The best cloud GPU for AI work is decided by one number before anything else: what an hour of compute costs. The same NVIDIA H100 rents for under two dollars an hour on a marketplace and about eleven dollars an hour on a hyperscaler. Same chip. Same 80 GB of memory. A gap of five times or more, and it’s got nothing to do with performance. It’s about where you rent it.
What follows is a best cloud GPU for AI comparison built the boring way: every rate comes from a published pricing page or a tracker that cites one, and every range is hedged, because these numbers move weekly.
How cloud GPU pricing actually works
Strip away the marketing and there are really only three pricing models in this market, plus three kinds of seller.
On-demand is the default: a published hourly rate, yours until you stop it. Spot, also called preemptible, rents spare capacity at 50 to 80 percent off, but the provider can reclaim it with little warning. It suits checkpointed training, and it’s fatal for anything that can’t be interrupted. If you’re hunting for the best cloud GPU for AI on a tight budget, spot is usually where the search ends; most practitioners who rent H100 capacity by the hour learned that first. Reserved pricing trades a commitment of one month to three years for roughly 30 to 60 percent off.
Sellers fall into three camps. Hyperscalers (AWS, Google Cloud, Azure) sell GPUs inside giant cloud ecosystems: expensive per hour, but with compliance certifications, private networking, and tooling enterprises already run on. Neoclouds (Lambda, CoreWeave, RunPod, Nebius, Crusoe) are GPU-first companies built around AI workloads, typically 40 to 60 percent cheaper than hyperscalers on identical NVIDIA hardware. Marketplaces (Vast.ai and similar) let independent hosts list spare GPUs: the cheapest option by far, with variable reliability and no meaningful support.
Prices are quoted per GPU-hour and move constantly. Recent snapshots put the volume-weighted H100 average in the mid-three-dollars range; the cheapest advertised on-demand rates sit near $1.79 while hyperscaler listings reach roughly $11. That’s the pricing landscape every best cloud GPU for AI buyer is navigating. Treat every figure below as a recent snapshot, and check live cloud GPU pricing before a long run.
Best cloud GPU for AI: the 8 providers, compared
All prices below are on-demand per-GPU-hour from recently published best cloud GPU for AI comparisons. Marketplace figures are ranges because hosts set their own prices. The H200 column matters more than it used to: NVIDIA’s 141 GB card is now the default recommendation for memory-bound inference.
| Provider | Type | H100 80GB ($/hr) | H200 141GB ($/hr) | A100 80GB ($/hr) | Best for |
|---|---|---|---|---|---|
| Lambda Labs | Neocloud | $3.29-$3.99 | — | $1.99-$2.79 | Straightforward ML-ready rentals |
| RunPod | Neocloud | $2.89-$2.99 | ~$4.59 | $1.00-$1.49 | Iterative experiments |
| Vast.ai | Marketplace | ~$1.40-$2.10 | — | from ~$0.53 | Cheapest possible prototyping |
| CoreWeave | Neocloud | $6.16 | ~$6.30 | ~$2.70 | Enterprise-scale training |
| Nebius | Neocloud | ~$3.85 | ~$4.50 | — | European data residency |
| Crusoe Cloud | Neocloud | ~$3.90 | — | — | Clean-energy procurement |
| AWS | Hyperscaler | ~$6.88 | ~$7.91 | ~$3.43 | Teams already on AWS |
| Google Cloud | Hyperscaler | ~$11.00 | ~$10.60 | — | GCP-native teams |
One more thing the table hides: the minimum purchase unit. CoreWeave’s $6.16 headline rate is node-locked to 8-GPU minimums, which is $49.24 an hour for the node, with no self-serve single-GPU option. Nebius at $3.85 and Lambda at $3.99 both allow single-GPU provisioning. If your job needs two GPUs, the effective per-GPU cost on CoreWeave’s on-demand tier jumps past $24. The headline rate is not the price. The minimum unit is part of the price. Any best cloud GPU for AI shortlist that ignores the minimum unit is comparing fiction.
I cross-checked the table against two sources that verify rates firsthand, because a best cloud GPU for AI comparison is only as good as its rates. Lyceum Technology’s team reads every rate off the provider’s own pricing page instead of copying aggregators, and when I laid their table next to FastGPU’s 28-cloud tracker, the same gaps showed up in both: RunPod Secure Cloud H100 at $2.99 against Lambda’s $3.99 and CoreWeave’s $6.16. Lyceum also flagged something I’d seen in older comparisons: the $12.29-an-hour Azure figure is a launch-era price, not a current one, so it’s excluded from every range above.
Provider by provider: who each one is for
Lambda Labs: the straightforward default
For researchers who want to train minutes after signing up, Lambda is the least painful on-ramp among fixed-price providers. Listed H100 rates run about $3.29 to $3.99 an hour in recent comparisons, with A100s from $1.99 to $2.79. Machines arrive with PyTorch and the usual ML stack preinstalled, there are no egress fees, and single-GPU provisioning is allowed, so there’s no Kubernetes to learn first. When people ask me where I would point a small research team hunting for the best cloud GPU for AI, Lambda is usually my first sentence.
RunPod: the experimenter’s workbench
Per-second billing with one-minute minimums. That’s the whole pitch, and for the loop most ML developers live in, spin up, run an experiment, tear down, repeat, it’s the right one. The H100 lists near $2.89 an hour on the vetted Secure Cloud tier, with cheaper community and spot options below it. An RTX 4090 starts near $0.34 an hour, AMD’s MI300X lists around $1.90, and the H200 sits near $4.59 on Secure Cloud. That combination makes RunPod my default for experiment-heavy work and a frequent winner in best cloud GPU for AI comparisons for iterative fine-tuning. The catch is the two tiers: Secure Cloud hosts are vetted and SLA-backed, Community Cloud hosts are cheaper and less predictable. Know which tier you’re on before a long run.
Vast.ai: the bargain basement, with the risk to match
Nothing else comes close on price, and nothing else asks you to accept this much risk. Vast.ai is a marketplace where independent hosts rent out spare GPUs, from gaming PCs to data-center racks. H100 listings show up around $1.40 to $2.10 an hour depending on host and timing, A100s from about fifty cents, L40S cards from about $0.47, and RTX 4090s from roughly $0.40. Hosts can disappear mid-job, there’s no support line, and multi-GPU setups are limited. Checkpoint aggressively and treat it as disposable compute. For students, hobbyists, and anyone whose training run can survive an interruption, it’s the cheapest cloud GPU you’ll find anywhere.
CoreWeave: built for hundreds of GPUs, priced like it
CoreWeave is the neocloud heavyweight: GPU-dense fleets, InfiniBand networking, Kubernetes-native orchestration, contracts built for companies training at real scale. Its listed H100 rate is $6.16 an hour in mandatory 8-GPU bundles, with spot near $2.46 and one-year reserved near $2.00 in tracked comparisons. It was among the first clouds to offer NVIDIA’s Blackwell cards at scale, with B200s listed around $8.60 and H200s near $6.30. This is the right call when you need hundreds of GPUs networked together with enterprise SLAs. It’s the wrong call for one card over a weekend. The node minimums punish small jobs.
Nebius: the European option
Geography is the differentiator. Nebius, the former Yandex cloud business rebuilt as an independent AI cloud in Amsterdam, runs EU data centers with GDPR-compliant data residency no US neocloud can match for regulated European workloads. H100s list around $3.85 an hour on-demand with preemptible capacity near $2.15; the H200 near $4.50 and the B200 near $7.15. Single-GPU provisioning is allowed. For background on the physical infrastructure behind providers like this, see our explainer on what AI data centers are.
Crusoe Cloud: flare gas into compute
Crusoe’s pitch is energy, not just silicon: modular AI data centers with clean-energy positioning built on captured flare gas. Its H100 lists near $3.90 an hour on-demand, squarely in the neocloud middle of the pack. If energy sourcing factors into procurement, it belongs on the shortlist.
AWS: you’re renting the ecosystem, not the chip
An H100 on AWS runs about $6.88 an hour per GPU on p5 instances, with per-second billing. You’re paying for compliance certifications, private networking, and petabyte-scale storage integration, not for faster silicon. If your company already runs on AWS and moving data out would be its own project, the premium can be rational. For everyone else, it’s roughly double the neocloud price for the same card.
Google Cloud: steepest list prices, deepest discounts
Google Cloud’s list prices reach roughly $11 an hour on-demand for top H100 instances, though committed-use discounts cut that sharply. Same logic as AWS: the premium buys the ecosystem, including TPU workloads for GCP-native teams, not faster NVIDIA silicon. One caution: verify the rate on the provider’s own pricing page before budgeting. Stale launch-era figures still circulate, and the gap between an old number and today’s rate can double the apparent cost difference.
Best cloud GPU for AI by workload: which card to rent?
Forget the provider: the card matters more than the logo on the invoice, and picking the best cloud GPU for AI by workload starts with what limits your job.
| GPU | VRAM | Typical cloud rate | What it is for |
|---|---|---|---|
| RTX 4090 | 24 GB | from ~$0.34/hr | Inference and QLoRA fine-tuning up to ~30B parameters; the budget sweet spot |
| L40S | 48 GB | $0.92-$1.19/hr | The value pick: most of an A100’s practical usefulness at a fraction of the cost |
| A100 80GB | 80 GB | $1.02-$2.79/hr | The default GPU for fine-tuning up to 70B; the most cost-efficient data-center card |
| H100 80GB | 80 GB | $1.79-~$11/hr | Serious training and large-scale inference; 2.2 to 2.5x faster than A100 on QLoRA workloads |
| H200 141GB | 141 GB | $2.60-~$11/hr | Memory-bound inference: one H200 holds 70B weights where two H100s were needed |
| B200 | ~180 GB | $3.69-$14.24/hr | Cutting-edge training and FP4 inference; typically needs liquid cooling |
| MI300X (AMD) | 192 GB | $1.71-$3.45/hr | Memory-bound inference on a budget; competitive for ROCm-compatible workloads |
The H200 deserves the attention it’s getting. With 141 GB of HBM3e at 4.8 TB/s, nearly double the H100’s 80 GB, it’s built on the identical Hopper die, so compute throughput is the same and the whole upgrade is memory. The H100 vs H200 decision is therefore simple: for memory-bound inference, the premium, roughly 15 to 50 percent over the H100 depending on provider, often pays for itself. A single H200 can hold a 70B model’s weights where two H100s were previously needed, which simplifies the entire serving setup. For compute-bound training, though, the H200 buys nothing the H100 doesn’t already do. Newer isn’t automatically better, and the cheaper card wins when memory isn’t the bottleneck. For memory-bound inference, the H200 is the best cloud GPU for AI you can rent right now.
The fine-tuning math is where most buyers overspend, and where I stopped trusting rules of thumb. With QLoRA, quantizing a model to 4-bit and training small adapters instead of the whole model, a 70B model fits in roughly 38 GB of VRAM. That’s why the A100 80GB became the default GPU for fine-tuning open models. These aren’t theoretical numbers, and I didn’t take them on faith: packet.ai, a GPU cloud that rents the same cards it writes about, puts a full 70B QLoRA run at $34 to $51 and 24 to 36 hours on a single A100 80GB, and measures the H100 at 2.2 to 2.5 times faster on those same QLoRA workloads. Note that’s QLoRA specifically, not every workload. Full fine-tuning without adapters is a different animal: a 70B model at full precision wants around 420 GB for weights, optimizer states, and gradients combined, roughly eleven times the 4-bit figure, which means at least six A100s or three H100s, and a cloud bill to match.
Data-center GPUs are a different species from gaming cards. An RTX 4090 tops out at 24 GB of GDDR memory. An H100 carries 80 GB of much faster HBM with error correction, NVLink for multi-GPU scaling, and hardware support for the FP8 precision modern training relies on. For shoppers comparing the best cloud GPU for AI on specs alone, this is the distinction that saves the most money. Consumer cards are fine for learning and small jobs. Production training wants data-center silicon.
Rent or buy: the breakeven math
An H100 costs roughly $25,000 to $35,000 new, and the urge to just own one is understandable. But the rent-versus-buy question, the other half of finding the best cloud GPU for AI, reduces to one variable: utilization. Below roughly 40 percent sustained use, renting wins; above 40 to 60 percent, ownership pulls ahead, and the line moves against buying once you count honestly.
A $30,000 card against a $2.50-an-hour rental breaks even near 14,000 GPU-hours once you count power, cooling, and hosting alongside the card, about nineteen months at full utilization. At 60 percent utilization, one detailed analysis puts ownership’s effective cost at $4.25 an hour against marketplace rental at $1.50 to $3.15, with breakeven at eighteen to twenty-two months. And depreciation is brutal: assume roughly half the card’s value gone by month eighteen. That’s the math behind every honest best cloud GPU for AI rent-versus-buy analysis. The unifying lesson: a build-versus-buy comparison that doesn’t state its utilization assumption hasn’t made an argument.
What I’d actually do
Six rules I’d hand a friend shopping for the best cloud GPU for AI:
- Match the card to the bottleneck. The best cloud GPU for AI is the one sized to your constraint, not the biggest one: memory-bound work wants the H200, compute-bound training the H100, fine-tuning at 70B and below the A100 80GB, experiments the RTX 4090 or L40S.
- Never accept hyperscaler on-demand without a reason. Shopping providers routinely saves 40 to 60 percent on the same silicon. The premium buys compliance and ecosystem, not speed.
- Check the minimum unit, not just the rate. An 8-GPU minimum turns a $6.16 headline into $49.24 an hour for a small job. Prefer single-GPU provisioning until you genuinely need scale.
- Use spot for anything checkpointed. Fifty to eighty percent off changes what you can afford to try. Reserve baseline capacity, burst on spot.
- Measure before committing. Track a month of real GPU-hours before signing anything longer than monthly. Cloud GPU pricing moves constantly; today’s best deal isn’t a contract.
- Price your first fine-tune before you rent a cluster. One A100 80GB with QLoRA handles a 70B model for roughly $34 to $51 all-in. Start there, and scale up when measurements, not anxiety, say so.
If you’re renting GPUs for AI agents or research workloads, our guides to AI agents for scientific research and AI coding tools for beginners cover what to do with the compute once you’ve rented it.
FAQ
Which cloud GPU provider is the cheapest?
Vast.ai, in every comparison I checked: H100 listings around $1.40 to $2.10 an hour, because it’s a marketplace of independent hosts rather than a company with data centers. RunPod is usually the cheapest vetted option. The tradeoff is reliability: marketplace hosts can vanish mid-job, so checkpoint your work and treat the compute as disposable.
How much does it cost to rent an H100 per hour?
Recent snapshots show on-demand H100 rates from about $1.79 an hour at the cheapest specialists to roughly $11 at hyperscalers, averaging in the mid-three-dollars range. To rent H100 capacity on a neocloud, expect roughly $2.89 to $3.99; RunPod, Lambda, and Nebius all sit in that band, and those are the bands that matter in any best cloud GPU for AI price comparison. Spot and marketplace rates go lower. Prices move, so check live rates before a long run.
Should I rent an H100 or an H200?
Rent the H200 when memory is the bottleneck: big-model inference, long context, large KV caches. Rent the H100 for training throughput, wider availability, and lower hourly cost. The H100 vs H200 question comes down to one thing. Same chip, different memory. Buy memory only when memory is what limits you.
Is it cheaper to rent or buy a GPU for AI work?
It depends on utilization, and only on utilization: roughly 40 to 60 percent sustained usage is the breakeven. Below that, renting wins; above it, buying or a long reservation wins. Practical tip: measure a month of real GPU-hours before signing anything. Most people overestimate how busy their card will be. It’s the cheapest advice here: free, and worth more than any pricing table.
What GPU do I need to fine-tune an LLM?
Less than you’d think. With QLoRA, a 70B model fits on a single A100 80GB, and a full run costs roughly $34 to $51. That’s the standard recipe: one card, quantized training, no cluster.
Can I use a consumer GPU like the RTX 4090 for AI training?
Yes, for many jobs. The 24GB RTX 4090 handles inference and QLoRA fine-tuning on models up to about 30B parameters and rents from around $0.34 an hour, which is why it’s the default answer when someone asks for the cheapest cloud GPU that can still train. Its limits are real, though: 24GB caps larger jobs, consumer GDDR is far slower than data-center HBM, and there’s no error correction or NVLink scaling. For production training, data-center GPUs remain the practical choice.
Sources
- Mercatus, H100 cloud price tiers and reserved-instance discounts
- Spheron, H100 buy-vs-rent prices and per-provider rental rates
- FastGPU, Live rental prices from 28 GPU clouds
- Lyceum Technology, First-party GPU cloud price comparison: Lambda vs RunPod vs Vast.ai
- RunPod, NVIDIA H200 specs, pricing, and H200 vs H100
- packet.ai, Fine-tuning cost guide: A100 vs H100, QLoRA VRAM math
- MillionMiner, Cloud GPU rent-vs-buy breakeven analysis
- GPUNex, Rent vs buy GPU server: TCO and depreciation
- Thunder Compute, Nebius GPU cloud pricing and overview
- MLQ.ai, GPU pricing dashboard: floor and ceiling rates by card
- TechPowerUp, NVIDIA H200 SXM 141GB specifications
- David Lawrence, Using spot instances on AI cloud GPUs
- AltStreet, GPU price comparison: spot and on-demand tracking
- GPUAdvisor, Cheapest H100 cloud: reserved and spot tiers
That’s the whole game in finding the best cloud GPU for AI: start from your bottleneck, memory, throughput, or budget, check the minimum billing unit alongside the hourly rate, and assume every price you’ve seen has already moved a little.
Leave a Reply