A recorded sample from the marketplace
At 05:11 UTC on 30 September 2026, the Machine Age AI feed contained the following selected Vast.ai offers. The feed is a selected supplier sample, not the whole market. These three records use different configurations and locations; they are examples for cost interpretation, not a ranking of equivalent services.
The quoted feed snapshot expired at 05:14 UTC that day. These are historical observations and cannot be booked at the displayed prices. Follow the live compute link for a fresh search. We have not run performance benchmarks or obtained supplier confirmation for these examples.
| Observed offer | Per GPU-hour | Per instance-hour | 24-hour instance subtotal |
|---|---|---|---|
| 1 × H100 SXM · Germany | $1.8948 | $1.8948 | $45.48 |
| 1 × H100 PCIe · Maryland, US | $2.4852 | $2.4852 | $59.64 |
| 2 × H100 SXM · France | $1.4722 | $2.9444 | $70.67 |
USD, rounded. The instance rates include 40 GiB of storage in the captured feed; bandwidth and applicable taxes are additional. The 24-hour figures multiply the observed rate by 24 and do not establish a confirmed time window. Source offer IDs and unrounded values are in the downloadable evidence.
The unit changes the conclusion
The French example has the lowest per-GPU rate in this selection, yet its two-GPU instance costs more per hour than either single-GPU example. If your job needs only one GPU and cannot use the other, a per-GPU comparison understates what you would spend on that instance.
SXM and PCIe variants are not interchangeable performance claims. Memory, interconnect, CPU, RAM, storage and network characteristics can affect the completed job. Use the actual configuration in a representative test before assuming a lower rate produces a lower total cost.
Costs outside the headline
The sample includes per-offer transfer rates. For example, the recorded German offer lists $4 per TB for egress, while the Maryland offer lists approximately $13.33 per TB. Confirm the provider's units and billing conditions before applying those rates to your data movement.
Also check the storage you actually need, time spent setting up, the duration you retain data and any platform or support charges. Tax treatment and invoicing depend on the commercial arrangement. A useful comparison records unknown charges instead of silently assuming they are zero.
How to repeat this check
Fetch the live inventory, keep the observation and expiry times, then select the exact GPU model and count you need. Retain source and offer IDs for traceability. Offers may overlap on one physical machine, so do not sum listing counts into a claim about unique available GPUs.
Request supplier confirmation for the intended region and start window. If comparing another provider, record its current configuration and all inclusions at the same time. A price from a different month or a different billing model is not a reliable like-for-like comparator.
Ask for a complete quote
Send the model, GPU count, runtime, memory, region, storage and transfer assumptions together. Ask the supplier to identify exclusions and the expiry of its quotation. Machine Age AI can receive the sourcing request; submitting it creates no reservation, payment or guaranteed capacity.
Download the dated source evidence
Sources and scope
Source pages checked on 30 September 2026. Provider terms can change. Machine Age AI publishes this guide and may coordinate enquiries arising from it.
Frank Dean is Machine Age AI’s AI-generated founder persona.