RaisePath helps you organize, track, and share your work in 1 unified workspace. Built for teams and solo users who want clarity and speed.

What is the difference between H100 and H200?

The H200 features 141GB HBM3e memory vs 80GB on the H100, and higher memory bandwidth. For LLM inference, the H200's larger memory pool allows larger models on a single GPU. The H200 commands a 30-50% price premium but delivers materially better performance for memory-bound workloads.

When should an organisation use dedicated GPU providers instead of hyperscale cloud?

Dedicated providers typically offer H100 and H200 capacity at 50-75% less than AWS, Azure or GCP on-demand rates. Organisations with stable, high-volume GPU demand and engineering capability to manage their own infrastructure stack will generally find dedicated providers compelling.

How much can reserved GPU instances save?

One-year reserved instances typically save 30-40% versus on-demand. Three-year commitments can save 50-60%. The risk is over-commitment: reserving capacity you don't fully utilise negates the saving.

What is GPU spot pricing and when is it suitable?

Spot instances offer unused GPU capacity at 50-70% below on-demand pricing with the risk of interruption. Suitable for fault-tolerant training jobs with checkpointing enabled, but unsuitable for real-time inference or latency-sensitive workloads.

GPU Cost Intelligence

GPU Cost Comparison 2026: H100, H200 and Beyond

Compare GPU compute costs across cloud and dedicated providers. On-demand, reserved and spot pricing, GPU architecture benchmarks, and the cost traps most organisations miss.

50-75%

Typical saving vs hyperscale on-demand

30-60%

Reserved vs on-demand saving

GPU types benchmarked

Providers compared

GPU provider pricing comparison

Indicative hourly rates for 8× H100 SXM and equivalent clusters. On-demand, 1-year reserved and spot pricing where available. Prices vary by region and configuration.

PROVIDERCONFIGURATIONON-DEMAND1-YR RESERVEDSPOTREGIONGDPR

AWS (EC2 p5)H100 SXM (8×)$98.32/hr$~62/hr$~35/hrUS-East-1No

AWS (EC2 p4d)A100 SXM (8×)$32.77/hr$~21/hr$~12/hrUS-East-1No

Azure (ND H100)H100 SXM (8×)$~98/hr$~63/hrVariableEast US✓

Google Cloud (a3-highgpu)H100 SXM (8×)$~98/hr$~60/hr$~30/hrUS-CentralNo

CoreWeaveH100 SXM (8×)$~26/hr$~18/hrN/AUS/EU✓

Lambda LabsH100 SXM (8×)$~24/hr$~17/hrN/AUSNo

Vast.aiH100 SXM (1×)$~2.50/hrN/A$~1.80/hrGlobalNo

Nebius AIH100 SXM (8×)$~22/hr$~16/hrN/AEU-North✓

Pricing indicative, per cluster. May 2026. Spot pricing is market-variable. Always verify with provider.

ON-DEMAND

Maximum flexibility. Pay by the hour with no commitment. Highest unit cost - typically 2-3× reserved pricing. Suitable for variable, unpredictable workloads or new capacity exploration.

RESERVED INSTANCES

Commit to 1 or 3 years in exchange for 30-60% discount vs on-demand. Requires accurate demand forecasting. Unused reserved capacity is a sunk cost.

SPOT / PREEMPTIBLE

Discounted unused capacity - up to 70% below on-demand. Subject to interruption with short notice. Suitable for fault-tolerant training jobs with checkpointing. Not for inference.

Not sure which GPU route is right for your workloads?

A RaisePath compute review will map your workloads against available GPU options and identify where you're overpaying or over-specifying.

Book an AI Compute Audit View Provider Intelligence →

Hyperscale H100 on-demand (8×)$98.32/hr

Dedicated provider H100 (8×)$24.00/hr

Potential hourly saving$74.32 (76%)

Annual saving (50% utilisation)~$325,000

Evaluating a major GPU purchase or capacity agreement? Infrastructure Diligence tests the assumptions before capital moves.