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

GPU Server Hosting 2026: CUDO Compute, Cherry Servers, Database Mart (and LumaDock GPU VPS)

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

For a fixed-term GPU server in 2026, the two hosts in this comparison whose readable official pages publish prices and allow online ordering are Cherry Servers (A100 80GB from $2.362/hr or $1,657.68/mo, mostly waiting-listed; base plans in Lithuania, GPU location not stated) and Database Mart (H100 80GB at $1,039.60/mo, renewing at $2,599.00/mo, USA). CUDO Compute states that its public on-demand platform closed on March 31, 2026; it now sells clusters starting at 128 GPUs through a request form. LumaDock sells GPU VPS plans, but its official pages returned HTTP 403 to automated checks, so nothing about its plans is verified here. Prices are from the providers' own pages, checked 2026-11-22, and change.

Key Takeaways

  • CUDO Compute: docs state the public on-demand platform closed on March 31, 2026; clusters begin at 128 GPUs through a request form, no prices published
  • Cherry Servers: GPU "from" rates run from $0.139/hr to $2.362/hr (A100 80GB); most models are on a waiting list; 15-day money-back stated for dedicated servers
  • Database Mart: 36 GPU plans, USA only, monthly pricing with 1, 3, 12 and 24-month terms; the H100 price renews at $2,599.00/mo after the first term
  • LumaDock: plans and prices not verified because its official pages returned HTTP 403 to automated checks
  • A 70B model at Q4 needs about 44 GB of VRAM per this site's VRAM cheat sheet, so a 48 GB or 80 GB card fits it
  • Prices were checked on 2026-11-22 on each official page and change often

GPU server hosting means renting a whole GPU machine (dedicated, bare metal or a GPU VPS) on a fixed term, instead of paying by the hour on one of the best cloud GPU providers. This page compares four hosts from their own official pages, checked 2026-11-22: CUDO Compute, Cherry Servers, Database Mart and LumaDock. The result is uneven. CUDO Compute states that its public on-demand platform closed on March 31, 2026 and now sells GPU clusters through a sales form. Cherry Servers and Database Mart publish prices and let you order. The official LumaDock pages returned HTTP 403 to automated checks, so its GPU VPS plans are not verified here. All scores are PromptQuorum assessments, not a hands-on lab test.

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.

GPU Server Hosting 2026: CUDO Compute, Cherry Servers, Database Mart (and LumaDock GPU VPS)

GPU Server Hosting Scorecard

PromptQuorum assessment from each provider's official pages, checked 2026-11-22. This is not a hands-on lab test; PromptQuorum did not rent or benchmark these servers. Scores weigh fit for someone who wants to run an LLM on a rented GPU server, so a cluster-only seller scores low on self-service criteria. LumaDock is not scored because its official pages could not be read.

Price transparency

CUDO:
2/10, none published
Cherry:
9/10, GPU table
Database Mart:
9/10, per-plan prices

Reliability evidence

CUDO:
Not assessed
Cherry:
Not assessed
Database Mart:
Not assessed

Ease of setup

CUDO:
2/10, sales form
Cherry:
6/10, waiting lists
Database Mart:
8/10, order online

GPU choice

CUDO:
9/10, Blackwell clusters
Cherry:
5/10, older cards
Database Mart:
9/10, to H100

Contract flexibility

CUDO:
3/10, 128 GPUs+
Cherry:
9/10, hourly
Database Mart:
6/10, monthly terms

Support / docs

CUDO:
Not assessed
Cherry:
6/10, FAQ only read
Database Mart:
Not assessed

Overall

CUDO:
4.0/10
Cherry:
7.0/10
Database Mart:
8.0/10

PromptQuorum assessment, not a hands-on lab test. Reliability (uptime, network speed, real-world throughput) was not measured, so it carries no score. Overall is the average of the scored criteria, rounded: four for CUDO and Database Mart (Support / docs is not assessed), five for Cherry.

Verdict: Who Should Use Which Host

Pick Cherry Servers for hourly or fixed-term dedicated servers with optional GPUs, from a provider with EU sites (which site hosts which GPU is not stated), and Database Mart for the widest published GPU range of the three hosts whose pages could be read, on monthly terms in the USA. Both publish prices and let a developer order without a sales call. Cherry Servers offers flexible billing and a stated 15-day money-back guarantee on dedicated servers, but its A100, A40, A16 and A10 listings show a waiting list. Database Mart lists cards up to the H100 and RTX Pro 6000 96 GB, but its H100 price is a one-time discount that renews at $2,599.00 per month.

CUDO Compute is not a fit for an individual or small team in 2026. Its docs state that the public on-demand platform closed on March 31, 2026, and its cluster page says deployments begin at 128 GPUs. If that is your scale, it is worth a sales conversation, with ISO 27001 and GDPR-aligned operations stated on its cluster page, not verified by PromptQuorum.

LumaDock should be treated as unverified, so check it yourself. PromptQuorum could not read its official pages, so this page makes no claim about its GPU VPS plans, prices or terms.

How GPU Server Hosting Works

A GPU server host gives you a machine for a fixed period instead of an hourly slice of a shared pool. The trade is flexibility for a lower flat rate. For the hourly side of the comparison, see the cloud GPU cost per hour table.

  • Dedicated server: a whole physical machine for you, no virtualization. Cherry Servers describes each server as a private machine with full hardware for one user.
  • GPU VPS: a virtual machine with a GPU attached. Database Mart lists eight GPU VPS plans, from an RTX 5060 8 GB to an RTX Pro 6000 96 GB.
  • Cluster: many GPUs wired together and run by the provider. CUDO Compute sells this, starting at 128 GPUs.
  • Billing model: Cherry Servers lists hourly, fixed-term and spot billing. Database Mart lists 1, 3, 12 and 24-month terms. Check the minimum term per plan: some Database Mart VPS plans are offered only on 3, 12 and 24-month terms.
  • Delivery: Cherry Servers states deployment takes 24 to 72 hours depending on region. Database Mart advertises a 15-minute fastest deploy in general, with no per-plan times.

Pricing: What Each Host Published on 2026-11-22

Every number below comes from the provider's own page, checked 2026-11-22, in USD. Cherry Servers rates are "from" prices; Database Mart prices include sale prices that can end. For per-hour rental comparisons across other hosts, see RunPod vs Vast.ai pricing and the build-versus-rent calculator.

  • Cherry Servers also lists base servers where a GPU is optional: an AMD Ryzen 7700X plan at $0.318/hr or $158.13/mo (1 GPU maximum, 64 GB RAM) and an AMD EPYC 7402P plan that takes up to 2 GPUs.
  • Database Mart also lists multi-GPU plans, for example 2x RTX 4090 at $729/mo, 2x RTX 5090 at $859/mo and 4x A100 40GB at $1,899/mo.
  • The Database Mart H100 price is labelled a one-time discount: the page states it renews at $2,599.00 per month.
  • LumaDock: no price is given because the official pages could not be fetched; check lumadock.com directly.
  • CUDO Compute lists no prices; an older CUDO blog post from before the platform closed is not used as a current price.

CUDO Compute

GPU and VRAM:
Blackwell clusters
Price:
Not published, request form

Cherry Servers A100

GPU and VRAM:
A100 80GB HBM2
Price:
$2.362/hr, $1,657.68/mo, waiting list

Cherry Servers A40

GPU and VRAM:
A40 48GB GDDR6
Price:
$0.801/hr, $472.66/mo, waiting list

Cherry Servers A10

GPU and VRAM:
A10 24GB GDDR6
Price:
$0.518/hr, $303.03/mo, waiting list

Cherry Servers A2

GPU and VRAM:
A2 16GB GDDR6
Price:
$0.239/hr, $139.23/mo

Database Mart GPU VPS

GPU and VRAM:
RTX 5090 32GB
Price:
$419/mo

Database Mart GPU VPS

GPU and VRAM:
RTX Pro 6000 96GB
Price:
$649/mo, 3/12/24-month terms

Database Mart dedicated

GPU and VRAM:
RTX 4090 24GB
Price:
$409/mo

Database Mart dedicated

GPU and VRAM:
A100 80GB
Price:
$764.55/mo (was $1,699.00)

Database Mart dedicated

GPU and VRAM:
H100 80GB
Price:
$1,039.60/mo, renews $2,599.00

LumaDock GPU VPS

GPU and VRAM:
Not verified
Price:
Not verified

Which Models Fit These GPUs

VRAM decides the model size, and this site's VRAM cheat sheet puts DeepSeek-R1 distills at about 9.5 GB (14B), 20.5 GB (32B) and 44 GB (70B) at the Q4_K_M quantization. Those figures are weights plus a small allowance; a long context needs more. For the general method, see how much VRAM a local LLM needs.

  • PromptQuorum assessment from VRAM arithmetic, not a measured benchmark: no tokens-per-second figure is given because PromptQuorum did not test these servers.
  • A rented server runs the same software as a local machine. To try a model first, follow how to install Ollama; for a shared API server, see vLLM explained.
  • Two 24 GB cards do not behave like one 48 GB card: the model must be split across GPUs, which adds setup work.

A2 (16 GB)

Fits at Q4_K_M:
Up to 14B (~9.5 GB)
Host listing it:
Cherry Servers

RTX 4090 (24 GB)

Fits at Q4_K_M:
32B, tight (~20.5 GB)
Host listing it:
Database Mart

RTX 5090 (32 GB)

Fits at Q4_K_M:
32B with context headroom
Host listing it:
Database Mart

A40 / RTX A6000 (48 GB)

Fits at Q4_K_M:
70B, tight (~44 GB)
Host listing it:
Cherry, Database Mart

A100 / H100 (80 GB)

Fits at Q4_K_M:
70B with headroom
Host listing it:
Cherry, Database Mart

RTX Pro 6000 (96 GB)

Fits at Q4_K_M:
70B at higher precision
Host listing it:
Database Mart

Use Cases

A fixed-term GPU server fits workloads that run for long periods; a short experiment is cheaper per hour elsewhere.

  • Always-on inference API: a 14B to 70B model serving an internal tool all day, where a flat monthly rate is predictable.
  • Long fine-tuning runs: multi-day jobs on a single 48 GB to 96 GB card, where hourly billing would also add up.
  • Private GPU for sensitive tests: a whole machine for one customer, as Cherry Servers describes its dedicated servers (not a data-protection guarantee, see Risk and Trust).
  • Large cluster training: the one case where CUDO Compute applies, at 128 GPUs or more through its sales process.
  • Not a fit: a few hours of GPU time, or a burst of parallel jobs; hourly providers suit those better.

Risk and Trust

Facts only, from the providers' own pages checked 2026-11-22. PromptQuorum did not read every terms page and did not test uptime or support.

  • Billing and refunds, Cherry Servers: hourly, fixed-term and spot billing; a 15-day money-back guarantee on dedicated servers, with no conditions stated on the pages read; 20+ payment methods including crypto. The GPU page itself states no refund guarantee.
  • Billing and refunds, Database Mart: monthly price per plan with 1, 3, 12 and 24-month terms; no refund policy appears on the plan pages read; terms are in its Terms of Service and SLA, which were not read.
  • Billing, CUDO Compute: no prices or contract terms are published; the cluster page names an SLA but its contents were not seen.
  • Where the hardware comes from: CUDO Compute states capacity comes from independently operated data centers; Cherry Servers lists Lithuania, Stockholm, Amsterdam, Frankfurt, Chicago, Singapore and Tokyo; Database Mart states US Tier-2 data centers without naming cities.
  • Data handling, CUDO Compute: states ISO 27001 certified facilities, GDPR-aligned operations and sovereign data residency enforcement on its cluster page, as vendor self-claims not verified by PromptQuorum.
  • Data handling, Cherry Servers: states each server is a private machine with no virtualization, default DDoS protection and 100 GB of free backup; no data retention policy appears on the GPU page.
  • Data handling, Database Mart: no data-handling statement appears on the pages read.
  • LumaDock: unverified; nothing about its data handling, locations or refunds is verified here. Read its terms yourself before moving data to it.
  • Do not put regulated or personal data on any host until you have read its privacy policy and data processing terms.

Pros & Cons

Fixed-term hosting trades flexibility for a flat price.

  • Pros
  • Cherry Servers: hourly or fixed-term billing, EU sites listed (GPU location not stated), 15-day money-back stated for dedicated servers
  • Database Mart: 36 published GPU plans from RTX 5060 to H100, with online ordering
  • Dedicated servers: Cherry Servers describes each as a private machine for one customer
  • Flat monthly cost suits an always-on model
  • Cons
  • Cherry Servers: most listed GPU models show a waiting list, and delivery is 24 to 72 hours
  • Database Mart: US locations only, and the H100 discount renews at $2,599.00 per month
  • CUDO Compute: public on-demand platform closed on March 31, 2026; clusters start at 128 GPUs
  • LumaDock: plans not verified
  • Paying for idle hours if the GPU is not busy

Where to Sign Up

Check the live price and the minimum term on the provider's own page before ordering.

  • 1. Pick the model size you need, then the VRAM class from the fit table above.
  • 2. Open the provider page and check the exact GPU, location and minimum term.
  • 3. Cherry Servers: GPUs marked waiting list need you to join it; "Configure Server" models can be ordered.
  • 4. Database Mart: pick a plan and billing cycle; check whether a sale price renews higher.
  • 5. CUDO Compute: fill the request form on the cluster page; expect a sales conversation.
  • 6. LumaDock: open the site and verify the plan, price, location and refund terms yourself.
Check CUDO Compute clusters →product link · disclosedCheck Cherry Servers GPU pricing →product link · disclosedCheck Database Mart GPU plans →product link · disclosedVisit LumaDock (unverified) →product link · disclosed

Alternatives to Consider

If a fixed term is the wrong fit, there are hourly options, bigger clouds and buying hardware.

How We Evaluate

PromptQuorum did not rent, benchmark or hands-on test any server on this page. Everything is read from the providers' own official pages on 2026-11-22.

  • Provider-stated: GPU models, prices, billing terms, locations, refund guarantees and security statements, read from Cherry Servers, Database Mart and CUDO Compute.
  • Not verified: LumaDock, whose official pages returned HTTP 403 to automated checks, and any terms of service or SLA pages that were not read.
  • PromptQuorum assessment: the scorecard, the VRAM fit table (arithmetic from this site's own cheat sheet) and the buy or skip guidance.
  • Not measured: uptime, network speed, support response time and tokens per second.

Frequently Asked Questions

Does CUDO Compute still rent GPUs by the hour?

CUDO Compute's documentation states that no public on-demand rental is offered and that its public on-demand platform closed on March 31, 2026, and its cluster page says deployments begin at 128 GPUs through a request form (checked 2026-11-22).

What does a Cherry Servers GPU server cost?

Cherry Servers lists "from" rates of $2.362/hr ($1,657.68/mo) for an A100 80GB, $0.801/hr for an A40 48GB, $0.518/hr for an A10 24GB and $0.239/hr for an A2 16GB (checked 2026-11-22). Several models show a waiting list.

Where are Cherry Servers data centers?

Cherry Servers lists Lithuania (Šiauliai), Stockholm, Amsterdam, Frankfurt, Chicago, Singapore and Tokyo. The GPU page shows Lithuania for its base server plans and does not say which GPU is offered where.

Does Cherry Servers have a money-back guarantee?

Its homepage and pricing page list a 15-day money-back guarantee for dedicated servers, with no conditions stated on the pages read. The GPU server page itself does not state a refund guarantee.

How much does an H100 cost at Database Mart?

Database Mart lists an Enterprise Dedicated H100 80GB at $1,039.60 per month, labelled a one-time discount that renews at $2,599.00 per month (checked 2026-11-22).

Where are Database Mart GPU servers located?

All 36 listed plans are in the USA. The homepage describes US Tier-2 data centers without naming cities.

Can a rented GPU server run a 70B model?

A 70B model at Q4_K_M needs about 44 GB of VRAM according to this site's VRAM cheat sheet, so a 48 GB card is tight and an 80 GB card (A100, H100) fits with headroom. This is VRAM arithmetic, not a speed measurement.

What could be verified about LumaDock's GPU VPS plans?

Nothing yet. Its official pages returned HTTP 403 to automated checks, so PromptQuorum makes no claim about its plans, prices, locations or terms. Read its terms and refund policy yourself.

Is a GPU server cheaper than hourly cloud GPUs?

Only if the GPU is busy most of the day. A flat monthly price wastes money on idle hours, while hourly billing wastes it on long-running jobs. Compare both using the cost per hour table and the build-versus-rent calculator linked on this page.

Where do the prices and scores on this page come from?

Prices come from each provider's own pages checked 2026-11-22 (or are marked not verified), and the scores are PromptQuorum's assessment. The disclosure shown on this page explains how its links are handled.

← Back to Power Local LLM

MCP