Buy If / Skip If
Quick decision guide. PromptQuorum's assessment: these three fit teams that want GPUs inside an existing general-purpose cloud account, not buyers optimizing only for hourly price.
- Use one of these three if:
- Your app, database or Kubernetes cluster already runs on Vultr, DigitalOcean or UpCloud and you want GPUs in the same account
- You want hourly billing with no hardware to manage, and a price list you can read (DigitalOcean)
- You need EU-based infrastructure and want a provider that describes its EU data handling (see the regions section)
- Skip them if:
- You only want the lowest hourly GPU price: compare marketplaces in the best cloud GPU providers overview first
- You need a verified price today for Vultr or UpCloud GPUs: their pricing pages returned an error to PromptQuorum's automated fetch on 2026-11-22, so confirm live prices manually
- You need guaranteed uptime on spot or interruptible GPUs: they can be reclaimed
Scorecard
PromptQuorum assessment, based on each provider's own pricing pages and documentation, not a hands-on lab test of any GPU server. Scores rate fit for LLM inference and fine-tuning on a general-purpose cloud. Price clarity, reliability and support / docs are not rated because PromptQuorum has no data to score them.
Price clarity
- Vultr:
- Not rated
- DigitalOcean:
- Not rated
- UpCloud:
- Not rated
Reliability
- Vultr:
- Not rated
- DigitalOcean:
- Not rated
- UpCloud:
- Not rated
Ease of setup
- Vultr:
- 8/10
- DigitalOcean:
- 8.5/10
- UpCloud:
- 7.5/10
GPU choice
- Vultr:
- 9/10
- DigitalOcean:
- 8/10
- UpCloud:
- 7/10
Support / docs
- Vultr:
- Not rated
- DigitalOcean:
- Not rated
- UpCloud:
- Not rated
PromptQuorum assessment. These scores are editorial judgments from published provider information; PromptQuorum did not benchmark or operate GPU servers at these providers.
Verdict
Choose DigitalOcean if you want GPU pricing you can verify and a simple Droplet workflow; choose Vultr if its documented GPUs (for example NVIDIA GH200, A16 or AMD Instinct) fit your workload; choose UpCloud if a Finland-based provider matters to you. All three are sensible when your other infrastructure is already there.
If the goal is the lowest possible hourly price, marketplaces are the usual first stop for the lowest hourly price. Compare the marketplaces in best cloud GPU providers and the RunPod vs Vast.ai pricing comparison instead.
PromptQuorum could read only DigitalOcean's full price table, so no price comparison is made. That is not a statement that DigitalOcean is cheaper.
How the Three Compare
All three rent GPU servers by the hour inside a general-purpose cloud, so networking, storage and Kubernetes sit next to the GPU. The product shapes differ.
- DigitalOcean (GPU Droplets): virtual servers with 1 or 8 GPUs (1 GPU only for the RTX 6000 Ada, L40S and RTX 4000 Ada). The pricing page lists on-demand, 12-month reserved and spot prices. The product page says GPU Droplets integrate with DigitalOcean Kubernetes. Separate 8-GPU bare-metal servers (H100, H200, MI300X) are sold by contract through DigitalOcean sales.
- Vultr (Cloud GPU): the Vultr docs list NVIDIA A16, A40, A100, L40S, HGX B200 and GH200, plus AMD Instinct MI300X, MI325X and MI355X (AMD ROCm). Vultr also sells bare metal and Kubernetes (VKE) as separate products; their GPU details were not verified.
- UpCloud (GPU Servers): on-demand and spot GPU servers. The UpCloud changelog lists an L40S and B200 spot plan (2026-07-06) and newer additions: H100, B200 and L4 (2026-05-18), RTX Pro 6000 (2026-09-16) and B300 (2026-10-15). Managed Kubernetes has a documented GPU Nodes page.
- Billing differs: Vultr bills GPU instances on actual calendar-month hours (a 730-hour model since 2024-10-01 per Vultr docs); DigitalOcean bills per second with a 5-minute minimum and still bills powered-off GPU Droplets; UpCloud bills spot servers hourly while powered on (see its spot docs).
Pricing
Only prices read from a provider's own page on 2026-11-22 are shown. "Not verified" means the page could not be fetched or listed no price. All prices are USD per hour and change without notice; the DigitalOcean figures are per GPU.
- Sources: DigitalOcean GPU Droplet pricing and Vultr Cloud GPU docs, both checked 2026-11-22. "DO" in the table means DigitalOcean.
- DigitalOcean reserved prices need a 12-month contract and sales contact; its 8-GPU bare-metal servers list no price and are sold through DigitalOcean sales.
- UpCloud: no price is shown here. Check its pricing page directly. UpCloud documents that spot plans cost "significantly" less than on-demand but gives no percentage in its docs.
Provider / GPU | Price per GPU-hour | Status |
|---|---|---|
| DO RTX 4000 Ada (20 GB) | $0.76 on demand | Verified |
| DO RTX 6000 Ada (48 GB) | $1.57 on demand | Verified |
| DO L40S (48 GB) | $1.57 on demand | Verified |
| DO AMD MI300X (192 GB) | $2.59 on demand / $1.91 12-mo | Verified |
| DO AMD MI325X (256 GB) | $3.80 on demand / $2.88 12-mo | Verified |
| DO HGX H100 (80 GB) | $4.41 on demand / $3.26 12-mo | Verified |
| DO HGX H200 (141 GB) | $4.47 on demand / $3.40 12-mo | Verified |
| DO spot B300 / MI355X / MI350X | $6.17 / $4.50 / $3.00 | Verified, interruptible |
| Vultr NVIDIA A16 (16 GB) | From $0.471 per hour | Verified (Vultr docs) |
| Vultr A100, L40S, B200, GH200 | Not verified | Fetch failed 2026-11-22 |
| UpCloud L40S, H100, B200, L4 | Not verified | Fetch failed 2026-11-22 |
Which Models Fit
Match the GPU memory (VRAM) to the model size first; the provider comes second. PromptQuorum's planning rules below follow the repo's VRAM guides (how much VRAM local LLMs need, DeepSeek-R1 distill VRAM cheatsheet); they are estimates, not benchmark results.
- Serving stack: vLLM runs on NVIDIA GPUs; Vultr states its AMD Instinct GPUs work with AMD ROCm. For a quick start on any Linux GPU server see how to install Ollama.
- No tokens-per-second figures are given here because PromptQuorum did not measure these servers.
GPU memory | Listed examples | Typical fit |
|---|---|---|
| 16 GB | Vultr A16 | 7B-8B models at 4-bit |
| 20 GB | DO RTX 4000 Ada | 8B-14B at 4-bit, tight |
| 48 GB | DO L40S, RTX 6000 Ada | 30B-class; 70B at 4-bit is tight |
| 80 GB | DO HGX H100 | 70B at 4-bit with headroom |
| 141-256 GB | DO H200, MI300X, MI325X | Large models, long context |
Use Cases
Best fits are workloads that sit next to other cloud services you already run.
- Private inference endpoint: a single L40S or RTX 6000 Ada Droplet behind your existing app, in the same private network.
- Kubernetes GPU nodes: DigitalOcean documents GPU integration with its Kubernetes service and UpCloud has a GPU Nodes page for Managed Kubernetes. Vultr Kubernetes GPU support was not verified.
- Burst training or fine-tuning: use on-demand or spot capacity, and keep checkpoints on attached storage. Spot servers can be reclaimed, so save often.
- Large-model hosting: 8-GPU H100, H200 or MI300X bare-metal servers from DigitalOcean (contract, sales-priced) for teams with steady load.
- Not a fit: cheapest-possible experimentation. A marketplace is usually cheaper for that.
Regions, GDPR and Billing
Check region, data-processing terms and billing rules before moving real data. Facts below come from provider documentation read on 2026-11-22; this is not legal advice.
- Hardware source: according to their own descriptions, all three operate their own cloud infrastructure; none is a peer-to-peer marketplace of third-party hosts.
- DigitalOcean regions: the GPU Droplet page lists NYC2, TOR1, ATL1, RIC1 and AMS3 (Amsterdam) and says more data centers are coming. Bare-metal GPU servers are listed in NYC1 and AMS3. Availability can vary by GPU, so confirm in the control panel.
- DigitalOcean data processing: its data processing agreement covers the GDPR; for transfers outside the EEA it relies on the EU-U.S. Data Privacy Framework certification and otherwise on Standard Contractual Clauses. The GPU Droplet page itself makes no GDPR statement, and the page says GPU Droplets are SOC 2 compliant and HIPAA-eligible.
- UpCloud: PromptQuorum could not read UpCloud's security and privacy pages on 2026-11-22 (the fetch returned an error), so confirm directly that your chosen GPU location is in the EU before relying on it. Its docs do not state which regions offer spot GPUs.
- Vultr: GPU region availability and GDPR terms were not verified; the Vultr docs link to a separate regions page.
- Billing and cancellation: DigitalOcean bills powered-off GPU Droplets until you destroy them. Vultr bills GPU instances on calendar-month hours. UpCloud spot servers keep billing attached storage and reserved IPs while stopped and give credit for the unused hour only if UpCloud stops the server.
- Spot risk: DigitalOcean spot Droplets can be reclaimed at any time (it aims for at least 2 hours notice, less in emergencies). UpCloud sends a shutdown request and allows 30 seconds, and spot servers are excluded from the standard SLA.
- Refunds and credits: no refund policy was verified for any of the three; check each provider's terms.
Pros & Cons
All three trade the lowest price for a single account that also covers networking, storage and Kubernetes.
- Pros
- DigitalOcean: readable public price table with on-demand, reserved and spot tiers
- Vultr: broad GPU list including AMD Instinct and NVIDIA GH200
- UpCloud: Finland-based provider with on-demand and spot GPU plans
- Hourly billing with no hardware to maintain
- Cons
- Price advantage over GPU marketplaces not verified
- Vultr and UpCloud prices could not be verified from their own pages here
- Spot GPUs can be reclaimed and are not suited to production serving
- Powered-off GPU Droplets still bill until destroyed
Where to Sign Up
Open each provider, check the live GPU price and region, then start with a single small GPU. PromptQuorum has no pricing control: links go to the providers' sites. For EU data questions also read EU cloud GPU and GDPR, DigitalOcean vs AWS for AI companies and data residency and sovereign AI in the EU.
- 1. Choose the model size and note the VRAM it needs.
- 2. Open the provider page and pick a region that matches your data rules.
- 3. Start one on-demand GPU; avoid spot for anything users depend on.
- 4. Install your serving stack, run a test, and destroy the server when done so billing stops.
Alternatives
If these three do not fit, step sideways to a specialist or move up to a hyperscaler.
- Lowest price per hour: RunPod or Vast.ai; see also RunPod vs Vast.ai pricing.
- Hyperscaler comparison: DigitalOcean vs AWS for AI companies.
- EU data rules: EU cloud GPU and GDPR and data residency and sovereign AI.
- Hourly price context: cloud GPU cost per hour and build vs rent calculator.
- Provider overview: best cloud GPU providers.
How We Evaluate
This page is built from provider documentation and pricing pages read on 2026-11-22; PromptQuorum did not run workloads on these servers.
- Verified from provider pages: DigitalOcean GPU Droplet prices, regions listed and bare-metal terms; Vultr GPU model list, A16 starting price and billing rule; UpCloud spot rules and GPU changelog entries.
- Not verified: Vultr and UpCloud pages returned an error to PromptQuorum's automated fetch on 2026-11-22, so their other prices, regions and Vultr GDPR terms are left out rather than guessed.
- PromptQuorum assessment: the scorecard, model-fit planning and recommendations are editorial judgment, not benchmark results.
- Refresh: prices are re-checked monthly; the checked date appears next to every price.
Frequently Asked Questions
How much does a DigitalOcean GPU Droplet cost per hour?
On demand, DigitalOcean listed $0.76 (RTX 4000 Ada, 20 GB), $1.57 (RTX 6000 Ada and L40S, 48 GB), $2.59 (AMD MI300X), $3.80 (AMD MI325X), $4.41 (HGX H100) and $4.47 (HGX H200) per GPU per hour on 2026-11-22. Billing is per second with a 5-minute minimum.
What GPUs does Vultr offer?
The Vultr docs list NVIDIA A16, A40, A100, L40S, HGX B200 and GH200, plus AMD Instinct MI300X, MI325X and MI355X. The documented starting price is $0.471 per hour for an A16 configuration; other Vultr GPU prices were not verified.
What GPUs does UpCloud offer?
UpCloud documents L40S and B200 spot plans and lists newer GPU additions in its changelog: H100, B200 and L4 (2026-05-18), RTX Pro 6000 (2026-09-16) and B300 (2026-10-15). UpCloud prices were not verified on 2026-11-22.
Does UpCloud have spot GPUs?
Yes. UpCloud documents spot GPU Servers for L40S and B200. They are priced below on-demand, can be reclaimed, give a 30-second shutdown window, and are excluded from the standard SLA, so UpCloud advises against them for production.
Does DigitalOcean have EU GPU regions?
The GPU Droplet page lists AMS3 (Amsterdam) among its data centers, and its bare-metal GPU servers are listed in NYC1 and AMS3. Confirm availability for your chosen GPU in the control panel before planning an EU deployment.
What does DigitalOcean's DPA say about GDPR?
DigitalOcean publishes a data processing agreement that references the GDPR and relies on the EU-U.S. Data Privacy Framework or Standard Contractual Clauses for transfers outside the EEA. The GPU Droplet page itself has no GDPR statement. This is not legal advice; ask your data protection officer.
Can I run Kubernetes with GPUs on these providers?
DigitalOcean says GPU Droplets integrate with its Kubernetes service, and UpCloud documents a GPU Nodes page for Managed Kubernetes. Vultr Kubernetes GPU support was not verified here.
Does DigitalOcean offer bare-metal GPU servers?
Yes, 8-GPU servers with NVIDIA H100, H200 or AMD MI300X, listed in NYC1 and AMS3. They are provisioned by contract through DigitalOcean sales, list no price, and are billed through a Paperspace account.
Are these providers cheaper than RunPod or Vast.ai?
PromptQuorum has not verified that. They are general-purpose clouds aimed at teams that want GPUs in the same account as other services; marketplaces are often where buyers look for the lowest hourly rate. Compare live prices on each page.
Which is best for a 70B model?
A 70B model at 4-bit generally needs around 40 GB or more of VRAM, so a 48 GB GPU is tight and an 80 GB or larger GPU gives headroom. DigitalOcean lists the HGX H100 (80 GB) at $4.41 per GPU-hour on demand (checked 2026-11-22).
Does a stopped GPU server still cost money?
On DigitalOcean, powered-off GPU Droplets are still billed until destroyed. On UpCloud, attached storage and reserved IPs keep billing while a spot server is stopped. Destroy servers you no longer need.
Are the scores a lab test?
No. The scores are PromptQuorum assessments from published provider information. PromptQuorum did not benchmark GPU servers at these providers.
