· via dev.to (home feed)
Reserved H100 capacity beats AWS on-demand pricing above 63% utilization
A dev.to pricing analysis finds a 1-year AWS EC2 Instance Savings Plan on an 8x H100 p5 instance only pays off above 63% utilization, and shows why committing bursty peak capacity erases the savings.

The break-even rule
A pricing analysis published on dev.to reduces the reserved-versus-on-demand GPU decision to a single division: reserved rate divided by on-demand rate. The result is the share of a term's hours a machine must actually run before the commitment costs less than paying by the hour. For an 8x NVIDIA H100 instance under a 1-year Amazon EC2 Instance Savings Plan, that break-even lands at 63.0% — the GPUs must be busy for roughly two-thirds of all hours in the year, nights and weekends included, or the reservation loses money.
The mechanics are straightforward: a commitment bills every hour of the term whether or not the machine does work, while on-demand bills only the hours used. Reserved pricing wins once the committed rate times all term hours falls below the on-demand rate times the hours actually consumed.
The post also stresses a figure teams tend to skip: the effective rate. Divide the reserved rate by real utilization to see what each productive hour costs. A $4.33-per-GPU-hour commitment used only half the time works out to $8.67 per hour of actual work — more than the $6.88 on-demand rate it was meant to undercut.
The AWS numbers
Working from AWS's official us-east-1 price lists, checked November 18, 2026 (Linux, no OS fee), the analysis prices a p5.48xlarge, which carries eight H100s:
| Option | Per instance-hour | Per GPU-hour | Break-even utilization |
|---|---|---|---|
| On-demand | $55.04 | $6.88 | — |
| EC2 Capacity Block | $47.757 | $5.97 | 86.8% |
| Compute Savings Plan, 1 year | $43.344 | $5.42 | 78.8% |
| EC2 Instance Savings Plan, 1 year | $34.6752 | $4.33 | 63.0% |
| EC2 Instance Savings Plan, 3 years | $23.77728 | $2.97 | 43.2% |
The pattern: the deeper the discount and the longer the term, the less of the year you need to fill. Over a full 8,760-hour year, the 1-year instance plan is a flat $303,754.75 and the 3-year plan $208,288.97 per year for three years, against $482,150.40 for around-the-clock on-demand. At 60% utilization on-demand is still $14,464.51 cheaper; at 80% the 1-year plan is ahead by $81,965.57.
Reserve the floor, not the peak
The recommended strategy is to tier a fleet by how busy each machine stays across the term and commit only the tiers above their break-even. The worked example: eight H100s serving a model around the clock, plus eight more that run 10 hours a day on weekdays — 50 of 168 weekly hours, or 29.8% utilization, well below every plan's threshold.
For that fleet, running everything on demand costs $625,647.54 a year; reserving all 16 GPUs costs $607,509.50; reserving only the always-on floor and leaving the peak on demand costs $447,251.89. Committing the steady tier saves $178,395.65, but committing the bursty tier hands back $160,257.61 — which is why blanket reservations barely beat list price. The post also prices the peak tier at the $2.60 per H100 SXM GPU-hour listed on Nodus's pricing page ($357,983.32 total), while cautioning that this is a rate comparison, not like-for-like, since the AWS rate bundles CPUs, memory and networking. Worth noting the post promotes Nodus's own compute products.
Reserved is not always a discount
Some reservations cost more per hour than the same vendor's on-demand rate because they sell guaranteed capacity rather than savings. Lambda's published pricing lists 1-Click Clusters of 16 H100s at $6.16 per GPU-hour while an on-demand 8x H100 instance is $4.59 — a ratio of 134.2% that never breaks even on rate alone. AWS Capacity Blocks, with an 86.8% break-even and an up-front reservation fee, sit in the same category: short placements of tightly co-located GPUs, not a price cut. Runpod and Nebius route commitment pricing through sales, so their long-term rates went unpriced.
Spot sits at the other end
According to the post, spot and reserved bookend one trade. AWS advertises spot at up to 90% off on-demand, suited to interruptible work that checkpoints, such as training and batch inference; reserved suits a steady floor like production inference or a fixed-start training run. Plan choice also hinges on flexibility: per AWS's Savings Plans guide, instance plans lock a single instance family in one region, while Compute Savings Plans flex across family, size and region — flexibility that costs about 16 points of break-even here (78.8% versus 63.0%). The 3-year plan's 43.2% line buys a lower threshold at the cost of locking one GPU generation for three years.
Why it matters
GPU capacity is now among the largest line items in many cloud budgets, and the instinct that "reserved is cheaper" fails quietly whenever utilization is modest. This analysis turns that instinct into one division any team can run against its own telemetry: measure utilization across all 8,760 hours of a year — idle nights and weekends included — and commit only what clears the line. The stakes are concrete: on a single 8-GPU machine, reserving the steady floor rather than reserving everything is worth roughly $160,000 a year, and a 3-year commitment signed against a workload that fades in year two keeps billing regardless.
- #aws
- #gpu
- #cloud-costs
- #h100
- #pricing