Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM
Overview & Reference

Best Cloud GPU Providers 2026: RunPod, Vast.ai, Clore, Novita, Vultr and More Compared

·12 min readBy · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

For most people running open-weight LLMs, start with RunPod (per-second Pods and Serverless; RTX 4090 at $0.34/hr on Community Cloud and $0.89/hr on Secure Cloud on its pricing page, checked 2026-11-22), use Vast.ai or Clore.ai only for non-sensitive work where marketplace prices matter, and use a data-center provider (DigitalOcean, UpCloud, Cherry Servers) when data location or uptime matters. This is PromptQuorum's assessment from each provider's published pricing and documentation, not a hands-on lab test. Several prices (Vast.ai, Clore.ai, Vultr beyond its entry plan, UpCloud, LumaDock) could not be verified and are marked as such.

Key Takeaways

  • Default pick for LLM work: RunPod, with RTX 4090 at $0.34/hr (Community) or $0.89/hr (Secure) on its pricing page, checked 2026-11-22
  • Vast.ai and Clore.ai are marketplaces with host-supplied hardware; keep sensitive or regulated data off their host-supplied community tiers
  • Novita AI is the pick if you want a model API (Llama 3.1 8B from $0.02 per 1M input tokens) rather than a rented GPU
  • EU data: DigitalOcean (Amsterdam), Cherry Servers (Lithuania, Stockholm, Amsterdam, Frankfurt) and UpCloud (page not verified)
  • CUDO Compute closed its public on-demand platform on 2026-03-31 per its docs; deployments start at 128 GPUs
  • LumaDock and several other prices could not be verified from official pages and are marked "not listed"

The best cloud GPU provider depends on the workload: per-second Pods and Serverless (RunPod), peer-to-peer marketplaces (Vast.ai, Clore.ai), a model API (Novita AI), or data-center GPU servers (Vultr, DigitalOcean, UpCloud, Cherry Servers, Database Mart). This page compares eleven providers on billing, entry price, hardware source and risk, using each provider's own pricing page and documentation, checked 2026-11-22. Where a price could not be verified from an official page, the table says so instead of guessing. Each provider has its own review, linked in the table.

This page contains links to third-party products for reference. PromptQuorum is not enrolled in any affiliate program — these are plain links that earn no commission. Clicking links and your next steps are entirely your own responsibility. Any recommendation on this page reflects our own reading of the vendors' public pages, not a vendor endorsement or a test.

Best Cloud GPU Providers 2026: RunPod, Vast.ai, Clore, Novita, Vultr and More Compared

Cloud GPU Providers Compared

Eleven providers, one row each: type, billing, the cheapest GPU price that could be verified, and where each fits. Entry prices were checked 2026-11-22 on each provider's own pages; "not verified" means the official page could not be read or shows no price. Prices change, so confirm before you deposit.

RunPod

Type:
Cloud, Community tier
Billing:
Per second, prepaid
Entry GPU/hr (checked):
RTX 3090 $0.22 (Community)
Best for:
General LLM work, Serverless

Vast.ai

Type:
Marketplace
Billing:
Per second, prepaid
Entry GPU/hr (checked):
Not verified (site: from $0.14 spot)
Best for:
Cost-sensitive, non-sensitive jobs

Clore.ai

Type:
Crypto marketplace
Billing:
Per rental term, crypto
Entry GPU/hr (checked):
Not verified (live market)
Best for:
Paying in crypto

Novita AI

Type:
Model API + GPUs
Billing:
Per token / per second
Entry GPU/hr (checked):
GPU not verified; API $0.02/1M in
Best for:
Model API, serverless GPUs

Vultr

Type:
Data-center cloud
Billing:
Hourly, 730-hour basis
Entry GPU/hr (checked):
A16 16 GB from $0.471
Best for:
Existing Vultr users

DigitalOcean

Type:
Data-center cloud
Billing:
Per second, 5-min minimum
Entry GPU/hr (checked):
RTX 4000 Ada 20 GB $0.76
Best for:
Production, Amsterdam region

UpCloud

Type:
Data-center cloud
Billing:
Hourly; spot plans
Entry GPU/hr (checked):
Not verified (page blocked)
Best for:
EU-based GPU servers

CUDO Compute

Type:
Cluster provider
Billing:
Contract, request quote
Entry GPU/hr (checked):
No public price
Best for:
Clusters of 128+ GPUs

Cherry Servers

Type:
Dedicated servers
Billing:
Hourly or monthly
Entry GPU/hr (checked):
A2 16 GB from $0.239
Best for:
EU dedicated GPU servers

Database Mart

Type:
GPU VPS / servers
Billing:
Monthly terms
Entry GPU/hr (checked):
No hourly; RTX 5060 VPS $85/mo
Best for:
Fixed monthly cost, US

LumaDock

Type:
GPU VPS (small)
Billing:
Not verified
Entry GPU/hr (checked):
Not verified (page blocked)
Best for:
Not assessed
Check RunPod GPU pricing →product link · disclosedBrowse Vast.ai GPU offers →product link · disclosedBrowse Clore.ai GPU marketplace →product link · disclosedSee Novita AI model API and GPUs →product link · disclosedSee Vultr Cloud GPU plans →product link · disclosedSee DigitalOcean GPU Droplets →product link · disclosedSee UpCloud GPU servers →product link · disclosedRequest CUDO Compute capacity →product link · disclosedSee Cherry Servers GPU servers →product link · disclosedSee Database Mart GPU plans →product link · disclosedSee LumaDock GPU VPS →product link · disclosed

Which Provider for Which Workload

Match the provider to the workload, not to the lowest headline price. Facts below come from provider pages checked 2026-11-22; the picks are PromptQuorum's assessment.

7B–8B inference

Why:
A 7B–8B model at Q4 needs about 5.5–6 GB of VRAM, so a 24 GB card is plenty; RunPod Community RTX 3090 is $0.22/hr. Novita lists Llama 3.1 8B at $0.02 / $0.05 per 1M tokens if you prefer an API.

70B inference

Why:
A 70B model at Q4 needs about 44 GB of VRAM. RunPod Serverless lists 48 GB GPUs at $1.89/hr and A100 80 GB at $2.99/hr; Novita lists Llama 3.3 70B at $0.135 / $0.40 per 1M tokens.

LoRA fine-tune under $1/hour

Pick:
RunPod
Why:
RTX 4090 is $0.34/hr on Community Cloud and RTX 3090 is $0.22/hr. See fine-tuning under $1 per hour for the full cost breakdown. Vast.ai interruptible prices could not be verified.

Serverless endpoint, scale to zero

Why:
RunPod Serverless defaults to 0 active workers and documents a vLLM worker with an OpenAI-compatible API. Novita states its serverless GPUs scale to zero in 30 seconds; neither publishes un-warmed cold-start timings.

EU data residency

Why:
DigitalOcean lists an Amsterdam (AMS3) GPU data center and a GDPR-covering DPA; Cherry Servers lists Lithuania, Stockholm, Amsterdam and Frankfurt. UpCloud's own page could not be read. Background: EU cloud GPU and GDPR.

Pay by crypto

Why:
Clore.ai docs list CLORE, BTC, USDT and USDC deposits. Vast.ai billing docs list crypto via BitPay and Crypto.com, and Cherry Servers lists crypto among 20+ payment methods. Novita does not support crypto.

Always-on production API

Why:
DigitalOcean GPU Droplets are on-demand and billed per second; the page states SOC 2 and HIPAA eligibility; eligibility is not compliance. Database Mart sells fixed monthly GPU plans. Avoid spot or interruptible tiers for always-on serving. Compare with DigitalOcean vs AWS.

Price Reference: RTX 4090, A100 80 GB, H100 80 GB

Only prices read on a provider's own official page on 2026-11-22 are filled in; everything else says "not listed". "Not listed" means the page showed no price for that GPU or could not be read, not that the provider lacks it. Prices are USD list prices, excluding VAT. Hourly prices are USD per GPU per hour unless marked /mo. Marketplaces (Vast.ai, Clore.ai) are not listed because their live prices are rendered dynamically.

  • RunPod Secure Cloud lists the RTX 4090 at $0.89/hr and the H100 SXM at $4.39/hr; the Community prices above are its lower tier. RunPod's Pod list does not state the memory size of its A100 and H100 Pod entries in the fetched data, so the A100 80 GB cell uses the Serverless listing; the H100 cell is the H100 SXM Pod.
  • Database Mart prices are monthly and promotional; its H100 price renews at $2,599.00 per month after the first term.
  • All cells checked 2026-11-22. Prices change often; open the provider page before you deposit.

RunPod

RTX 4090 24 GB:
$0.34/hr (Community)
A100 80 GB:
$2.99/hr (Serverless)
H100 80 GB:
$2.69/hr (SXM, Community)

Vast.ai

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
not listed

Clore.ai

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
not listed

Novita AI

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
not listed

Vultr

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
not listed

DigitalOcean

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
$4.41/hr (HGX H100)

UpCloud

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
not listed

CUDO Compute

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
not listed

Cherry Servers

RTX 4090 24 GB:
not listed
A100 80 GB:
$2.362/hr (waiting list)
H100 80 GB:
not listed

Database Mart

RTX 4090 24 GB:
$409/mo (dedicated)
A100 80 GB:
$764.55/mo (promo)
H100 80 GB:
$1,039.60/mo (renews $2,599)

LumaDock

RTX 4090 24 GB:
not listed
A100 80 GB:
not listed
H100 80 GB:
not listed

Marketplace Risk: Community Hosts vs Data-Center Tiers

On Vast.ai and Clore.ai the hardware belongs to individual hosts, so reliability and security vary by host; data-center tiers and data-center providers give more predictable conditions. This section states only what each provider's documentation says; PromptQuorum has not audited any host.

  • Clore.ai: its docs describe servers "owned and operated by individual hosts around the world". The docs state no security or privacy guarantee for host-side data, say partial compensation for an offline server may be provided but is not guaranteed, and list no formal dispute process. Container data is lost when the order ends.
  • Vast.ai: its security FAQ says individual hosts may have less formal security measures and recommends encrypting sensitive data, not storing credentials in instances and destroying instances when done. Its compliance page says verified hosts are manually tested for non-regulated workloads and describes a Secure Cloud tier of vetted data centers for regulated or enterprise workloads.
  • RunPod: Secure Cloud runs in T3/T4 data centers; Community Cloud uses vetted individual compute providers, and its docs do not describe Community Cloud security.
  • Interruptions: interruptible and spot tiers (Vast.ai, Clore.ai Spot, DigitalOcean and UpCloud spot) can be reclaimed; use them only for fault-tolerant, checkpointed work.
  • Practical rule (PromptQuorum assessment): do not run sensitive, personal or regulated data on community hosts. Use a data-center tier or provider and read the data processing terms first.

Build vs Rent: When Owning a GPU Wins

Renting wins for bursts, experiments and one-off fine-tuning runs; owning tends to win when a GPU runs most hours of most days. Break-even depends on your hourly rate, utilization and electricity cost, so run your own numbers in the build vs rent cost calculator. For the wider cost picture see local LLM vs cloud GPU cost, and for a mixed setup see the hybrid local and cloud LLM strategy. To size either option, use the VRAM guide.

Other Providers, Listed as Text Only

These providers are listed for completeness only; no prices were compared here. Descriptions are one line each; check each vendor directly.

Lambda

What it is:
GPU cloud for AI training and inference
On this page:
Text link only

Hyperstack

What it is:
GPU cloud for AI workloads
On this page:
Text link only

CoreWeave

What it is:
AI-focused cloud with large GPU clusters
On this page:
Text link only

Nebius

What it is:
AI cloud with GPU infrastructure
On this page:
Text link only

Together AI

What it is:
Hosted open-model inference, dedicated endpoints
On this page:
Text link only

Fireworks

What it is:
Hosted open-model inference, on-demand GPUs
On this page:
Text link only

AWS

What it is:
Hyperscale cloud with GPU instances
On this page:
Text link only

Google Cloud

What it is:
Hyperscale cloud with GPU instances
On this page:
Text link only

Azure

What it is:
Hyperscale cloud with GPU instances
On this page:
Text link only

Descriptions are brief orientation only, not a review.

Frequently Asked Questions

What is the cheapest way to rent a GPU for a 7B model?

A 7B model at Q4 needs about 5.5 to 6 GB of VRAM, so a 24 GB card is plenty. The lowest price read on a provider page for this comparison is RunPod Community Cloud: RTX 3090 at $0.22 per hour and RTX 4090 at $0.34 per hour (checked 2026-11-22). Marketplaces such as Vast.ai may be lower, but their live prices could not be verified. If you only need answers, Novita AI lists Llama 3.1 8B at $0.02 input and $0.05 output per 1M tokens.

RunPod vs Vast.ai: which is better?

RunPod is a managed cloud with Pods, Serverless and a documented vLLM worker; Vast.ai is a marketplace where individual hosts set prices and supply hardware, with on-demand, interruptible and reserved tiers. RunPod is simpler to start and has verified prices; Vast.ai can be cheaper but varies by host. See the RunPod vs Vast.ai pricing comparison for the pair in detail.

Is Clore.ai safe?

Clore.ai is a marketplace where, per its docs, servers are owned and operated by individual hosts, payments are crypto-denominated, and no security guarantee for host-side data is stated. Container data is lost when an order ends, and compensation for an offline server is not guaranteed. PromptQuorum has not audited it; keep sensitive data off it and read the Clore.ai review.

Is cloud GPU rental cheaper than buying a GPU?

It depends on utilization. Renting suits bursts, experiments and short fine-tuning runs, because you pay only while the GPU runs. A card used most hours of most days tends to favor owning. Use the build vs rent cost calculator with your own hourly rate and electricity price.

Which cloud GPU provider is best for EU data?

DigitalOcean lists an Amsterdam (AMS3) GPU data center and a data processing agreement covering GDPR; Cherry Servers lists data centers in Lithuania, Stockholm, Amsterdam and Frankfurt. UpCloud's own page could not be read for this comparison, so verify its regions and terms directly. A data center location alone does not make a deployment compliant; see EU cloud GPU and GDPR. This is not legal advice; check the provider's data processing terms and ask your data protection officer.

Which provider supports serverless GPU endpoints that scale to zero?

RunPod Serverless defaults to 0 active workers and documents a vLLM worker with an OpenAI-compatible API. Novita AI states that its serverless GPUs scale to zero in 30 seconds. Neither publishes cold-start timings for a worker that is not pre-warmed, and PromptQuorum did not measure any.

Can I pay for cloud GPUs with crypto?

Clore.ai docs list CLORE, BTC, USDT and USDC deposits. Vast.ai billing docs list crypto via BitPay and Crypto.com, and Cherry Servers lists crypto among its payment methods. Novita AI states cryptocurrency payments are not supported. Always check the network and minimum deposit before sending funds, because a wrong network can be unrecoverable.

Can I run a 70B model on a rented GPU?

Yes, on a card with enough memory. A 70B model at Q4 needs about 44 GB of VRAM, so use a 48 GB card (tight) or an 80 GB A100 or H100. RunPod Serverless lists 48 GB GPUs at $1.89 per hour and A100 80 GB at $2.99 per hour as checked 2026-11-22. Speed was not measured for this page.

Why are some prices marked "not listed" or "not verified"?

Several provider pages render live prices with JavaScript or block automated reading, and this page only prints a price when it was read on the provider's own official page on 2026-11-22. "Not listed" does not mean the provider lacks that GPU. Open the provider page for current rates.

Does CUDO Compute offer on-demand GPU rental?

Not publicly. Its docs state the public on-demand platform was closed on March 31, 2026, and its cluster page says deployments begin at 128 GPUs with pricing on request. It is included here for clusters, not for hourly rental.

How are the picks on this page chosen?

Picks follow the facts published on each provider's own pages and PromptQuorum's assessment; providers such as Lambda and the hyperscalers are covered the same way, as text only.

How We Evaluate

This comparison was not hands-on tested by PromptQuorum. PromptQuorum did not rent GPUs, run benchmarks or measure speed, cold starts or reliability for this page. It is built from each provider's own pricing pages and documentation, fetched 2026-11-22, and from the six provider reviews linked above.

  • Provider-stated: prices, billing granularity, regions, payment methods and security statements, each taken from the provider's own official pages. Source URLs are listed in each provider review.
  • PromptQuorum assessment: the picks, the workload matrix and the risk guidance, applied to the provider-stated facts. VRAM sizes come from PromptQuorum's own cheatsheets.
  • Not verified: live marketplace prices (Vast.ai, Clore.ai), Novita AI GPU hourly rates, Vultr prices beyond its entry plan, any UpCloud price and region claim from its own page, and all LumaDock details; those pages were blocked or render prices dynamically.
  • Basis of the picks: provider-stated facts checked on 2026-11-22 and PromptQuorum's assessment, not hands-on tests. The disclosure shown on this page explains how its links are handled.
  • Criteria used: billing model, verified entry price, hardware source, documented regions and risk.

← Back to Power Local LLM

MCP