Get the latest on AI, LLMs & developer tools
New MCP servers, model updates, and guides like this one — delivered weekly.
The Short Answer
The honest version: pick the provider that fits the job, not the one with the best homepage.
- Just tinkering or fine-tuning on a budget? Rent from Vast.ai for the lowest price, or RunPod if you want the low price with a friendlier setup.
- Serious training where you want it to just work? Lambda Labs. Not the cheapest, but the pricing is published, there are no egress fees, and the workflow is built for people who train models.
- Serving a model or running bursty jobs? Modal. Per-second billing that scales to zero beats paying for an idle GPU.
- Training across 100+ GPUs where interconnect decides throughput? CoreWeave, for the InfiniBand fabric and reserved capacity.
If I had to hand one provider to a small team doing a mix of training and experiments and walk away, it would be Lambda Labs — reliable, transparent, no surprises on the bill. But I would send a solo tinkerer to Vast.ai or RunPod, and I would not put a 200-GPU run anywhere but CoreWeave. The rest of this page is the why.
At a Glance
| Provider | Best for | ~H100 price | Main caveat |
|---|---|---|---|
| RunPod | Cheap self-serve, tinkering, serverless inference | ~$2.39/hr (spot ~$1.99) | Single pod caps at 8 GPUs; InfiniBand is not the default |
| Lambda Labs | Reliable training with a research-friendly workflow | ~$2.49–$3.29/hr, no egress | Not the cheapest; capacity tightens on spikes, so reserve ahead |
| CoreWeave | Large distributed training (100+ GPUs) | ~$2.70–$3.50/hr (from 8-GPU node) | Enterprise contracts and reserved capacity, not casual rentals |
| Vast.ai | The lowest price, if reliability can flex | from ~$1.87/hr (marketplace) | Reliability varies host to host; an instance can vanish mid-job |
| Modal | Serverless inference and bursty batch jobs | ~$3.95/hr equiv (per-second) | Scales to zero, but wrong tool for big multi-node training |
Prices are on-demand, single-GPU rates as of mid-2026 and move constantly. Treat them as ballpark, not a quote — check each provider's live pricing page before you commit. Spot and marketplace rates swing the most.
RunPod — Cheapest Self-Serve
RunPod is where I would send someone who wants cheap compute without a sales call. On-demand H100s run about $2.39/hr, and spot instances dip near $1.99/hr (as of mid-2026), with per-second billing so you pay only while the GPU is actually yours. The catalog is wide — dozens of GPU types from consumer cards up to H100s — and the serverless endpoints are genuinely useful for inference that scales to zero between requests.
The honest limit: a single pod tops out at 8 GPUs, and while RunPod now offers multi-node clusters, InfiniBand is not the default the way it is at CoreWeave. For tinkering, fine-tuning, and inference, that ceiling rarely bites. For a 64-GPU training run, look elsewhere. If you are wiring RunPod into a coding agent that provisions its own compute, the clean API and per-second billing make it a comfortable fit.
Lambda Labs — The Reliable, Transparent One
Lambda Labs is the one I trust for training when I do not want surprises. Its pricing is published openly — H100s land around $2.49–$3.29/hr depending on the config and which tracker you read, with A100s from roughly $1.99/hr — and there are no egress fees, so shuttling checkpoints and datasets around does not quietly inflate the bill.
It is not the cheapest, and capacity can tighten when demand spikes, so reserve ahead for a big run. What you get in return is a research-friendly workflow and reliability that lets you focus on the model instead of the infrastructure. For most individual and small-team training, that trade is worth it. This is a deliberate pick, not the default winner: if rock-bottom price is your only metric, RunPod and Vast.ai undercut it.
CoreWeave — Enterprise and Hyperscale
CoreWeave is built for the other end of the spectrum. It offers bare-metal nodes and Kubernetes-native orchestration, and — the part that matters — InfiniBand interconnect (roughly 400–800 Gb/s) across GPU clusters, which is what keeps a hundred-GPU job from stalling on gradient sync.
Normalized from an 8-GPU HGX node, H100s work out to roughly $2.70–$3.50/hr per GPU (as of mid-2026), but the real model is reserved capacity and enterprise contracts, not a credit card and a checkbox. If you are running large distributed training with SLAs, this is the serious option. If you are fine-tuning a 7B model on a weekend, it is overkill.
Vast.ai — Cheapest, With Asterisks
Vast.ai is a peer-to-peer marketplace: independent hosts rent out their GPUs and you pick from them. That makes it the cheapest way to get an H100, from around $1.87/hr (as of mid-2026), and lower still on consumer cards.
The trade is reliability. Hosts vary, network and uptime can be uneven, and an instance can disappear mid-job. Vast surfaces reliability scores for a reason — stick to hosts rated 4.5+/5.0. For experiments, batch jobs that can retry, and cost-sensitive tinkering, it is hard to beat. For a production endpoint that has to stay up, it is the wrong tool.
Modal — Serverless for Inference and Jobs
Modal is not a raw GPU rental at all — it is serverless Python. You write a function, decorate it, and Modal runs it on a GPU, billing per second and scaling to zero when idle. H100 time is about $0.0011/sec (~$3.95/hr equivalent) and A100 80GB about $0.0007/sec (as of mid-2026).
The hourly number looks high next to a raw instance, but that comparison misses the point. If your traffic is bursty — an inference endpoint that is quiet at 3am, a batch job that runs twice a day — you pay for seconds of compute instead of a GPU that sits idle. For steady, hours-long training a dedicated instance is cheaper; for event-driven inference and jobs, Modal often wins on total cost. It also slots neatly into an agent-native pipeline where code, not a person, decides when to run something. (Paperspace and TensorDock are reasonable secondary options in this same tier if you want to shop around.)
Picks by Job and Budget
Just tinkering or learning. Start on Vast.ai or RunPod. Rent an A100 or a single H100 by the hour, run a fine-tune or a notebook, and shut it down. If you are mostly running inference to see what a model can do, an A100 at roughly $1–$2/hr is plenty — and the models themselves are free, so start with the best free open-weight models before you pay for anything bigger.
Serving a model / inference. Modal. Per-second billing and scale-to-zero mean a quiet endpoint costs almost nothing between requests. RunPod serverless is a close second if you prefer container-level control over your runtime.
Distributed training (dozens to hundreds of GPUs). CoreWeave. Once you cross roughly 16–32 GPUs, InfiniBand interconnect stops being a nice-to-have and starts deciding your throughput. Lambda handles the single-node and small multi-node tier well; CoreWeave owns the top.
Use Each One If…
| Provider | Reach for it when… |
|---|---|
| Use RunPod if | you want cheap compute with a clean UI, per-second billing, and serverless endpoints for inference — and you do not need a large InfiniBand cluster. |
| Use Lambda Labs if | you want prices you can read off a page, no egress fees, and a setup that treats researchers well — and you will pay a little more for that reliability. |
| Use CoreWeave if | you are training across dozens or hundreds of GPUs and interconnect speed is your bottleneck — and you can sign an enterprise contract. |
| Use Vast.ai if | price is the thing that matters most, the job can survive a host disappearing, and you are willing to vet host reliability scores yourself. |
| Use Modal if | you are serving a model or running bursty jobs and want to pay per second and scale to zero instead of babysitting an idle instance. |
One rule cuts across all of them: price the whole job, not the sticker. A GPU that is $0.50/hr cheaper but charges egress, or sits idle between requests, or dies mid-run and forces a restart, is not cheaper. Storage, egress, cold starts, and reliability all land on the same invoice.
FAQ
What is the cheapest cloud GPU for an H100 in 2026?
On paper, peer-to-peer marketplaces win: Vast.ai lists H100s from around $1.87/hr and RunPod spot dips near $1.99/hr (as of mid-2026). The catch is reliability — marketplace hosts vary and a spot instance can be reclaimed mid-job. For throwaway experiments that is fine; for anything you cannot afford to lose, pay a bit more for a dedicated instance from Lambda Labs or RunPod on-demand.
Lambda Labs or RunPod — which should I pick?
RunPod is usually cheaper on-demand and adds serverless endpoints and per-second billing, which suits tinkering and bursty inference. Lambda Labs costs a little more but publishes transparent pricing, charges no egress fees, and has a research-friendly setup that tends to just work for training. Pick RunPod to save money and experiment; pick Lambda when a reliable, predictable bill matters more than the last dollar.
Do I actually need InfiniBand?
Only for large multi-node distributed training. A single node with up to 8 GPUs connected by NVLink covers most fine-tuning and small training runs, which RunPod, Lambda, and Vast.ai all handle. Once a job spans dozens or hundreds of GPUs across nodes, interconnect speed decides how well it scales — that is when CoreWeave's InfiniBand fabric earns its price.
Do these providers charge for data egress?
Lambda Labs and RunPod advertise no egress fees, which matters if you move large datasets or checkpoints in and out. Hyperscalers like AWS and GCP charge roughly $0.08–$0.12 per GB, which can quietly dwarf the GPU bill on data-heavy work. Marketplace and enterprise pricing varies, so check the specific terms before you assume.
Which is best for inference versus training?
For bursty or unpredictable inference, Modal's per-second, scale-to-zero model means you do not pay for idle GPUs. For steady training or fine-tuning, a dedicated instance from Lambda or RunPod is simpler and cheaper per hour of real work. For very large training runs, CoreWeave. Match the billing model to your traffic shape, not just the hourly sticker price.
Are these prices current?
No — treat every number here as a mid-2026 snapshot. GPU cloud pricing moves week to week with supply, and spot and marketplace rates swing the most. Always check the provider's live pricing page before you commit, and price the whole job (storage, egress, idle time), not just the per-hour GPU rate.