 2026, split strictly by instanceType before taking a minimum (dedicated on-demand offers can't be confused with spot offers, since spot listings carry a matching price field too). They're frozen here as the inputs to the cost-per-token math below; current rates are on the pricing page:

Line those 15 Jul 2026 rates up against Azure's MI300X range and the comparison isn't a clean sweep either way. On-demand, Azure's cheapest MI300X rate ($6.00/GPU-hr) is still 2.4x Spheron's H100 rate and 1.3x Spheron's H200 rate, and Azure's priciest region ($12.00/GPU-hr) is 4.7x H100 and 2.6x H200. But flip to spot, and the gap partially inverts: Azure's cheapest MI300X spot rate ($1.11/GPU-hr) actually undercuts Spheron's own H100 spot rate ($1.44/GPU-hr), because this SKU's spot discount is unusually steep (81-82% off) rather than because AMD hardware is inherently cheaper.

Applying the same ROCm-versus-CUDA throughput estimate we used in the MI300X pricing roundup (92% of an assumed 19,500 tokens/sec H100 baseline, so roughly 17,900 tokens/sec for MI300X) to those 15 Jul 2026 rates:

Formula: ($/hr / 3,600 sec) x (1,000,000 / tokens_per_sec)

There's no Spheron MI300X row here on purpose. Its rate is live, and a per-token figure frozen at one day's rate would go stale. Because it's the same chip as Azure's, the same 17,900 tokens/sec estimate applies, so run the current Spheron MI300X rate from the comparison table above through the formula.

Treat the tokens/sec figures as a directional estimate carried over from prior coverage, not a benchmark run on this exact configuration, and rerun the math against your own workload before committing budget. The honest read on those July rates: on-demand, Spheron's H100 beat Azure's MI300X on cost per token by a wide margin. Spot-to-spot, Azure's own discounted MI300X capacity was the cheapest path in this table, if your workload tolerates reclaimable capacity and your stack already runs on ROCm. That's a real, specific case where Azure wins outright, and it's worth saying plainly rather than smoothing it over.

The other side of the ledger is memory. A single MI300X's 192GB can host models that need two H100s (80GB each) to fit in FP16, which cuts interconnect complexity and multi-GPU serving overhead for large models. Whether that consolidation is worth more than Spheron's per-token edge depends on your model size and whether you're compute-bound or memory-bound, not on price alone. For the fuller memory-versus-throughput trade-off, see AMD MI300X vs NVIDIA H200, and for the software side of running ROCm in production, our ROCm pretraining and deployment guide and the ROCm vs CUDA comparison cover framework compatibility in depth. For the wider market view beyond Azure and Spheron, the GPU cloud pricing comparison tracks H100, H200, and B200 rates across every major provider side by side. If you're weighing that same memory-versus-throughput tradeoff one generation up, our B300 vs B200 inference cost-per-token guide runs the equivalent analysis for the Blackwell Ultra chips that are starting to replace H100 and H200 in high-end deployments.

Pricing fluctuates based on GPU availability. The prices above are based on 15 Jul 2026 and may have changed. Check current GPU pricing for live rates.

When It's Worth Moving AMD Inference Workloads Off Azure

Stay on Azure's MI300X if your team is already deep in ROCm on Azure specifically, you're drawing down Enterprise Agreement credits that don't transfer elsewhere, your workload is genuinely spot-tolerant and can exploit that 81-82% low-priority discount, or FedRAMP/HIPAA compliance is a hard requirement tied to Azure's certifications. None of that changes just because the reserved tier is missing.

Rent MI300X on Spheron if you want the same AMD silicon without buying a whole 8-GPU VM or waiting on a quota request. One or two GPUs, self-serve, billed per minute after a 20-minute minimum (60 minutes on 2x), at the live rate in the comparison table above. It suits ROCm validation runs, single-GPU serving of models that need 192GB, and 2x jobs up to 384GB. It doesn't suit self-serve multi-node training, since those 1x and 2x instances are single VMs and anything larger is a reserved-capacity request rather than a launch, or spot-tolerant batch work, where Azure's low-priority tier had no Spheron MI300X counterpart as of 17 Sep 2026.

Move to an NVIDIA path on Spheron if you need reliable on-demand capacity rather than reclaimable spot, if quota approval timelines (Azure's ND-series process typically runs 1-4 weeks, the same pattern documented for H100 and H200) don't fit your schedule, or if CUDA-first tooling (TensorRT-LLM, FlashAttention 3) matters more to your serving stack than MI300X's memory headroom. A September 2024 Microsoft Q&A thread also shows a user unable to provision this SKU through the Azure Portal at all, only the H100 equivalent appeared; that's a signal nearly two years old and shouldn't be read as current 2026 capacity, but it's a reminder to confirm regional availability before you commit budget to any specific SKU, AMD or NVIDIA. Spheron's overview docs cover how the platform aggregates capacity across 5+ providers instead of a single region's inventory, which is the structural reason it provisions without a quota queue in the first place. If you're weighing a broader move off Azure entirely, the step-by-step migration guide covers auditing current spend and standing up a first workload elsewhere without a rewrite, and the hyperscaler hidden-cost breakdown walks through the egress and quota math in more general terms.

Azure's ND MI300X v5 is a genuinely unusual SKU on its own pricing terms: a real spot discount that doesn't exist yet on its H200 sibling, paired with a reserved-instance gap that mirrors it exactly. Whether that trade works for you depends on how spot-tolerant your AMD workload actually is, not on a single headline number.

If you want Azure's MI300X silicon without the 8-GPU VM or the quota queue, Spheron rents it on-demand one or two GPUs at a time, next to H100 and H200 on the same account.

Spheron MI300X instances →