Everyone comparing cloud GPU prices asks the same question: which provider is cheapest? We run a marketplace that aggregates eight of them, so we pulled every live offer and did the arithmetic. The answer turns out to be that the question is slightly wrong. On the GPUs most people actually rent, the gap between providers is small, and the gap between two machines inside the same provider is bigger.
TL;DR: 638 live offers, 8 providers, one snapshot. Median prices between providers differ by only 1.3x to 2.0x on most models. But on Vast.ai, the same RTX 5090 ranges from $0.37 to $2.00 per GPU-hour depending on which host you land on, a 5.5x spread inside one provider. Catalogue-priced providers land on one number and never move. Datacenter GPUs barely disperse at all; consumer GPUs disperse wildly.
Method, so you can argue with it
Every number here comes from a single snapshot of our own public marketplace feed, taken on August 4, 2026. It covers 638 offers across 8 providers: Vast.ai (338 offers), SimplePod (197), Massed Compute (39), Akash (29), Vultr (24), DataCrunch (5), Hyperstack (5) and Hot Aisle (1).
Every price is normalized to dollars per GPU per hour. That normalization is the part that goes wrong most often in published comparisons, so to be explicit: for almost every provider the feed's price is the total for the offer and the divisor is the GPU count. Akash is the exception — its listed price is already per-GPU and must not be divided. Get this backwards and you publish numbers that are wrong by exactly the GPU count, which is how a $0.20 card becomes a $1.80 card in a table nobody double-checks.
The sanity check we use, and which anyone reproducing this should use: a correct divisor makes the per-GPU rate collapse to a constant within a catalogue-priced provider. On SimplePod the CMP 170HX lands on exactly $0.200 across all 33 offers, the RTX 4070 Ti Super on exactly $0.160 across 20, the RTX 3060 on $0.152, the RTX 3070 on $0.045. If your arithmetic doesn't produce that collapse, it's wrong, and no amount of averaging will save it.
This is one point in time. Availability moves hourly, and a provider that looks cheapest today may have nothing in stock tomorrow. Treat it as a snapshot, not a league table.
The headline: providers are closer than you'd think
Here is the median price per GPU-hour by provider, for the models with enough offers to be meaningful.
| GPU | Offers | Cheapest provider (median) | Most expensive (median) | Cross-provider gap |
|---|---|---|---|---|
| RTX 5090 | 80 | Akash $0.420 | Vast.ai $0.708 | 1.7x |
| RTX 4090 | 78 | Akash $0.336 | Vast.ai $0.680 | 2.0x |
| V100 | 70 | SimplePod $0.070 | Vast.ai $0.317 | 4.5x |
| A100 | 38 | Vast.ai $1.101 | DataCrunch $1.790 | 1.6x |
| RTX 3090 | 36 | SimplePod $0.160 | Vast.ai $0.274 | 1.7x |
| RTX PRO 6000 | 19 | SimplePod $1.393 | Massed Compute $2.190 | 1.6x |
| H100 | 9 | Massed Compute $2.590 | Vast.ai $3.420 | 1.3x |
| L40S | 8 | Vast.ai $0.851 | Massed Compute $0.880 | 1.0x |
One caveat on that table, because it matters: several of those "cheapest provider" medians rest on very few offers. Akash lists only 2 RTX 5090s and 2 RTX 4090s, SimplePod only 1 RTX PRO 6000, Hyperstack only 1 A100. A median over two offers is a data point, not a market rate. The rows you can lean on are the ones with real depth on both sides — V100 (70 offers), RTX 4090 (78), RTX 3090 (36), A100 (38).
Most of those gaps are under 2x, and on the datacenter parts they're under 1.5x. If you switched providers purely to chase the H100 median you'd save about 24%, which is real money at scale but not the order-of-magnitude difference the marketing in this space implies. Note the H100 sample is thin too — 9 offers across 3 providers — which is itself a finding: most H100 supply is sold on contracts that never appear in an on-demand feed.
The V100 is the exception and it's instructive: 4.5x, because SimplePod is clearing old inventory at $0.070 while Vast.ai hosts price the same card at $0.317. Aging hardware is where provider choice actually pays.
The thing nobody puts in a comparison table
Now the same data cut the other way — the range within a single provider, for a single GPU model.
| GPU | Provider | Offers | Cheapest | Most expensive | Spread |
|---|---|---|---|---|---|
| RTX 5090 | Vast.ai | 77 | $0.367 | $2.004 | 5.5x |
| RTX 5080 | Vast.ai | 9 | $0.272 | $1.171 | 4.3x |
| RTX 4090 | Vast.ai | 58 | $0.360 | $1.274 | 3.5x |
| RTX 3090 | Vast.ai | 17 | $0.197 | $0.670 | 3.4x |
| RTX 4070 Super | Vast.ai | 10 | $0.171 | $0.545 | 3.2x |
| A100 | Vast.ai | 25 | $0.734 | $2.189 | 3.0x |
| RTX 4070 Ti S | SimplePod | 20 | $0.160 | $0.160 | 1.0x |
| CMP 170HX | SimplePod | 33 | $0.200 | $0.200 | 1.0x |
One row we deliberately left out, because it looks like a 4x spread and isn't: Vultr's A16 shows $0.059, $0.118 and $0.236 — exactly 1x, 2x and 4x of the same base. Those are almost certainly three fractional slices of one card, and every one of them reports a GPU count of 1, so the feed can't distinguish a quarter of an A16 from a whole one. That's a measurement artifact, not price dispersion, and treating it as a bargain would mean comparing a quarter card against someone else's whole one. When a "spread" lands on clean powers of two, suspect the unit before you celebrate the discount.
On an RTX 5090, the within-provider spread on Vast.ai (5.5x) is more than three times the entire cross-provider gap (1.7x). The decision that actually determines your bill is not which logo you pick. It's which machine you land on after you've picked.
This isn't a defect in Vast.ai. It's what a host marketplace is — independent operators set their own prices, and their hardware, bandwidth and reliability genuinely differ. But it means "Vast.ai is cheap" and "Vast.ai is expensive" are both true statements about the same platform on the same day, and a comparison article quoting a single Vast.ai number for the 5090 is quoting one draw from a distribution that spans $0.37 to $2.00.
The contrast with catalogue-priced providers is total. SimplePod's 170HX is $0.200 on all 33 offers. Not approximately $0.200 — exactly, every time, and the same holds for its RTX 4070 Ti Super ($0.160), RTX 3060 ($0.152) and RTX 3070 ($0.045). You give up the chance of finding a bargain host and you get a number you can put in a budget.
Akash deserves a footnote here, because it's easy to mislabel. Its listed price is quoted per GPU regardless of node size — an A100 lease is $1.6275 per GPU whether you take 8 or 68, which is why it must never be divided by GPU count. That is not the same as every Akash offer costing the same: it's a decentralized market, and its two RTX 5090 listings sit at $0.158 and $0.683. Invariant to node size, not invariant across sellers.
Consumer GPUs disperse, datacenter GPUs don't
Sort the models by spread and a clean pattern falls out.
The wide ones are all consumer cards: RTX 5090, 5080, 4090, 3090, 4070 Super. The tight ones are almost all datacenter parts: H100 (1.3x cross-provider), H200 (1.3x), L40S (1.0x), RTX PRO 6000 (1.6x), A6000 on Massed Compute (1.1x).
That makes sense once you say it out loud. An H100 is sold by a small number of operators running real datacenters with roughly comparable cost structures, competing on a commodity against a published market rate. A 4090 might be in a rack in Virginia or under someone's desk, and those two things do not cost the same to run. Consumer GPU pricing is a distribution of operators; datacenter GPU pricing is a market rate.
Practical consequence: if you rent consumer GPUs, shopping within a provider is worth more than shopping between providers. If you rent datacenter GPUs, neither saves you much, and you should be optimizing for availability instead.
Where the money actually is
Three findings worth acting on.
1. On consumer cards, catalogue-priced providers beat the marketplace median. SimplePod's RTX 4070 Super is $0.090 against a Vast.ai median of $0.252. The RTX 4070 is $0.080 against $0.186. The V100 is $0.070 against $0.317. You can beat those numbers by finding the cheapest Vast.ai host, but you have to actually go find it, every time, and it may be gone tomorrow.
2. The B200 has one supplier in this feed. Eight offers, all Vast.ai, $5.095 to $7.635 per GPU-hour. Newest silicon means no competitive floor yet — you're paying whatever the small number of people who have them decide to charge. The MI300X is thinner still: a single node.
3. The A100 is the most contested part on the board. Five providers, medians from $1.101 to $1.790, a 1.6x gap and a 3.0x within-provider spread on top. That's a mature market with real competition, which is exactly where careful shopping pays for itself.
Why a 5x spread only matters if you can move
Here's the catch that makes all of this less useful than it looks. A 5.5x price spread is only worth something if acting on it is cheap. If moving to the cheaper box means an afternoon reinstalling CUDA, rebuilding a Python environment, re-downloading eighty gigabytes of model weights and rediscovering which flag made it work last time, then a 30% saving is not a saving. It's a worse hourly rate on your own time.
That gap is the entire reason we work on this. Price transparency without portability is a list of things you can't act on — and it's why the honest version of this post has to end on the constraint rather than the table. If you want the mechanics of moving a workload between providers without rebuilding it, we wrote that up separately in moving a GPU workload to another cloud provider and snapshot and restore across providers.
The live version of every number above is on the marketplace and the GPU index, which is where to look before acting on a table that was accurate on one August afternoon.
FAQ
Which cloud GPU provider is cheapest? It depends on the GPU, and less than you'd expect. In this snapshot, medians differed by only 1.3x to 2.0x on most models. SimplePod was cheapest on consumer cards and older datacenter parts, Massed Compute on H100, Vast.ai on A100 — but Vast.ai's own range on a single model is often wider than the gap between providers.
Why does the same GPU cost different amounts on the same provider? On host marketplaces like Vast.ai, independent operators set their own prices, and their hardware, network and reliability differ. Fixed-price providers such as SimplePod and Akash quote one rate that doesn't vary between offers.
How do you calculate price per GPU? Divide the offer's total price by its GPU count, with the exception of Akash, whose listed price is already per-GPU. The check that your arithmetic is right: on a catalogue-priced provider such as SimplePod, every offer of the same model must land on exactly the same per-GPU number.
How much can I save by switching providers? On datacenter GPUs, typically 20 to 30%. On consumer GPUs and older cards, sometimes 3 to 4x — but the same or better is often available within a provider by picking a different host.
Is a cheaper host worse? Sometimes. Price on a host marketplace reflects hardware generation, bandwidth, storage and reliability alongside the GPU itself. The cheapest offer is not automatically the best value, which is part of why the spread persists.
What this snapshot changed for us
We expected to find one cheap provider and one expensive one. What we found is that provider choice is a second-order decision on most GPUs, and the first-order decision is which specific machine you take — a choice that no comparison table can make for you, because it changes hourly.
That's an argument for watching the market continuously rather than picking a home and staying there. Which only works if leaving is cheap.
About the author
I'm Ansh Saxena. I work on the infrastructure layer under rented GPU boxes, mostly on making a machine's whole state portable so that acting on a price difference doesn't cost you a day of rebuilding. I have an obvious interest in you believing that portability matters, which is why every number above is reproducible from a public feed rather than something you have to take on trust.
Sources
- Aquanode public marketplace feed, snapshot taken 2026-08-04. 638 offers across Vast.ai, SimplePod, Massed Compute, Akash, Vultr, DataCrunch, Hyperstack and Hot Aisle. Live data at /marketplace and /gpu-index.
- Per-GPU normalization follows the shared helper used across our own pricing surfaces: offer total divided by GPU count, with Akash treated as a flat per-GPU rate.
