best 10 gpu cloud providers

runpod lists h100 sxm at $2.99 on secure cloud and $2.69 on community cloud, with the pcie variant from $1.99, and b200 at $5.89 across both tiers. compute and pod storage bill per second, and there are no ingress or egress fees at all — a combination nothing else here matches at this price.

the trap is in the storage table, and it inverts the advice everyone gives. volume disk costs $0.10 per gb per month while a pod runs and $0.20 while it is stopped. shutting a pod down to save money doubles the storage line, and network volumes bill hourly even though compute bills per second.

the gap is accountability. there is no published sla or uptime percentage for standard pods; 'sla-backed uptime' appears only against enterprise reserved clusters at ten-thousand-gpu scale, with no figure attached. at this price that is a reasonable trade, but it should be a decision rather than a surprise.

$1.14 per gpu-hour for a spot h100 is the lowest verified figure in this ranking, and b200 spot at $2.14 is likewise the cheapest blackwell access we found. the on-demand rate of $3.25 is mid-field, so the entire case here rests on tolerating interruption — which for checkpointed training is usually the right trade.

the name is the practical hazard. datacrunch is now verda, and datacrunch.io redirects to verda.com, so every existing comparison, bookmark and cited price points at a brand that no longer exists. if you are checking our figures against another source, make sure that source knows.

two things to read carefully. storage is $0.20 per gibibyte per month — gibibyte, not gigabyte, a 7% difference that is easy to mis-transcribe. and egress, billing increment and sla are all absent from the pricing page, so the total cost of a long-running job needs a sales conversation rather than a calculator.

$2.50 per gpu-hour on-demand undercuts everything here except the marketplaces and runpod's community tier, and it comes with free ingress and egress, per-minute billing, and ssd volumes at roughly $0.07 per gb per month. b200 is $6.00 on-demand or from $5.10 reserved.

the funding rule is the thing to internalise before you rely on it: hyperstack terminates resources after a one-week grace period if the account balance isn't topped up. that is a reasonable policy for a prepaid service and a genuinely bad surprise for a long training run funded by a card that expired.

one ambiguity we could not resolve: the pricing page does not disambiguate h100 pcie from sxm at the on-demand rate, and the $2.00 figure isn't clearly labelled as a true interruptible tier rather than a discounted sku. those are meaningfully different products and the page treats them as one.

99.9% uptime, stated plainly on the pricing page. that sounds unremarkable until you notice that eight of the other nine providers here publish nothing at all, including the ones selling to enterprises. combined with per-minute billing, no minimum commitment, storage bundled into the hourly rate and sub-two-minute deployment, it is a coherent offer for bursty work.

the price needs care. the pricing page says $2.54 on-demand and $1.67 spot; spheron's own comparison blog posts advertise $2.01 for the same thing. we have quoted the pricing page because that is what you will be charged from, but a vendor undercutting itself in its own marketing is a reason to confirm at checkout.

the structural caveat is that this is a decentralised network rather than a single operator, so the machine you get comes from an underlying host and specs, availability and real-world performance vary with it. the 99.9% figure is spheron's promise about that network, not about any particular node.

everything about data movement is free here: egress, iops, and transfer within coreweave. object storage runs $0.06 per gb monthly hot, $0.03 warm and $0.015 cold — the cheapest cold tier we found. for training pipelines that shuttle checkpoints and datasets constantly, that structure can matter more than the hourly rate.

the hourly rate needs arithmetic. coreweave prices per eight-gpu hgx node: $49.24 on-demand, which is $6.16 per gpu and expensive, and $19.71 spot, which is $2.46 per gpu and competitive. we have seen the node figure quoted as though it were per-gpu, and the error runs in the direction that makes coreweave look eight times worse than it is.

the omissions are the usual ones: no published sla, no billing increment, no idle-billing policy, and reserved capacity discounts of up to 60% that require a sales conversation with no published terms. b200 is $68.80 per node on-demand, $34.11 spot, and gb200 has no spot option at all.

vast.ai advertises h100 sxm from $1.92 and pcie from $1.72, with interruptible listings described as 50%+ cheaper again. billing is per second with no minimum hours and no rounding, which is the strictest granularity here.

the important qualification is that this is a peer-hosted marketplace, not a cloud with a price list. those figures are floors that move with supply and demand, individual hosts set their own terms, and some listings quote a rate 'plus bandwidth' without publishing what bandwidth costs. we could not confirm live rates because the pricing console renders through javascript.

so the honest framing is that vast is cheaper than everything here and less certain than everything here, at the same time. for checkpointed batch jobs that is often the right trade; for anything with a deadline and an sla requirement, the eight providers around it are safer.

$3.85 on-demand and $2.15 preemptible for h100, with b200 at $7.15 and $3.95, published clearly per gpu per hour. b200 pricing at all is worth noting — several providers here still gate blackwell behind a reservation or don't list it.

european operation is the differentiator. for teams whose blocker is where the data sits rather than what the compute costs, that narrows the field considerably, and nebius is among the few at this scale.

what the page doesn't say is most of what determines your bill. no egress rate, no billing increment, no idle-billing policy, no sla, and no statement of whether you can rent a single gpu or must take a node. object storage is priced between $0.0147 and $0.11 per gibibyte monthly, but block storage for compute isn't confirmed. reserved discounts reach 35% through sales only.

lambda's per-gpu h100 rate is $4.29 if you rent one and $3.99 if you rent eight — a volume curve rather than a single price. that matters because comparisons routinely quote $3.29 for lambda h100, a figure we could not find on lambda's own pages at either end of the curve. b200 follows the same pattern at $6.99 down to $6.69.

egress is free, which at this price point is a real offset, and on-demand access is self-serve on a first-come basis rather than gated behind a reservation. longer commitments and one-click clusters exist at lower rates through sales.

the disclosure gaps are wide for a provider of this profile: no storage pricing, no billing increment, no idle-billing policy and no sla on either pricing page we read. there is also no spot or preemptible tier at all, so the cheap-and-interruptible strategy that makes verda and coreweave attractive isn't available.

two things work well. you can rent a single h100 rather than a node, and compute billing stops the moment the machine is powered off, with only block storage continuing. bandwidth is free in both directions. for exploratory work with long idle gaps that model beats hourly-committed rivals even at a higher sticker.

the sticker is the problem. $5.95 per gpu-hour is second only to coreweave's on-demand node rate here, and digitalocean's own docs label it a promo rather than the list price. we could not establish what it reverts to, which makes it a poor basis for planning a multi-month project.

there is no confirmed spot tier, no published sla, and the per-second billing claimed in secondary sources isn't corroborated on the docs page — which describes hourly rates and power-state billing only. a three-year committed rate of $2.24 appears in third-party summaries and not on digitalocean's own page, so we haven't published it as fact.

capacity blocks is a genuinely differentiated product: you reserve a defined block of h100 time on defined dates and aws guarantees it. nothing else in this ranking sells certainty as a product. at $41.528 an hour for eight h100s in us regions — $5.19 per gpu — and $37.76 internationally, you pay for that certainty.

the constraint most comparisons omit is that aws does not sell a single h100. the p5.48xlarge is a fixed eight-gpu instance, so the smallest purchase is roughly eight times the smallest purchase at runpod or vast. a per-gpu figure placed beside a specialist's in a table describes a transaction you cannot make.

we could not verify classic on-demand p5 pricing: aws's on-demand pages render through javascript and produced no readable table across attempts, so the figure above is explicitly the reservation product. in practice that may matter less than it sounds — on-demand p5 capacity is frequently unavailable, which is why capacity blocks exists.