Best GPU Cloud for AI Models (2026): Prices Compared

Disclosure: CodeToolPilot may earn a commission if you buy through links on this page, at no extra cost to you. How we make money.

This is an editorial research buyer’s guide based on official documentation, pricing pages and the sources listed below — how we review. Prices and plans change often; confirm on the vendor’s site.

Status update: Paperspace now operates as part of DigitalOcean. DigitalOcean changed GPU Droplet pricing effective August 1, 2026.

Running your own AI models, whether that is an open-weight LLM behind a private API, a fine-tuning job, or an image model for a product feature, means renting GPUs. The market has split into very different kinds of provider: GPU marketplaces, developer clouds with on-demand VMs, serverless GPU platforms that bill by the second, and large AI clouds that sell mainly through sales teams. Prices for the same GPU vary by a factor of two or more between them.

This roundup is for developers, ML engineers and small teams who want to self-host or fine-tune models without signing a long contract. It is part of our developer hosting hub, and it focuses on providers you can sign up for and use straight away, with one sales-led AI cloud included for comparison. If you only need CPU servers next to your GPU work, our DigitalOcean, Vultr and Hetzner comparison covers general-purpose VPSs.

This is an editorial research article. It is based on official pricing pages, documentation and company announcements checked on October 17, 2026, plus a snapshot of a public marketplace. We did not run benchmarks or training jobs, so there are no throughput, latency or reliability claims here. GPU prices change constantly, sometimes daily; treat every figure as a dated snapshot.

Quick picks

Runpod — best for flexible, low-commitment GPU pods, with a lower-priced Community Cloud tier and consumer RTX cards.
Lambda — best for straightforward on-demand H100, B200 and A100 instances from a GPU-focused cloud.
Vast.ai — best for bargain-hunting on a marketplace where hosts set prices, if you can tolerate variable hosts.
Modal — best for serverless inference and batch jobs that scale to zero, written in Python.
DigitalOcean GPU Droplets — best for teams already on DigitalOcean who want GPUs in the same account.
CoreWeave — best for large, committed multi-node capacity; not self-serve.

How we chose

We did not lab-test these providers. We selected them from the GPU clouds developers most often shortlist for self-hosted models and compared them on documented criteria: whether you can sign up and launch GPUs without talking to sales, which GPUs are listed for on-demand use, the published on-demand price per GPU-hour, billing granularity, storage and egress charges, and stated limits such as preemption, minimum sizes or capacity constraints. All prices are in USD, on-demand, as listed on October 17, 2026, unless noted.

The picks

Runpod

What it is: A GPU cloud built around “Pods” (on-demand GPU containers), plus serverless endpoints and instant clusters. Pods run in two tiers: Secure Cloud in vetted data centers, and Community Cloud on third-party hosts at lower prices. In June 2026 Runpod announced a $100 million growth round led by Summit Partners.

Best for: Developers who want to spin up a single GPU quickly, try several GPU types, or run consumer cards such as the RTX 4090 and RTX 5090 for inference.

Starting prices: H100 SXM at $3.99 per hour on Secure Cloud or $2.69 on Community Cloud; H100 PCIe at $2.89 / $1.99; A100 SXM 80 GB at $1.79 / $1.39; L40S at $1.09 / $0.79; RTX 4090 at $0.89 / $0.34. Runpod’s advertised “H100 from $1.99” refers to Community Cloud PCIe pricing. Pods are billed by the second, and Runpod’s docs state there are “no fees for data ingress or egress”. Volume storage costs $0.10 per GB-month while a Pod runs and $0.20 when it is idle.

Notable limits: You need at least one hour’s worth of credit to deploy an on-demand Pod. Community Cloud hosts are third-party machines, so in our assessment they suit experiments and stateless inference better than sensitive data. Larger clusters and reserved capacity go through sales. Availability of specific GPU types fluctuates.

Lambda

What it is: A GPU cloud from Lambda (formerly Lambda Labs) offering on-demand instances with 1, 2, 4 or 8 GPUs, plus “1-Click Clusters” of 16 to 2,000+ GPUs booked for two weeks to a year, and larger reserved Superclusters.

Best for: Teams that want data-center GPUs on demand with a simple per-GPU price and preinstalled ML tooling.

Starting prices: Per GPU-hour, plus applicable sales tax or VAT: H100 SXM at $3.99 in an 8-GPU instance or $4.29 as a single GPU; H100 PCIe at $3.29 (single GPU); B200 at $6.69 (8x) to $6.99 (1x); A100 SXM 80 GB at $2.79 (8x only); A100 40 GB at $1.99; GH200 at $2.29. The lowest-priced listed option is a Quadro RTX 6000 at $0.69. Lambda’s docs say instances are billed in one-minute increments and that there are no ingress or egress charges.

Notable limits: Lambda describes on-demand as “self-serve, first-come access”, so popular configurations can be unavailable when you want them. We did not see H200, L40S or consumer RTX cards in the on-demand table. 1-Click Clusters cost more per GPU-hour (for example $6.16 for H100 at 16 GPUs) and are booked through Lambda’s team.

Vast.ai

What it is: A GPU marketplace. Hosts, from individuals to data-center partners, list machines and set their own prices for GPU time, storage and bandwidth. Vast says “Prices are set by the market, not by Vast.” You can rent on-demand (fixed price, not interrupted), interruptible (bid-based, which Vast says is “50%+ cheaper” and may be reclaimed), or reserved for longer terms.

Best for: Cost-sensitive experiments, batch jobs and hobby inference where the lowest hourly price matters more than consistency.

Starting prices: There is no fixed price list. In a snapshot of on-demand offers from Vast’s public offer search on October 17, 2026, RTX 4090 listings ranged from about $0.14 to $1.34 per GPU-hour (median around $0.51), H100 SXM from about $2.13 (median around $2.40), A100 SXM4 from about $0.43 (median around $0.80) and L40S from about $0.43 (median around $0.64). These figures moved within minutes and covered a limited sample, so check live prices yourself. Billing is per second, and storage is charged while an instance exists, even if stopped.

Notable limits: Hardware quality, network speed and host reliability vary by listing. Interruptible instances can be paused. In our assessment, avoid putting sensitive data on non-data-center hosts.

Modal

What it is: A serverless platform where you define GPU functions, containers or sandboxes in Python, and Modal runs them on demand, scaling to zero when idle. Containers can use up to 8 GPUs for most GPU types.

Best for: Bursty inference APIs, batch processing and scheduled jobs where you do not want to pay for idle GPUs or manage servers.

Starting prices: Modal prices GPUs per second. Converted to hourly rates: H100 at about $3.95, H200 about $4.54, B200 about $6.25, A100 80 GB about $2.50, L40S about $1.95, L4 about $0.80 and T4 about $0.59. CPU and memory are billed on top. The Starter plan is $0 a month plus compute and includes $30 of free credit each month; the Team plan is $250 a month with $100 of credit.

Notable limits: Modal’s pricing page lists multipliers: choosing specific regions costs 1.15–1.75x base prices, and non-preemptible execution costs 3x base prices. In our assessment, check which multipliers your workload needs before comparing Modal’s base rates with other providers. Egress is billed at $0.04 per GiB. Starter is limited to 10 concurrent GPUs. There are no consumer RTX cards.

DigitalOcean GPU Droplets

What it is: GPU virtual machines with 1 or 8 GPUs inside DigitalOcean’s general cloud, alongside Droplets, managed databases and Kubernetes. DigitalOcean also owns Paperspace; Paperspace’s site now says it “is now part of DigitalOcean”, and its products are documented in DigitalOcean’s docs.

Best for: Teams that already host their app on DigitalOcean and want GPUs in the same account, network and bill.

Starting prices: On-demand per GPU-hour: HGX H100 at $4.41, HGX H200 at $4.47, AMD MI300X at $2.59, L40S at $1.57 and RTX 4000 Ada at $0.76. 12-month reserved pricing is lower (for example H100 at $3.26). DigitalOcean changed GPU pricing effective August 1, 2026. DigitalOcean’s billing docs say Spot GPU Droplets are in public preview at variable prices.

Notable limits: You keep paying for a powered-off GPU Droplet until you destroy it. DigitalOcean’s pricing FAQ states a 5-minute minimum billing period, while its billing docs describe a 60-second minimum, so expect short runs to carry a minimum charge. GPU Droplets are offered in a subset of data centers, and spot capacity can be reclaimed: the billing docs promise at least two hours’ notice, while the pricing page says DigitalOcean aims for two hours but may give less or none.

CoreWeave

What it is: A Kubernetes-native AI cloud and a public company (Nasdaq: CRWV), selling mostly 8-GPU HGX nodes on-demand, as spot capacity and on committed contracts.

Best for: Companies that need large, multi-node clusters and are ready to negotiate commitments.

Starting prices: North America, per 8-GPU instance-hour: HGX H100 at $49.24 (about $6.16 per GPU), HGX H200 at $50.44 (about $6.31 per GPU), HGX B200 at $68.80 (about $8.60 per GPU), A100 80 GB at $21.60 (about $2.70 per GPU) and L40S at $18.00 (about $2.25 per GPU). Spot prices are much lower (H100 at $19.71 per instance-hour). CoreWeave says committed use can save “up to 60%” and lists egress as free.

Notable limits: CoreWeave’s docs say a new organization is activated “after the CoreWeave Sales team approves your organization”, so this is not a sign-up-and-launch option for GPU capacity. Most SKUs come as 8-GPU nodes.

Also worth checking

Together AI sells on-demand GPU clusters (it lists HGX H100 at $3.99 per GPU-hour) alongside its hosted model APIs. Vultr offers Cloud GPU and bare-metal GPU servers in many regions. If you want a hosted API for an open model rather than raw GPUs, inference APIs can be cheaper than renting a GPU that sits idle.

Comparison table

Provider	Self-serve?	Model	Billing	Egress	Consumer GPUs
Runpod	Yes (prepaid credit)	Pods (Secure / Community), serverless, clusters	Per second	No ingress/egress fees	Yes (RTX 3090, 4090, 5090)
Lambda	Yes for instances	On-demand VMs (1–8 GPUs); 1-Click Clusters	Per minute	No egress charges	No
Vast.ai	Yes (prepaid credit)	Marketplace: on-demand, interruptible, reserved	Per second	Set by each host	Yes
Modal	Yes	Serverless functions and containers	Per second (GPU + CPU + memory)	$0.04/GiB	No
DigitalOcean	Yes for on-demand and spot	GPU VMs (1 or 8 GPUs)	Per second with a minimum charge	Transfer allowance included	No (RTX 4000/6000 Ada workstation cards)
CoreWeave	No (sales approval)	Kubernetes-based nodes, mostly 8 GPUs	Per instance-hour	Free	No

Pricing (as of October 2026)

On-demand price per GPU-hour in USD. “—” means the GPU was not listed for on-demand use on the provider’s pricing page.

GPU	Runpod (Secure / Community)	Lambda	Modal (converted)	DigitalOcean	CoreWeave (per GPU, 8x node)
H100 SXM	$3.99 / $2.69	$3.99 (8x) – $4.29 (1x)	~$3.95	$4.41	~$6.16
H200	$5.29 / $3.59	—	~$4.54	$4.47	~$6.31
B200	$7.99 / $5.98	$6.69 (8x) – $6.99 (1x)	~$6.25	—	~$8.60
A100 80 GB	$1.79 / $1.39 (SXM)	$2.79 (SXM, 8x)	~$2.50	—	~$2.70
L40S	$1.09 / $0.79	—	~$1.95	$1.57	~$2.25
RTX 4090	$0.89 / $0.34	—	—	—	—

Caveats: Lambda prices exclude sales tax/VAT. Modal’s figures are its per-second base rates multiplied by 3,600; they exclude CPU and memory charges and the region and non-preemptible multipliers. CoreWeave lists prices per 8-GPU instance; we divided by eight. Vast.ai is omitted from this table because its prices are set by individual hosts and change continuously (see the snapshot above). Prices can differ by region and change frequently. Check current rates: Runpod pricing, Lambda pricing, Vast.ai pricing, Modal pricing, DigitalOcean GPU Droplets, CoreWeave pricing.

Prices verified against each vendor’s official pricing page (listed under Sources). Plans, promotions and renewal prices change often — confirm on the vendor’s page before you buy.

How to choose

Match the GPU to the model. VRAM is usually the deciding constraint. As a rough rule, a quantized model in the 7–14B-parameter range typically fits on a 24 GB card such as an RTX 4090; larger models or long contexts need 48–80 GB cards (L40S, A100, H100) or several GPUs.
Always-on or bursty? For an API that receives steady traffic, an hourly VM or Pod is simpler to reason about. For sporadic requests, Modal’s scale-to-zero model avoids paying for idle hours, but compare its effective rate after CPU, memory and multipliers.
Price vs predictability. Marketplace and community tiers (Vast.ai, Runpod Community Cloud) offer the lowest listed prices; data-center tiers (Runpod Secure Cloud, Lambda, DigitalOcean) cost more but give more consistent hosts.
Hidden costs. Count storage while stopped, egress, minimum billing periods and the fact that most providers keep billing a stopped instance’s disk.
Scale and commitment. If you need dozens or hundreds of GPUs for weeks, Lambda’s 1-Click Clusters and CoreWeave’s committed contracts are designed for that, at different price points.
How you will use it. If your goal is to point a coding agent at your own model, tools such as Cline and Aider can talk to OpenAI-compatible endpoints; see our Cline review and Aider review for how they handle custom model providers.

For the rest of your stack, such as the web app that calls your model, see our comparison of developer hosting options.

Bottom line

For most developers starting out, in our assessment Runpod and Lambda are the simplest self-serve places to rent data-center GPUs by the hour, with Runpod adding consumer cards and a lower-priced community tier. Vast.ai’s median listings in our October 17 snapshot were below the other providers’ list prices for the GPUs we checked, in exchange for host-to-host variability. Modal is a good fit for scale-to-zero serverless inference written in Python, once you account for its multipliers. DigitalOcean makes sense if your app already lives there. CoreWeave is built for large committed capacity and requires sales approval.

FAQ

What is the lowest-cost way to run an open-weight LLM?

For small models with occasional use, a consumer GPU on a marketplace or community tier is usually the lowest hourly cost, and a hosted inference API may be cheaper still if your traffic is low. For steady production traffic, compare a reserved or committed price with on-demand rates; Lambda, DigitalOcean and CoreWeave all list lower prices for commitments.

Do I pay when my GPU instance is stopped?

Often, yes. DigitalOcean bills GPU Droplets until you destroy them. Runpod charges $0.20 per GB-month for idle volume storage, and Vast.ai charges storage while an instance exists. Delete what you are not using.

Is per-second billing important?

It matters for short jobs and frequent restarts. Runpod, Vast.ai and Modal bill per second, Lambda in one-minute increments, and DigitalOcean per second with a minimum charge. For a GPU that runs all month, the billing increment makes little difference.

Best GPU Cloud Hosting for Running Your Own AI Models