Provider record · updated 2026-09-10
Microsoft Azure — GPU pricing & deliverability
9 tracked offerings · grade mix A:4 · B:5 · C:0 · D:0 · U:0. Azure publishes per-GPU InfiniBand bandwidth on Microsoft Learn: ND H100 v5 and ND H200 v5 each give every GPU a dedicated 400 Gbps NVIDIA Quantum-2 ConnectX-7 InfiniBand link (3.2 Tbps per 8-GPU VM) with GPUDirect RDMA, scaling 'up to thousands of GPUs'. ND GB200 v6 (4 GPUs/VM) has 4x 400 Gb/s Quantum-2 CX7 IB per VM, NVL72 NVLink domains of 72 GPUs (18 VMs/rack, 28.8 Tb/s rack scale-out), and a stated InfiniBand fabric reach of up to 100,000 GPUs. Prices from the official Azure Retail Prices API (Consumption meters, Linux).
Tracked offerings
| GPU | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| H200-SXM | Standard_ND96isr_H200_v5 (3yr) | $5.28 | reserved · min: 3yr | B | publ | InfiniBand NDR: dedicated 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand per GPU, 3.2 Tb/s per VM, GPUDirect RDMA (Microsoft Learn) | ↗API |
| H100-SXM | Standard_ND96isr_H100_v5 (3yr) | $5.40 | reserved · min: 3yr | B | publ | InfiniBand NDR: 'dedicated, topology-agnostic 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand connection' per GPU, 3.2 Tbps per VM, GPUDirect RDMA (Microsoft Learn) | ↗API |
| H200-SXM | Standard_ND96isr_H200_v5 (1yr) | $5.82 | reserved · min: 1yr | B | publ | InfiniBand NDR: dedicated 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand per GPU, 3.2 Tb/s per VM, GPUDirect RDMA (Microsoft Learn) | ↗API |
| H100-SXM | Standard_ND96isr_H100_v5 (1yr) | $7.87 | reserved · min: 1yr | B | publ | InfiniBand NDR: 'dedicated, topology-agnostic 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand connection' per GPU, 3.2 Tbps per VM, GPUDirect RDMA (Microsoft Learn) | ↗API |
| H200-SXM | Standard_ND96isr_H200_v5 eastus2 | $10.60 | on-demand | B | publ | InfiniBand NDR: dedicated 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand per GPU, 3.2 Tb/s per VM, GPUDirect RDMA (Microsoft Learn) | ↗API |
| GB200-NVL72 | Standard_ND128isr_NDR_GB200_v6 (3yr) | $11.90 | reserved · min: 3yr | A | publ | InfiniBand NDR: 4x 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connections per VM (one 400 Gb/s NIC per GPU); 28.8 Tb/s scale-out networking per 72-GPU rack (Microsoft Learn) | ↗API |
| H100-SXM | Standard_ND96isr_H100_v5 eastus | $12.29 | on-demand | A | publ | InfiniBand NDR: 'dedicated, topology-agnostic 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand connection' per GPU, 3.2 Tbps per VM, GPUDirect RDMA (Microsoft Learn) | ↗API |
| GB200-NVL72 | Standard_ND128isr_NDR_GB200_v6 (1yr) | $17.31 | reserved · min: 1yr | A | publ | InfiniBand NDR: 4x 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connections per VM (one 400 Gb/s NIC per GPU); 28.8 Tb/s scale-out networking per 72-GPU rack (Microsoft Learn) | ↗API |
| GB200-NVL72 | Standard_ND128isr_NDR_GB200_v6 eastus, eastus2 | $27.04 | on-demand | A | publ | InfiniBand NDR: 4x 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connections per VM (one 400 Gb/s NIC per GPU); 28.8 Tb/s scale-out networking per 72-GPU rack (Microsoft Learn) | ↗API |
Grades measure what sellers publish — deliverability of the written claim, not measured performance. No fleet holds a measured (Verified) badge yet; that tier arrives with the verification harness. Source classes: API = provider's live pricing API · vendor = provider's published page · mkt API = marketplace API · 3rd-party = aggregator (provider publishes no price on its own pages) · quote = quote-gated or "starting at", not a walk-up price. Quote-class prices are shown in tables but excluded from headline ranges, spreads, and fair-price bands.
Terms & trust (as published, 2026-09-14)
| Uptime SLA | 99.99% (2+ VMs across 2+ Availability Zones); 99.95% (2+ VMs in Availability Set / Dedicated Host Group); single instance: 99.9% (Premium SSD/Ultra Disk), 99.5% (Standard SSD), 95% (Standard HDD) — measured on: 'Virtual Machine Connectivity' — connectivity to at least one instance for multi-instance deployments; single-instance tier depends on disk type for all OS and data disks. No GPU-specific SLA; standard VM SLA applies to ND/NC-series by deployment configuration. |
| SLA remedy | credits — multi-instance: <99.99%/<99.95% = 10%, <99% = 25%, <95% = 100%; single-instance tiers step down similarly by disk type. 'Service Credits are Customer's sole and exclusive remedy.' Claim-based. |
| Capacity guarantee | Explicitly published but NOT for flagship AI GPUs: On-demand Capacity Reservations 'Provides capacity guarantee in the specified location' with an SLA, but docs state 'Support for other VM series isn't currently available: GPU-accelerated compute: ND-series, NCads H100 v5-series.' Supported GPU series limited to NCv3/NCasT4_v3/NC_A100_v4/NV-series and similar. Also published: 'A quota is a credit limit on Azure resources, not a capacity guarantee.' |
| Egress fees | First 100 GB/month free; internet egress from North America/Europe: next 10 TB $0.08/GB, next 40 TB $0.065, next 100 TB $0.06, next 350 TB $0.04 (higher rates from Asia/Australia/MEA and South America); >500 TB/month contact Microsoft |
| Billing granularity | per-minute — 'We charge for the number of full minutes your virtual machine is running'; no sub-minute billing published for VMs |
| Self-serve signup | partial — pay-as-you-go credit-card signup, but GPU families (NCasT4_v3, NC_A100_v4, NDv2, etc.) 'start with a default of zero cores'; quota increase via portal, may escalate to a support request; Free Trial/Students subscriptions cannot get GPU quota |
| Security certifications | SOC 1/2/3, ISO 27001/27017/27018/27701, HIPAA, FedRAMP — listed on Azure compliance documentation hub (~100 offerings) |
Notes: Canonical global VM SLA text now lives in the downloadable 'SLA for Online Services' licensing document; exact figures verified via Microsoft's mirrored Azure China SLA page plus Microsoft Learn. Key gap: ND-series/H100-class excluded from the published capacity guarantee.
Sources: ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ · Fields marked "not published" are exactly that — the seller does not state them anywhere we could find. Absence is a finding.
Reliability record (public status feed, last 90 days)
0 incidents on Microsoft Azure's own status page in the trailing 90 days, collected 2026-09-15. Self-reported by the provider — a floor on true incident count, not a ceiling; it still beats the war-story thread you were going to read instead. High counts often reflect granular per-rack reporting, not worse infrastructure: read the incidents, not just the number.
No incidents disclosed in the window.
What Microsoft Azure does not publish
- Numeric maximum cluster size for ND H100 v5 / ND H200 v5 (Microsoft states only 'thousands of GPUs')
- ND96isr_H200_v5 price in eastus specifically (API listed eastus2/centralus and others, not eastus, at collection time)
Unpublished fields are recorded as unknown and grade conservatively. Sellers improve their grades by disclosing — that is the point.
Weekly context on every provider
The Delivered Compute Report tracks how these prices and grades move.
Subscribe free