The Cheapest GPU Cloud in 2026

The short answer

The lowest published rates almost always come from marketplaces and spot capacity, where an operator is passing through GPUs that would otherwise sit idle, with no guarantee the machine is there when you come back. That is genuinely cheap for work that tolerates interruption and a false economy for anything that has to stay up. The cheapest headline rate and the cheapest monthly bill are two different numbers, and the gap between them is where most of the money is won or lost.

01 · SourcingWhere a cheap GPU hour comes from

A rate that looks too good to be true usually has a structural reason behind it. The cheapest capacity on the market is spot or marketplace supply, where an operator with underused hardware lists it at whatever clears rather than leaving it dark and a platform passes that price through to you. There is no committed availability behind it and no support relationship beyond the platform, so the machine can be reclaimed when a higher bidder or the operator itself needs it back. For a batch job that checkpoints and resumes, that is a fair trade and a real saving.

The next tier down in price is resold capacity, where a provider rents from a larger cloud and marks it up or down depending on how it fills its own commitments. Neither of these is dishonest, yet both mean the low number you are looking at is a floor for a specific commercial structure rather than a like-for-like discount on the product a direct operator sells. Knowing which structure produced the quote tells you far more than the digits do.

02 · The billThe cheapest headline is rarely the cheapest bill

The per-hour price is the most visible line and often the smallest one. Many providers price the accelerator alone and meter the host CPU, the memory, the attached storage and the egress separately, so two quotes with the same headline can land a long way apart once a real month of work runs through them. A training job reads its dataset many times, writes checkpoints continuously and pushes a final artefact out to somewhere else and each of those touches a meter the cheap headline never mentioned.

Interruption has a cost too and it does not show up on any price page. If a spot instance is reclaimed mid-run and your pipeline does not resume cleanly, the wasted hours and the engineer time spent restarting it can quietly erase the saving that drew you to the cheap rate in the first place.

03 · SupplyWhere the cheap supply actually sits

Cheap follows competition. The widely-listed prior-generation accelerators are where the most operators compete for the same workload, which is where a budget buyer has the most room to move. The counts below are the number of tracked providers listing each accelerator, a stand-in for how contested each card is rather than any quoted rate.

The scarce flagship cards concentrate among fewer operators, so chasing only the newest silicon means competing for a thinner supply at a firmer price. The prior-generation A100, L40S, RTX 4090 and V100 are each listed by dozens of operators, which is exactly where the cheap end is deepest for a workload that does not need the frontier.

Where the cheap capacity is sourced

The cheapest rates come from operators reselling or passing through capacity, not from the operators that run their own metal. viabandwidth verifies operator type for 160 of the 1,009 indexed operators so far. The 144 network-verified direct operators are the premium, steadier-priced tier; resold and not-yet-classified capacity is where the cheapest pass-through rates tend to sit.

Sourcing is the single filter that separates a real quote from a reseller markup, and it is the one thing a buyer can almost never see on the open web. The verified-direct count is small on purpose, because it only includes operators whose own network evidence we could confirm.

04 · ProcurementHow to shop the cheap end without getting burned

• 01Start from the workload, not the price. If the job tolerates interruption and restarts cheaply, shop spot and marketplace capacity hard. If it has to stay up, a direct operator with committed availability is cheaper once you count the failures you avoided.
• 02Match the card to the job. A widely-listed prior-generation card that fits the work beats a rented flagship you only half-use.
• 03Read past the per-hour number. Ask for the all-in cost of a workload shaped like yours, storage and egress included, in writing.
• 04Layer the commitment. A committed floor for the work you will certainly run, with cheaper on-demand or spot on top for the variable part.