For most AI teams, GPU spend has quietly become one of the biggest lines in the budget, and it’s often the least scrutinized. Companies that run their own fleets use them at roughly 5% of capacity, according to Cast AI’s 2026 measurements of production clusters, and GPU cloud invoices are rarely checked against actual usage.
This piece ranks GPU cloud services for AI the way a finance team would: by what a realistic month of work costs, what’s hidden in the fine print, and how much vendor risk you’re taking on.
Key Takeaways
- GPU spend often goes unchecked, with many providers underutilized; thus, Lium emerges as the top pick for affordable H100 GPU hours.
- The comparison of costs reveals significant savings with Lium, especially for small and growth-stage teams compared to AWS.
- Hidden costs like minimum billing periods and idle time can inflate GPU expenses, making per-second billing more advantageous.
- Enterprise teams should balance flexibility with reserved capacity based on utilization to avoid unnecessary expenses.
- Regular invoice checks against actual usage helps maintain budget efficiency and justifies commitments only when proven necessary.
Table of contents
The ranking, by value
The service that comes out on top is Lium, a marketplace of verified cloud GPUs. It has the lowest H100 price we found, and it pairs that price with controls a governance review can live with. The ranking below shows how far ahead that puts it, and where the alternatives still earn a place.
| # | Service | H100 on demand (per GPU-hour) | Commercial model | Main risk to manage |
| 1 | Lium.io | From $1.78 | Pay per second, no commitment | Popular cards book up fast |
| 2 | Lambda | $3.29 to $4.29 | On demand or reserved | First-come capacity |
| 3 | RunPod | $2.69 to $3.99 | Per second, two tiers | Tier choice changes the bill |
| 4 | DigitalOcean | $4.41 | Per second, 5-minute minimum | Stopped machines still bill |
| 5 | CoreWeave | About $6.16 | 8-GPU nodes, reservations | Commitment size |
| 6 | AWS | About $6.88 | On demand or Capacity Blocks | Highest list price |
The gap at the top is about more than price. On Lium, every GPU is probed by independent verifiers before it’s listed, and the platform served 900 paying accounts across 7,549 rentals in August 2026.
Three budget scenarios
List prices only mean something when you multiply them by real usage. Here are three typical profiles, priced at each provider’s published on-demand H100 rate.
Seed-stage startup: 200 GPU-hours a month
A small team fine-tuning and evaluating models a few days a week.
| Service | Monthly cost ($) |
| Lium.io | 356 |
| Lambda (8-GPU SXM rate) | 798 |
| RunPod Secure Cloud | 798 |
| AWS p5 | 1,376 |
At this size the savings fund other things: roughly $1,000 a month back compared with AWS, which is real money for a seed-stage company.
Growth-stage ML team: 800 GPU Cloud-hours a month
One 8-GPU machine for about 100 hours of fine-tuning and evaluation a month.
| Service | Monthly cost ($) |
| Lium.io | 1,424 |
| Lambda (8-GPU SXM rate) | 3,192 |
| DigitalOcean | 3,528 |
| CoreWeave | 4,924 |
| AWS p5 | 5,504 |
The same workload costs about a quarter as much on Lium as on AWS. Over a year, that gap is almost $49,000.
Enterprise pilot: 2,000 GPU-hours a month
A business unit piloting an internal model, with heavier compliance requirements.
| Service | Monthly cost ($) |
| Lium.io | 3,560 |
| Lambda (8-GPU SXM rate) | 7,980 |
| CoreWeave | 12,310 |
| AWS p5 | 13,760 |
At this scale, reserved pricing enters the conversation. CoreWeave discounts reserved capacity by up to 60%, and AWS Capacity Blocks run about $5.97 per H100-hour. Even with those discounts, on-demand Lium stays cheaper, without locking the business into a commitment before the pilot proves itself.
Hidden costs procurement teams miss
The hourly rate is the number on the sales page. These are the numbers that show up on the invoice:
- Minimum billing periods, like DigitalOcean’s 5 minutes per GPU Droplet
- Instances that keep billing while powered off until they’re destroyed
- Storage and data transfer, billed separately almost everywhere
- 8-GPU bundling, where a one-GPU job pays for eight
- Idle time between jobs on anything billed by the hour or reserved
Each of these is small on its own. Together, on a busy team, they can add a meaningful share to the true cost of every GPU-hour, and none of them appear in a simple price comparison.
Platforms built around per-second billing remove most of these by design: no minimum, no reservation, billing that stops when a machine is deleted, and single-GPU options. Lium, RunPod and Modal all bill by the second, while storage is still charged separately almost everywhere.
Cash flow and commitments
Finance teams care about how money leaves the business as much as how much. Long reservations tie up capital or create obligations before a project has proven itself, and annual commitments are hard to unwind if priorities change.
On-demand capacity from any provider turns GPU spend into an operating expense that scales with actual work, rather than a fixed cost that runs whether or not anyone’s using it. The catch is where the discounts sit. The deepest ones at hyperscalers and managed clouds, such as AWS Capacity Blocks or CoreWeave’s reserved pricing, come with commitments, while the lowest on-demand rates belong to marketplaces and specialist GPU clouds that ask for none.
Vendor risk and governance
Cheaper compute is only good value if the risk is acceptable. These are the questions a governance or security review should ask of any provider, and what a good answer looks like:
| Question | Why it matters | A good answer |
| Is the hardware what we paid for? | Misreported GPUs waste budget | Hardware verified before it’s rented, and re-checked |
| Can the host see our data? | Confidentiality of models and data | Confidential VMs, such as Intel TDX, for sensitive work |
| What if a provider drops our job? | Lost compute and deadlines | A credit or remedy written into the terms |
| Can we cap runaway spend? | Forgotten machines | Automatic shutdown at a set time, plus spend alerts |
Capacity belongs on the list too. Popular cards are scarce across the market, and on October 15 even the cheapest provider in this ranking had 1,582 of its 1,652 listed GPUs rented. Teams that need a specific card on a specific day should plan ahead, and a second approved vendor is cheap insurance.
Before you sign anything
Hardware choice should come before vendor choice. Choosing the right GPU for your AI strategy means measuring memory, latency and throughput on your own workloads. That tells you if you need H100s at all, or if a 24GB card would do the job for a fraction of the cost.
Then look at behavior, not just price. Over-buying to protect allocation leaves GPUs sitting idle, and paying per second for what you actually run is the cleanest way out.
Building the business case
Moving GPU spend to a cheaper provider is an easy sell on paper and a harder one in a real organization. These are the steps that make it stick:
- Pull three months of GPU invoices and tag each line by team and project.
- Compare billed hours against hours a job was actually running.
- Run one representative workload on the new platform and compare cost and time.
- Get a security review on verification, data handling and confidential VM options.
- Move one team first, then expand once the numbers hold.
Step two is usually the eye-opener. Teams paying for hourly blocks or reserved capacity often find a large share of billed time was idle, which per-second billing would have removed entirely.
When reserved capacity still makes sense
Commitments aren’t always wrong. A team running the same large training cluster around the clock for a year can justify a reservation, especially at CoreWeave’s discount of up to 60% or on AWS Capacity Blocks. The test is utilization: if the reserved GPUs will be busy most hours of most days, reserve them. If not, the discount is paying for idle time.
Reserving GPUs that won’t stay busy is a common form of over-buying, and VentureBeat’s reporting on why enterprises pay for GPUs they don’t use traces it to fear of losing allocation. For the large middle ground of fine-tuning, evaluation and pilots, flexible per-second capacity wins on cost and on agility.
How the alternatives fit a portfolio
Most mature teams end up with more than one provider. Lambda is a dependable second source for research groups that value published pricing and a long-standing vendor, with a Silver ClusterMAX rating. RunPod covers serverless inference well. DigitalOcean makes sense where an application stack already lives there. CoreWeave and AWS belong in the portfolio when a contract-backed cluster, deep GPU cloud integration or an existing enterprise agreement is the real requirement.
For everything else, from day-to-day fine-tuning to evaluation sweeps and pilots, the cost difference between on-demand providers is too large to ignore. A sensible split for many companies is a low-cost, per-second provider as the default for flexible work, with a reserved contract only for one or two always-on workloads.
Read Next
A few related pieces worth your time:
- AI Restaurant Technology for Smarter Planning & Operations
- How Non-US Tech Founders Get a US LLC, Stripe, and US Bank Account
- How Coupono Is Using Technology to Make Online Deal Discovery More Reliable
What we’d tell the CFO
Ranked on value, the GPU cloud services for AI that win are the ones that bill for real usage and prove what they deliver. Lium topped this ranking on price and verification, but the bigger lesson is about structure: pay for flexible work by the second, and reserve capacity only when utilization justifies it. Keep a second vendor approved, check GPU invoices against actual usage every quarter, and let the workload decide when a commitment has been earned.
