Home › Guides › Compute & cloud credits › Guide › Article

Article · Compute & cloud credits · 5 min read

On-demand vs reserved GPUs for AI research

Paying the list on-demand rate is only one way to rent a GPU. Providers also sell interruptible capacity, which is cheaper if you can survive being preempted, and reserved or committed terms, which cut the unit price if you will actually use the hours. The labels sound similar across clouds; the terms do not.

How much free compute could you stack?Answer three questions in the credits calculator. Built from verified listings only.

Some links in this guide are affiliate links. If you sign up through them we may earn a small commission, at no extra cost to you. It never affects what we recommend.

The three ways most GPU clouds charge

Most of the specialist GPU clouds we track sell some mix of these three models. The names vary, but the trade-off is the same:

  • On-demand. You start a machine, you pay while it runs, and you can usually stop it when the job ends. The unit price is the highest of the three, and capacity is not guaranteed at busy times.
  • Interruptible, sometimes called spot. A lower rate, in exchange for the provider being able to reclaim the machine. This only works if you checkpoint and can restart.
  • Reserved or committed. You agree to a term, or a spend, in return for a lower unit price and, on some platforms, guaranteed capacity. You typically pay for the commitment whether or not you use every hour.

List prices move, and a credit programme can change the comparison entirely. If you have not already looked at credits, start with the cheapest GPU clouds for AI research and GPU and inference credits beyond the big three.

Runpod: on-demand now, savings plans withdrawn

Runpod's published Secure Cloud on-demand rates, from the pricing page updated 12 November 2026 and re-checked for this article on 18 November 2026, include H100 PCIe at $2.89/hr, H100 SXM at $4.39/hr and H100 NVL at $3.19/hr [source]. Pods are billed by the second for compute and storage, with no fees for data ingress or egress [source].

On-demand pods are dedicated and, Runpod says, cannot be displaced by other users. You must have at least one hour's worth of credits for the configuration you choose before you can deploy an on-demand instance [source].

Runpod's savings plans are deprecated and can no longer be purchased. If you already bought a plan, it stays active until its expiration date, applies to pods of the same GPU type, covers GPU compute only (storage is billed at standard rates), and is non-refundable [source]. For reserved or custom capacity, the pricing page and the pods pricing docs both direct customers to the sales team [source] [source].

Storage is a separate bill. Container disk is $0.10/GB/month while a pod is running and is not charged when it is stopped (it is also erased when the pod stops). Volume disk is $0.10/GB/month running and $0.20/GB/month when stopped. Network volumes are cheaper per gigabyte and persist until you delete them [source]. If the account balance hits $0, pods are stopped; without a network volume they are terminated and the data cannot be recovered [source].

Runpod Startup Program is the credit programme we currently list: Starter Tier: $1,000 in credits; Growth Tier: $25,000 bonus credits when you commit $50,000.

Vast.ai: on-demand, interruptible or reserved

Vast.ai does not publish a single list price. Hosts set rates across 40+ data centres, and the official pricing page offers three instance types, re-checked on 18 November 2026 [source]:

  • On-demand: described as best for production, with per-second billing, no interruptions, and the ability to spin up or down at any time.
  • Interruptible: described as 50%+ cheaper and best for batch training. Instances are preemptible and may be reclaimed, so the page recommends fault-tolerant workloads that can checkpoint and resume.
  • Reserved: described as up to 50% off for a long-term commitment, with 1, 3 or 6 month terms, guaranteed capacity and volume discounts, arranged through sales.

Because the price is a marketplace rate, a reserved quote can still be higher or lower than another provider's on-demand list price on the day you check. Compare the same GPU variant, and read the uptime terms, before you treat "50% off" as a saving against a different cloud.

Vast.ai Startup Program (GPU credit matching) is the credit programme we currently list: Dollar-for-dollar credit match on verified GPU compute spend, up to an approved amount.

DigitalOcean reserved plans, and a caution about powered-off machines

DigitalOcean publishes both on-demand and reserved rates for GPU Droplets. On 18 November 2026, the official GPU Droplets pricing page listed NVIDIA HGX H100 at $4.41 per GPU per hour on demand, and a 12-month reserved H100 plan at $3.26 per GPU per hour, which the page says to arrange with sales [source]. That reserved rate is about 26% below the published on-demand price, but only if you actually use the reserved hours.

DigitalOcean also lists spot plans for some other GPUs. The page says spot GPU Droplets are interruptible and can be reclaimed at any time, with an aim of at least two hours' notice, but with less notice in some cases including emergencies [source].

A detail that catches people out: when you power off a GPU Droplet, you are still billed for it, because the disk, CPU, RAM and IP address stay reserved [source]. That is the opposite of a "stop the meter" habit from some other clouds, and it can erase the apparent saving of a reserved or on-demand plan if you leave machines sitting idle.

How to choose, in practice

  1. See whether a credit programme covers the work first. Credits expire, but they are usually the cheapest official hour you will be offered. Runpod Startup Program and Vast.ai Startup Program (GPU credit matching) are the specialist GPU programmes we list; the large public clouds are covered in Google, AWS and Azure credits compared and in how to spend those credits without wasting them.
  2. If the job can be checkpointed, compare interruptible or spot rates, including Vast.ai's interruptible offers, and budget time for restarts.
  3. If the job cannot be interrupted, compare on-demand list prices with reserved quotes for the same GPU. Reserved capacity only wins if you will use enough of the term to beat the on-demand bill, after storage and powered-off charges.
  4. For reserved terms on Runpod and for Vast.ai reserved capacity, you will usually be speaking to sales rather than buying a self-serve plan. Ask what happens if you need to exit early, and what is guaranteed.
  5. Re-check the provider's pricing page on the day you start. The figures in this article are the official published rates on 18 November 2026, not a quote.

This comparison is about published rental terms, not a recommendation of any provider. For how GPU spend sits next to grants and equity, see the AI startup funding stack.

Open now: compute & cloud credits closing soonest

See all 28 open compute & cloud credits →

More on compute & cloud credits

Frequently asked questions

Is reserved GPU capacity always cheaper than on-demand?

Only if you use enough of the committed hours to beat the on-demand bill, after storage and any charges that continue when a machine is stopped. DigitalOcean's published 12-month reserved H100 rate of $3.26/GPU/hr is lower than its $4.41/GPU/hr on-demand list price, but a powered-off Droplet is still billed. Runpod no longer sells savings plans; reserved capacity is through sales.

What happened to Runpod savings plans?

They are deprecated and can no longer be purchased. Existing plans run until they expire, apply to pods of the same GPU type, cover GPU compute only, and are non-refundable, according to Runpod's pods pricing documentation, checked on 18 November 2026.

When is an interruptible GPU a bad idea?

When the job cannot be checkpointed, when a mid-run preemption would waste a lot of paid time, or when you have a hard deadline you cannot slip. Vast.ai describes interruptible instances as 50%+ cheaper and preemptible; DigitalOcean says spot GPU Droplets can be reclaimed at any time.

Do startup credits work on reserved instances?

It depends on the programme and the cloud. AWS Activate, for example, applies credits to Reserved Instances and Savings Plans in the same way as other eligible services, but does not cover upfront fees. Always read the credit programme's own exclusions, which we walk through in how to spend Google, AWS and Azure credits without wasting them.

Should I rent GPUs before I apply for credits?

Usually no, if you think you will qualify. Credits expire, so you should not apply years before you have work to run, but paying list price first can waste hours the credit would have covered. Check the open programmes on our compute-credits list.

Sources

Programme details are quoted from the organisers' pages on the dates shown. advance.to isn't affiliated with any organiser.

  1. Runpod GPU cloud pricing (page updated 12 November 2026; checked 18 November 2026)
  2. Runpod docs: Pods pricing (checked 18 November 2026)
  3. Vast.ai live GPU pricing (checked 18 November 2026)
  4. DigitalOcean GPU Droplets pricing (checked 18 November 2026)
  5. Runpod: Runpod Startup Program (checked 18 Nov 2026)
  6. Vast.ai: Vast.ai Startup Program (GPU credit matching) (checked 18 Nov 2026)

Share this page:

Free guide + the Monday email

Get the free guide. Never miss a deadline.

Your First AI Funding Application shows you which AI funding fits you and how to write an application that gets read. Then, every Monday, one short email with the deadlines closing in the next 14 days.