cd /news/ai-infrastructure/when-does-reserved-gpu-capacity-beat… · home › topics › ai-infrastructure › article
[ARTICLE · art-172274] src=dev.to ↗ pub= topic=ai-infrastructure verified=true sentiment=· neutral

When does reserved GPU capacity beat on-demand pricing?

An analysis of AWS GPU pricing shows reserved capacity only beats on-demand once utilization exceeds the reserved rate divided by the on-demand rate—63.0% of all hours for an 8×H100 p5.48xlarge on a 1-year EC2 Instance Savings Plan, and 43.2% for the 3-year plan. At 60% utilization the machine is still $14,464.51 cheaper on demand for the year, while at 80% the 1-year plan saves $81,965.57, leading to the recommendation to reserve only the steady floor and run the rest on demand.

recommendation

Compute break-even utilization as reserved rate divided by on-demand rate before committing, and reserve only GPUs that stay busy above that line.

more in playbook: Controlling LLM cost
by read8 min views7 publishedNov 19, 2026

Reserved GPU capacity beats on-demand once your utilization over the term exceeds the reserved rate divided by the on-demand rate: 63.0% of all hours for an 8 x NVIDIA H100 machine on a 1-year Amazon EC2 Instance Savings Plan. Nodus lists the H100 SXM at $2.60/hr on its pricing page, a GPU rate below even AWS's 3-year plan per GPU-hour, and Nodus Reserved Compute (Beta) quotes dedicated capacity for a steady floor. Commit only the GPUs that stay busy above the line and run the rest on demand.

Divide the reserved rate by the on-demand rate you would otherwise pay. The answer is the fraction of the term's hours you must actually use for the commitment to cost less.

The reasoning is short. A commitment bills every hour of the term, used or not. On-demand bills only the hours you run. So reserved wins when reserved rate x all hours < on-demand rate x used hours, which rearranges to:

break-even utilization = reserved rate / on-demand rate
effective reserved rate = reserved rate / your utilization

The second line is the one people skip. A $4.33 commitment that you use half the time really costs $8.67 per hour of work. That is more than the $6.88 on-demand rate it was meant to beat.

We worked this on AWS because it publishes every commitment rate for the same machine. From the official AWS price list for us-east-1, a p5.48xlarge costs $55.04 per hour on demand, or $6.88 per GPU-hour.

The 1-year EC2 Instance Savings Plan with no upfront payment is $34.6752 per hour, or $4.33 per GPU-hour. The break-even is 34.6752 / 55.04 = 0.630. We re-checked it the other way: 55.04 x 0.63 = 34.6752.

Option (8 x H100, AWS us-east-1) Per instance-hour Per GPU-hour Break-even utilization
On-demand $55.04 $6.88 n/a
EC2 Capacity Block $47.757 $5.97 86.8%
Compute Savings Plan, 1 year, no upfront $43.344 $5.42 78.8%
EC2 Instance Savings Plan, 1 year, no upfront $34.6752 $4.33 63.0%
EC2 Instance Savings Plan, 3 years, no upfront $23.77728 $2.97 43.2%

Prices checked November 18, 2026 (this applies to every price in this post). AWS rates come from the AWS EC2 on-demand price list, the AWS Compute Savings Plans price list file for us-east-1 (which also carries EC2 Instance Savings Plan rates) and the EC2 Capacity Blocks pricing page. Per GPU-hour is the instance rate divided by 8, rounded to cents. Linux, no operating system fee.

The deeper the discount and the longer the term, the less of the year you need to fill.

A commitment is a flat line and on-demand is a slope, so the only question is where your real usage lands. A full year is 8,760 hours (AWS defines a Savings Plan year as 365 days). The 1-year EC2 Instance Savings Plan costs $34.6752 x 8,760 = $303,754.75 for the year. The 3-year plan costs $23.77728 x 8,760 = $208,288.97 per year, for three years.

Utilization over the year On-demand cost 1-year plan 3-year plan, per year
0% $0.00 $303,754.75 $208,288.97
20% $96,430.08 $303,754.75 $208,288.97
40% $192,860.16 $303,754.75 $208,288.97
60% $289,290.24 $303,754.75 $208,288.97
80% $385,720.32 $303,754.75 $208,288.97
100% $482,150.40 $303,754.75 $208,288.97

On-demand is $55.04 x 8,760 x utilization. At 60% the machine is still cheaper on demand, by $14,464.51. At 80% the 1-year plan saves $81,965.57.

On-demand crosses the 1-year line at 63.0% utilization and the 3-year line at 43.2%.

Utilization here means hours of the whole year, nights and weekends included. A team that runs a training box 10 hours a day on weekdays uses 50 of 168 weekly hours, which is 29.8%. That team should not sign either plan for that machine.

Reserve the floor. Split your fleet into tiers by how busy each GPU is across the term, commit only the tiers above break-even, and run the rest on demand or on spot.

Here is a worked example with the AWS per-GPU rates above. A team serves a model on 8 H100s around the clock. It adds 8 more H100s for weekday business hours, 10 hours a day, 5 days a week. The second tier runs 8 x 8,760 x 50 / 168 = 20,857.14 GPU-hours a year.

Strategy First 8 GPUs (100% busy) Second 8 GPUs (29.8% busy) Year total
Everything on demand $482,150.40 $143,497.14 $625,647.54
Reserve all 16 (1-year plan) $303,754.75 $303,754.75 $607,509.50
Reserve the floor, peak on demand $303,754.75 $143,497.14 $447,251.89
Reserve the floor, peak on Nodus at $2.60 $303,754.75 $54,228.57 $357,983.32

The floor tier saves $178,395.65 by committing. The peak tier would waste $160,257.61 if you committed it too, which is why reserving everything barely beats paying list price. The last row prices the peak hours at the H100 SXM rate listed on the Nodus pricing page: 20,857.14 x $2.60. GPU rates only; the AWS rate includes the instance's CPUs, memory and networking, so treat that row as a rate comparison, not a like-for-like quote.

Committing the always-busy floor saves money; committing the peak gives most of it back.

No. Some reservations cost more per GPU-hour than the same vendor's on-demand rate, because you are buying guaranteed, co-located capacity for fixed dates rather than a discount.

Lambda's pricing page lists 1-Click Clusters of 16 H100s for terms of 2 weeks to 1 year at $6.16 per GPU-hour, while an on-demand 8 x H100 instance is $4.59 per GPU-hour.

That ratio is 134.2%, so on rate alone the cluster never breaks even. You buy it for a cluster of 16 or more GPUs held for a fixed term, not for a lower rate. The AWS Capacity Blocks documentation describes the same trade: blocks are for GPUs "for days or weeks at a time", placed close together in EC2 UltraClusters. The Capacity Blocks pricing page adds that the reservation fee is charged up front when you schedule it.

Lambda also says on that page to contact them "for reserved capacity at our lowest prices". Runpod's pricing page and the Nebius prices page likewise route reserved and commitment pricing through sales. Their long-term rates are not published, so we have not priced them here.

They sit at opposite ends of the same trade. Spot is the cheapest per hour with no commitment, but the provider can take the machine back. Reserved costs a commitment and bills every hour, but the capacity is yours for the term.

AWS advertises Spot Instances at up to 90% off on-demand prices. Spot fits interruptible work that checkpoints, such as training and batch inference.

Reserved fits the always-busy floor: production inference, a long training run with a fixed start date. On-demand covers the space between. Our spot vs on-demand guide works through the rework math for spot, so we do not repeat it here.

Pick the one whose flexibility matches what you might change. AWS's Savings Plans guide says EC2 Instance Savings Plans commit you to one instance family in one region, while Compute Savings Plans apply regardless of family, size or region. For p5 that flexibility is priced in: the 1-year Compute plan breaks even at 78.8%, against 63.0% for the instance plan.

A 3-year plan drops the line to 43.2%, but it locks one GPU generation for three years. Sign it only if you are confident you will still want that hardware in year three. If you mostly want to stop overpaying AWS list rates, our cheaper AWS alternatives post compares on-demand H100 rates across platforms.

We cover both halves of the floor-and-peak split. For the peak, Nodus Jobs run your command on demand, with logs, outputs, spending limits and recovery, and the H100 SXM is listed at $2.60/hr on the Nodus pricing page. That is below the $2.97 per GPU-hour of AWS's 3-year EC2 Instance Savings Plan, so for GPU rate alone no AWS commitment in the table beats it at any utilization.

For the floor, Nodus Reserved Compute (Beta) supplies dedicated GPU capacity for your configuration and reservation window. You describe the GPU model, quantity, duration, location and start timing. Nodus reviews each request and sends a proposal covering availability, price, access and delivery. Submitting a request does not reserve hardware or authorize payment.

The Reserved Compute page also shows a planning index built only from providers' published reservation rates for the exact term you select, with the range and provider count.

It is a Beta and a quote, not a self-serve checkout. If you need a price this minute, use the published commitments above.

Take at least a few weeks of billing or scheduler data for the exact GPU type, and divide GPU-hours used by GPU-hours in the period. Count nights and weekends. If the number moves a lot week to week, commit only to the lowest steady level.

Yes, slightly. In the same AWS file, the 1-year EC2 Instance Savings Plan for p5.48xlarge is $34.6752 per hour with no upfront. It drops to $33.024 partial upfront and $32.36352 all upfront, so more cash up front lowers the line a little.

Yes. The Reserved Compute guide says you can request a single GPU, and the final offer confirms how GPUs are arranged across servers. The console request form accepts durations up to 365 days.

Growth is the easy case: the commitment keeps covering the floor and new load runs on demand until it proves steady. Shrinking is the costly case, because the commitment bills every hour whether or not you use it. Our cost per finished job post shows how idle hours inflate the real bill.

Originally published on the Nodus blog.

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @aws 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/when-does-reserved-g…] indexed:0 read:8min 2026-11-19 · —