Best VPS providers in 2026: 48 services compared

The best VPS provider in 2026 is not the company with the smallest number on its pricing card. It is the provider whose CPU policy, storage behavior, network limits, transfer accounting, billing model, support boundary, and regions match the workload.

My practical shortlist is simple. Hetzner is the value default for many European workloads. OVHcloud is the clearest mainstream choice when unmetered transfer under an explicit port cap matters. DigitalOcean and Akamai Cloud/Linode are approachable global developer clouds. Civo and Gcore deserve attention when separately billed VM egress would dominate the bill. Clouding.io is the strongest Spain-based option in this comparison. AWS Lightsail packages AWS into a simpler VPS product, while AWS EC2, Google Compute Engine, Microsoft Azure, and Oracle OCI make sense when the surrounding cloud platform matters more than a flat VPS bill.

This guide compares 48 major globally marketed VPS and cloud-VM services, focused on machines around 0.5–64 GB RAM and 1–16 vCPU. It includes unmanaged VPS, developer clouds, hyperscalers, high-transfer specialists, and managed VPS products. It does not pretend to include every local reseller, temporary low-stock offer, white-label host, or custom enterprise contract.

Prices, limits, and official documentation were checked on July 23, 2026. Prices are before tax unless stated otherwise. Regional calculators, currencies, processors, promotions, and available stock can change the result. There are no affiliate links.

Table of contents

The short answer

What you needShortlistWhyMain catch
Low-cost general-purpose VPS in EuropeHetzner, netcup, ContaboAggressive RAM, storage, and transfer for the priceShared CPU, location-specific terms, setup or commitment differences
Predictable high-transfer VPSOVHcloud, Civo, GcoreUnmetered or no separately billed VM egress under published conditionsPort speed, fair use, regional availability, and product-specific limits still matter
Simple global developer cloudDigitalOcean, Akamai Cloud/Linode, VultrClear APIs, images, documentation, hourly billing, broad regionsIncluded transfer and overage differ; compute may continue billing while powered off
Spain-hosted cloud VPSClouding.ioBarcelona infrastructure, Spanish IPs, custom sizing, hourly billingOne primary geography; excess transfer is still metered
European cloud with custom sizingUpCloud, Scaleway, Exoscale, Clouding.ioFlexible infrastructure and European data-location optionsPricing and bandwidth rules differ substantially by region and instance family
Tiny or intermittent development machineClouding.io, AWS Lightsail, Oracle OCISmall entry sizes and/or a meaningful free allowance“Stopped” does not always mean free; disks and IPs can continue billing
High-frequency or compute-sensitive workloadVultr, UpCloud, Gcore, Akamai dedicated CPUDedicated or high-frequency product familiesDo not infer dedicated CPU from a fast processor model or benchmark
Large public egress on a small VMOVHcloud, Hetzner, Civo, Gcore, Oracle OCIBetter included transfer or egress economics than ordinary hyperscaler networkingAllowances, port caps, geography, and abuse policies are not interchangeable
AWS integration without full EC2 complexityAWS LightsailBundled compute, disk, IP, and transferAllowance counts traffic in both directions; only excess outbound is charged
Full hyperscaler ecosystemAWS EC2, Google Compute Engine, Microsoft Azure, Oracle OCIManaged databases, private networking, IAM, observability, global servicesVM price is only one line; egress, disk, IP, NAT, load balancers, and zones can dominate
Managed VPS with a control panelScalaHosting, Liquid Web, InMotion, KnownHostProvider manages more of the OS and hosting stackHigher price; “managed” scope and supported software vary
Managed cloud control planeCloudwaysOperated server layer across several infrastructure choicesYou pay a management premium and still need to size the underlying server
Promotional long-term VPSHostinger, Time4VPS, IONOS, NamecheapLow initial effective rateRenewal, upfront term, setup, panel, backup, and IPv4 details decide the real cost

This is not a universal speed ranking. A short benchmark from a fresh VM cannot tell you how aggressively a shared vCPU is contended next month, whether storage latency degrades under a noisy neighbor, or what support will do during an incident.

How this comparison works

I separated the market into five groups:

  1. Developer clouds: self-service VM platforms with hourly or minute-level billing, APIs, images, snapshots, and block storage.
  2. European and high-transfer providers: often stronger on transfer economics, local data residency, or flat monthly value.
  3. Budget and promotional VPS: attractive headline resources, frequently tied to prepayment, renewal, fair-use, or support tradeoffs.
  4. Hyperscalers: compute is one service inside a much larger system; the surrounding network and managed services are part of the decision.
  5. Managed VPS: the provider takes more responsibility for patching, panels, migrations, or troubleshooting, but “managed” is not a standard specification.

Claims have three layers:

  • Verified facts: current official prices, resource allocations, transfer rules, and documented billing behavior.
  • Accepted technical reality: a shared vCPU is not a dedicated core; port speed is not guaranteed throughput; a snapshot is not an independent backup.
  • Editorial inference: which services appear to fit particular workloads.

Where a provider publishes only a live calculator, regional checkout, or temporary offer, this article says so instead of turning one observed configuration into a universal list price.

The VPS bill is larger than the VPS

The usable monthly cost is:

compute + boot disk + extra storage + public IPv4 + backups + snapshots + outbound transfer + NAT + load balancers + cross-zone traffic + licenses + management + tax

A €4 VM with 10 TB of internet delivery can be cheaper than a €20 VM—or several hundred euros more expensive—depending on how transfer is measured. A stopped instance can still reserve CPU, disk, IPv4, or licensed software. A backup toggle may create snapshots in the same account and region, which is a different risk profile from an encrypted offsite backup with tested restoration.

Normalize every offer to the same questions:

  • What cash is charged today?
  • Is the rate hourly, monthly, annual-effective, introductory, or tied to several years prepaid?
  • Does the vCPU represent shared scheduling, a burst allowance, a dedicated thread, or an entire physical core?
  • Is local disk ephemeral or persistent? Is the advertised disk replicated?
  • What IOPS and throughput are included, capped, or merely “up to”?
  • Is transfer inbound-only, outbound-only, or bidirectional?
  • Is the allowance pooled, prorated, reset monthly, or attached to each VM?
  • What happens after the allowance: charge, throttle, suspend, or manual review?
  • Is the network figure a NIC ceiling, public-port cap, expected range, or guarantee?
  • Does deleting the VM also delete its disks, snapshots, backups, and IPs?

Shared, burstable, and dedicated CPU are different products

The “v” in vCPU matters. Providers can expose one logical hardware thread, a scheduled share of a thread, a burstable baseline, or a dedicated allocation under the same label.

Shared vCPU

Shared plans are economical because multiple tenants can use the same physical CPU capacity. They are sensible for:

  • cached websites;
  • low-duty APIs;
  • VPN endpoints;
  • small monitoring systems;
  • development and test environments;
  • cron jobs with short peaks.

They are a poor default for sustained video encoding, busy databases, game servers with strict tick latency, continuous compilation, or any application whose economics depend on holding 100% CPU for hours. A shared plan can benchmark beautifully when the host is quiet and still be the wrong product.

Burstable CPU

Burstable instances accumulate credits or receive a documented baseline, then run faster while allowance remains. They fit services with low average usage and short spikes. They become dangerous when a production workload silently becomes sustained: performance can fall to the baseline or surplus credits can become a bill.

AWS EC2 T-family and comparable hyperscaler families expose this model explicitly. Many conventional VPS companies do not publish a credit system; “shared” should not be relabeled “burstable” unless the provider documents it.

Dedicated vCPU

A dedicated vCPU product normally promises that the virtual CPU allocation is not competing with unrelated tenants for scheduled CPU time. It does not automatically mean:

  • a whole physical core instead of a hardware thread;
  • a fixed processor generation in every region;
  • bare-metal cache and memory behavior;
  • identical performance across host generations.

Use dedicated CPU when sustained compute or latency predictability justifies the premium. Still benchmark the actual region and instance family.

RAM: allocation is not application headroom

A 1 GB Linux VM does not give an application a clean 1 GB. The kernel, filesystem cache, monitoring agent, SSH, control panel, database, web server, language runtime, and background jobs all consume memory.

Practical lower bounds:

  • 0.5–1 GB: VPN, bastion, tiny static service, monitoring probe, or carefully tuned single-purpose daemon.
  • 2 GB: small web application, low-traffic WordPress, lightweight container host, or development server.
  • 4 GB: comfortable small application server, modest database, or several low-duty containers.
  • 8–16 GB: busier databases, stores, search, CI runners, game servers, or multi-service stacks.
  • 32–64 GB: memory-heavy databases, caches, build systems, analytics, or consolidated workloads.

Swap can prevent an abrupt out-of-memory kill, but storage-backed memory is not RAM. Heavy swapping on network block storage can turn a memory shortage into a latency incident and an I/O bill.

Storage: NVMe is not a performance specification

“NVMe SSD” describes an interface and medium. It does not disclose:

  • local versus network-attached storage;
  • replication;
  • read and write IOPS;
  • sequential throughput;
  • queue-depth behavior;
  • burst duration;
  • noisy-neighbor policy;
  • snapshot pause or consistency behavior;
  • durability and failure domain.

Local disk

Local NVMe can deliver excellent latency and throughput. It may also disappear when a host fails or an instance is moved. Use replication and backups when the provider describes local storage as ephemeral.

Persistent block storage

Network block storage survives VM replacement and can often move between instances. Performance may scale with volume size or provisioned IOPS. The volume can keep billing while detached. For example, DigitalOcean Volumes are charged for as long as the volume exists, attached or not.

Backups and snapshots

A snapshot is a point-in-time storage image. A backup strategy answers different questions:

  • Is the copy in another account, region, or provider?
  • Is it encrypted with a recoverable key?
  • How many versions exist?
  • Can one compromised credential delete the VM and every snapshot?
  • Was application consistency coordinated with the database?
  • Has restoration been timed and tested?

Provider backups are useful. They are not permission to skip an independent recovery copy for important data.

Bandwidth, transfer, and egress: the vocabulary trap

Three different numbers are routinely collapsed into “bandwidth”:

  1. Port speed: the maximum rate the interface or product permits, such as 500 Mbit/s or 1 Gbit/s.
  2. Throughput: the rate the VM actually delivers under current CPU, host, route, protocol, distance, and congestion conditions.
  3. Transfer allowance: the number of GB/TB moved during a billing period before another rule applies.

A 10 Gbit/s port with 1 TB included can consume its monthly allowance in minutes. A 500 Mbit/s unmetered port can deliver roughly 160 TB in a 30-day month at perfect continuous utilization, but real traffic, protocol overhead, provider policy, and end-to-end paths reduce that theoretical figure.

Port cap is not a guarantee

When OVHcloud lists 500 Mbit/s, 1 Gbit/s, or 2 Gbit/s public bandwidth on a VPS, that is an explicit product cap. It is not a contractual promise that every destination will receive that speed.

Hetzner makes the opposite caveat especially clear: cloud-server throughput is not guaranteed. Its technical FAQ gives a typical shared-host-network expectation rather than an SLA. Treat “up to,” “typical,” and “guaranteed” as different words.

Inbound, outbound, and bidirectional accounting

Common models:

  • Outbound-only: inbound traffic is free; only internet egress consumes the allowance. DigitalOcean Droplets use this model.
  • Both directions consume allowance, only excess outbound is billed: AWS Lightsail counts inbound and outbound against the bundle, but does not charge excess inbound.
  • Unmetered under a port cap: OVHcloud VPS publishes unlimited traffic plus plan-specific public bandwidth.
  • No separate VM ingress or egress charge: Civo states data transfer is free and unlimited for its compute pricing.
  • Large included pool: Hetzner Cloud includes a location-dependent traffic allowance.

Always check private traffic too. Same-region traffic over a public IPv4 address can be billed differently from VPC/private traffic. Cross-zone and cross-region traffic can incur charges even when ordinary ingress is free.

GB, GiB, TB, and TiB

Decimal units:

  • 1 GB = 1,000,000,000 bytes
  • 1 TB = 1,000 GB

Binary units:

  • 1 GiB = 1,073,741,824 bytes
  • 1 TiB = 1,024 GiB

10 TB is about 9.095 TiB. 10 TiB is about 10.995 TB. A provider charging per GiB and another advertising TB are not using interchangeable units. This difference matters when calculating thousands of units.

Pooled, prorated, and non-rollover allowances

DigitalOcean pools accrued Droplet transfer at team level and prorates allowance for the time each Droplet exists. Akamai Cloud/Linode also uses transfer pools, with regional exceptions. AWS Lightsail aggregates allowances only across instances of the same bundle in a region.

Deleting and recreating a VM usually does not create free transfer. Unused monthly allowance normally does not roll forward. Read the accounting rules before designing a fleet around pooling.

The 10 TB egress test

This scenario isolates public internet outbound transfer only. It is not a complete infrastructure quote. It excludes compute, disk, IPv4, NAT, load balancers, request charges, taxes, CDN, cross-zone traffic, and negotiated discounts.

The example uses 10 TB decimal, converted to the billing unit each provider publishes. That is 10,000 GB or about 9,313.23 GiB. The result remains approximate because account-level free allowances, prorating, regional rates, and other services can change the invoice.

Provider/productPublished allowance or modelApproximate network-only charge at the stated scenarioInterpretation
Hetzner Cloud EU20 TB included on relevant shared-cloud plans$0 incrementalVerify plan and location; US and Singapore allowances differ
OVHcloud VPSUnlimited traffic under the plan’s public-bandwidth cap$0 incrementalPort cap and acceptable-use/security controls still apply
Civo ComputeFree, unlimited ingress and egress$0 incrementalConfirm the chosen product and region, not a different Civo service
Gcore VMsFree egress on current VM offer$0 incrementalRegion, VM family, active-instance billing, and network profile still matter
Oracle OCIFirst 10 TB/month public egress free$0 at 10 TBOver 10 TB, current North America/Europe/UK list rate is $0.0085/GB
Akamai Cloud/LinodeExample assumes 5 TB included; overage starts at $0.005/GB in core regionsAbout $25 for the extra 5 TBJakarta, São Paulo, and distributed regions use different rates
DigitalOcean DropletsExample assumes 5,000 GiB included; excess $0.01/GiBAbout $43.13Actual pooled allowance is prorated and depends on active plans
AWS LightsailExample assumes a 5 TB allowance; Europe/most US regions $0.09/GB overAbout $450 for the extra 5,000 GBBoth directions consume allowance; only chargeable excess outbound is billed
AWS EC2100 GB free, then $0.09/GB for the first paid 10 TB in US East exampleAbout $891 for 10,000 GBCross-AZ, NAT, and load-balancer charges can add more
Microsoft AzureFirst 100 GB free, then $0.087/GB from Europe/North America in the next tierAbout $861.30 for 10,000 GBRouting tier, source geography, and service path matter
Google Cloud Premium Tier1 GiB free; $0.12/GiB through 1,024 GiB, then $0.11/GiB through 10,240 GiB to Europe/North AmericaAbout $1,034.57 for 10 TB decimalStandard Tier can be materially cheaper; destination and source region matter

The point is not that a hyperscaler is “bad.” AWS EC2, Google Compute Engine, and Azure VMs offer ecosystems that ordinary VPS companies do not reproduce. The point is that a $10 compute line can coexist with a four-figure network line. Architecture and traffic shape decide whether that trade is rational.

Market comparison: provider-by-provider

Developer clouds and general-purpose VM platforms

ProviderCurrent small-VM referenceCPU and storageTransfer and networkBilling traps
DigitalOceanBasic Regular: $4/month, 1 shared vCPU, 512 MiB, 10 GiB SSD, 500 GiB transferBasic is shared; dedicated families cost more. Volumes are separate persistent block storageInbound free; outbound pooled and prorated; $0.01/GiB excess (bandwidth rules)Powered-off Droplets still bill. One IPv4 included. Backups add 20% weekly or 30% daily
Akamai Cloud/LinodeNanode 1 GB: $5/month, 1 shared vCPU, 25 GB SSD, 1 TB transferShared and dedicated CPU families; dedicated suits sustained full-core useInbound and qualifying private traffic free; pooled transfer; core-region excess starts $0.005/GB (rules)Powered-off instances bill. Regional overage and pooling exceptions apply
VultrCloud Compute uses shared CPU; live price matrix must be checkedShared, high-frequency, high-performance, and dedicated VX1/optimized familiesInbound unmetered; outbound only counted; $0.01/GB excessHourly with a monthly cap on ordinary servers; stopped servers bill until destroyed; pricing page is dynamic
UpCloudStarter: $3.50/month, 1 core, 1 GB, 10 GB SSD, 250 Mbit/sStarter uses earlier-generation AMD; Premium and Cloud Native differZero-cost egress inside an account Fair Transfer Limit; warning then possible 100 Mbit/s throttle, not a per-GB billStarter/Premium bill while stopped. Cloud Native compute can pause, but storage/IP continue
ScalewaySTARDUST1-S: about €0.43/month, 1 vCPU, 1 GB, 100 Mbit/s; DEV1-S: about €6.55, 2 vCPU, 2 GBDevelopment families use economical virtual CPU; storage is separate unless statedCurrent Virtual Instance pricing advertises no egress fee; plan bandwidth remains cappedCompute pauses when powered off; volumes and Flexible IPv4 continue billing
CivoStandard xSmall: $5.43/month, 1 core, 1 GB, 25 GB NVMeSeveral standard and optimized families; standard core dedication is not promisedCurrent compute pricing states free, unlimited ingress and egressHourly rounded up to a full hour. Stopped-instance and public-IPv4 billing need dashboard confirmation
ExoscaleStandard Micro: 1 core, 512 MB; current numeric rate requires calculatorVM family, local storage, and snapshots are separate dimensionsEach running instance-hour contributes 1.42 GiB outbound, about 1 TiB/month; pooled at organization levelCompute stops when powered off, but local storage and software licenses continue
KamateraType A: $4/month, 1 vCPU, 1 GB, 20 GB NVMe, 5 TBType A budget, Type B reserved CPU threads, Type T burstable, Type D dedicated5 TB package with $0.01/GB excess; a 50 Mbit/s unmetered option also appearsMinute-granular hourly billing. Powered-off price and IPv4 fee require calculator confirmation
Leaseweb Public CloudCompute rate is live-calculator dependentMulti-tenant compute; local NVMe or redundant network SSD with different IOPS ceilings1 TB pooled outbound; ingress free; excess begins around €2.20/TB and falls by volumePowered-off instance, storage, and IPv4 keep billing until destroyed

European value, high-transfer, and regional specialists

ProviderCurrent small-VM referenceTraffic and throughputBest useCaveat
HetznerCX23: €5.49/month after the June 15 increase; 2 shared vCPU, 4 GB, 40 GB SSD; IPv4 extraEU shared plans include at least 20 TB; US 1 TB; Singapore 0.5 TB. No throughput guarantee; official expectation about 300–500 Mbit/sLow-cost EU web, API, development, and moderate database workloadsShared CPU is not for sustained full-core use. Stopped servers and IPs keep billing
OVHcloudVPS-1 Spain: €3.81/month before VAT; 2 vCores, 4 GB, 40 GB NVMe, one-day daily backupUnlimited traffic; 500 Mbit/s public cap. Larger plans list 1, 2, and 3 Gbit/sPredictable high-transfer workloads where a capped port is acceptable“vCore” is not a dedicated-core promise. Regional price, IP, and backup options differ
Clouding.ioCustom entry: €3/month cap; 0.5 vCore, 1 GB, 5 GB triple-replicated NVMe4 TB included; €0.004/GB or €4/TB excess. Platform says 80 Gbit/s network, not per-VM guaranteeSpain-hosted projects, tiny machines, custom sizing, intermittent developmentArchive removes CPU/RAM cost but disk and backups remain. One fixed Spanish IPv4 included
Cherry ServersCloud VPS 1 Gen 2: $3.51/month; 1 shared vCore, 1 GB, 20 GB SSD1 Gbit/s port, 1 TB egress; ingress unmetered; published excess €1.69/TBLow-cost European/API-driven VPSShared/burstable CPU; provider can reduce bandwidth near/over limits; APAC allowances differ
Aruba CloudVPS O1I1: €1.99/month before VAT; 1 vCPU, 1 GB, 20 GB, 2 TB1 Gbit/s interface; current plans offer 2–100 TB. Older official docs say inbound + outbound countLow-cost EU VPS with a public IPv4 includedOver-limit path is upgrade rather than a published per-GB fee; current and older docs need reconciliation
CloudSigmaPricing is calculator/partner-quote dependentFree-form sizing and consumption billingBuyers needing unusual CPU/RAM ratios or provider-specific arrangementsCurrent public page does not expose safe retail CPU, RAM, egress, IPv4, or backup unit prices
Atlantic.NetG3.2GB: about $10/month; read exact current specs from matrixInbound unlimited; outbound plan allowance; $0.02/GB excessStraightforward hourly public cloud with optional managed servicesExtra-IPv4 official pages conflict ($2.19 versus $3.96/month); confirm checkout
GcoreStandard VM from €8.93/month before VAT in the advertised Luxembourg exampleDedicated vCPU on Standard/optimized families; free VM egress; regional network profile variesSustained compute, European sovereignty, and egress-heavy servicesThe same page contains a conflicting €14.98 entry line. Use regional calculator
HostwindsUnmanaged entry: $4.99/month; 1 CPU, 1 GB, 30 GB SSD, 1 TB1 Gbit/s port; 1–9 TB by plan. Terms say throttle to 3 Mbit/s after capBuyers choosing between unmanaged and provider-managed variantsOverage terms use unusual “$0.01 per Gigabit” wording; confirm unit in writing

Budget and promotional VPS providers

ProviderCurrent offer snapshotTransfer/networkTerm and operational catch
ContaboCloud VPS 10 €4.50/month; 3 vCPU, 8 GB, 75 GB NVMe or 150 GB SSDPricing says 32 TB; support describes no default traffic limit and fair-use throttlingSelf-managed fixed subscription. Treat the traffic wording as an official-documentation ambiguity
netcupVPS 500 G12 €5.91/month including 19% VAT; 2 shared vCores, 4 GB, 128 GB NVMeTraffic included; interface and rolling 24-hour fair-use thresholds vary by productSome hourly products, others 12-month terms. Do not transfer one plan’s throttle rule to another
IONOSVPS S+ $2/month for 3 months, then $5; 2 vCores, 2 GB, 90 GB NVMeUnlimited traffic, up to 1 Gbit/s; one IPv4 and IPv6 /80One-year term; self-managed. Additional IPv4 $5/month. Existing accounts may face a 2026 adjustment
Time4VPSLinux 4 €4.38/month effective for 2 years, renewal €8.76; 2 CPU, 4 GB, 50 GB8 TB; dedicated 100 Mbit/s; €1.33/month 1 Gbit/s add-on; overage reduces port by 10×Standard storage documented at 75 MB/s and 300 IOPS; optional SSD component raises limits
HostingerKVM 1 $6.49/month on 2 years, renewal $11.99; 1 vCPU, 4 GB, 50 GB NVMe4 TB and 1 Gbit/s; one IPv4 plus IPv6Headline is prepaid-term monthly equivalent. Weekly backups and a manual snapshot included
NamecheapSpark $3.88/month billed annually; 1 CPU, 1 GB, 20 GB RAID-10 SSD, 1 TBExtra transfer $10 per 100 GB; one or two IPv4 by planSelf-managed included; management and offsite backups add $10–$25/month on eligible plans
SSD NodesSmall advertises 2 vCPU, 8 GB, 160 GB SSD, 4 TB at $66/yearOutbound allowance; IPv4 can depend on the promotionOfficial footnotes and a May 2026 comparison disagree on term totals. Verify cart, prepayment, IP, and backups
InterServerOne slice $3/month; 1 core, 2 GB, 40 GB SSD, 2 TBAllowance scales by sliceKVM/full root; current IPv4, port, and backup price were not reliably published
BuyVMSlice 512 $24/year; 1 fair-share core, 512 MB, 10 GB, unmetered, one IPv4Unmetered subject to policy; five free snapshots per locationStock is constrained. Nightly seven-day backups cost 10% of base, $0.50 minimum
HostHatchNVMe 2 $4/month; 1 fair-share core, 2 GB, 10 GB, 1 TBBoth inbound and outbound count; private traffic excluded; APAC differs“40 GbE” describes platform fabric, not guaranteed VM throughput
V.PSStarter €6.95/month; 2 Xeon cores, 1 GB, 20 GB, 1 TBBoth directions count; exhaustion can throttle to 1–10 Mbit/s or suspend by productOne IPv4/IPv6. Official enforcement FAQ is dated 2023; reconfirm for critical traffic
GreenCloudVPSRegular entry $6/month; 1 CPU, 1 GB, 15 GB, 1.5 TB. Stock sales can be far cheaperAPAC allowance lower; advertised 10 Gbit/s port is sharedSpecials are stock-sensitive, frequently annual, and can be no-refund
DreamHostSelf-managed Stack 4 $8.99 for 3 months, then $15.99; 2 EPYC vCPU, 4 GB, 75 GB NVMeUnmetered under fair use; IPv4/IPv6Separate managed line changes root, backups, resources, term, and support. Do not mix them
RackNerdRegular 512 MB $26.99/year; specials can offer 1 GB from $21.99/yearPlan allowances and ports vary; one IPv4 commonly listedUnmanaged; special inventory, renewal, and backup inclusion require cart/TOS confirmation

Hyperscaler virtual machines

ProviderSmall-machine modelNetwork and egressStopped/IP/storage traps
AWS LightsailLinux IPv4 bundles from $5/month: 2 burstable vCPU, 0.5 GB, 20 GB SSD, 1 TB. The $44 8 GB bundle includes 5 TBBoth inbound and outbound consume allowance; only chargeable excess outbound bills. Common US/EU excess $0.09/GBSame-bundle allowance pools inside one region. Snapshot/delete when cost cessation matters
AWS EC2Region/family/OS-dependent; T3/T3a/T4g are credit-based burstable families100 GB account-level internet egress free, then common US first tier $0.09/GB. Cross-AZ often $0.01/GB each directionStopping EBS-backed VM ends compute, not EBS/IP. Public IPv4 $0.005/hour. NAT example adds $0.045/hour + $0.045/GB
Google Compute EngineE2 shared-core exposes 2 vCPU but fractional sustained physical CPU on micro/small/medium; ordinary E2 scales to 16 vCPU/64 GB in scopePremium Tier default; 1 GiB free then $0.12/$0.11 tiers to Europe/North America. Standard Tier is opt-inStop ends CPU, but disks and retained external IPv4 continue. Cloud NAT adds gateway and $0.045/GiB processing
Microsoft AzureDynamic regional pricing; B-family is credit-based and spans small shapes through 16 vCPU/64 GiB100 GB free; Europe/North America Premium Global Network then $0.087/GB. ISP routing can be cheaper with a routing tradeoffGuest shutdown stays allocated/billable. Only deallocation ends compute; disks/static IP/NAT can continue
Oracle OCIFlexible shapes; x86 OCPU equals 2 hardware threads, Ampere A1 OCPU equals 1 Arm coreFirst 10 TB/month public egress free; intra-region movement and ingress free under current pricingControl-plane stop pauses ordinary VM compute, guest shutdown does not. Boot/block volumes continue. Free A1 totals 4 OCPU/24 GB
IBM Cloud VPCProfiles cover 2 vCPU/1 GB through 16/64; published profile bandwidth is shared between network and storageCurrent egress and floating-IP rates require regional calculatorStop suspends compute/license billing; persistent storage, floating IP, load balancer, and network services remain
Alibaba ECSRegional calculator; x86 vCPU generally hardware thread, Arm vCPU physical core; T6 uses CPU creditsIngress free; outbound can use fixed-bandwidth or traffic billing. Peak bandwidth is a cap, not SLASubscription and standard PAYG keep billing when stopped. Eligible PAYG “economical mode” is conditional; disks/EIPs continue
Tencent CVMRegional, promotional, and family-dependent calculatorFixed-bandwidth or traffic billing; selected peak is maximum, not committed throughputPAYG stop normally keeps billing. Only eligible cloud-disk instances support explicit stop-charging mode

Managed VPS and managed cloud

ProviderCurrent managed/mixed offerIncluded operating layerWatch for
Liquid Web1 GB VPS $5/month self-managed reference; 4 GB offer $8.50 for two months then $17Self-managed or fully managed; control panel and OS change price; managed tiers include backup allocation10 Gbit/s network is attachment capacity; outbound allowance is metered. Powered-off relief unverified
InMotion4 vCPU/8 GB $9.99/month for 24 months, renewal $16.99; 160 GB NVMe, 5 TB, two IPsRoot, DDoS protection, migration help; management/panel selections varyLarger plan can say unlimited transfer. Confirm exact management, backup, term, and renewal
KnownHostUnmanaged $5/month: 1 core, 1 GB, 25 GB NVMe, 2 TB, two IPv4; managed NVMe starts much higherManaged plans add complementary backups and support; unmanaged restores can cost extraCompare high-frequency core wording, port, backup retention, and management boundary
RoseHostingNVME 2 $43.99 monthly or $39.59 annual-effective; 2 dedicated Xeon cores, 2 GB, 50 GB NVMeFully managed with root, dedicated IP, weekly automatic backupsUnmetered traffic remains subject to acceptable use; port speed unpublished
ScalaHostingManaged cloud from $29.95 introductory, renewal $54.95; unmanaged from $19.95, renewal $24.95Managed: SPanel/operations and offsite backups. Unmanaged: root and free snapshotCPU/RAM/storage sliders are dynamic; backup retention add-ons and renewal change the bill
CloudwaysFlexible Small example $11/month; 1 vCPU, 2 GB, 50 GB, 2 TBManaged application platform, not raw-root VPS; infrastructure/provider choiceExcess bandwidth changes by underlying cloud. Required offsite backup storage $0.033/GB/server
hosting.comUnmanaged from $4.99 introductory, regular $9.99; exact resources dynamically renderedUnmanaged root or managed Linux/cPanel without root; managed line advertises daily backups“Up to 1 Gbit/s” and unmetered fair use are not guaranteed throughput; verify exact checkout configuration
GoDaddy VPSLocale-specific. Canadian example: 1 vCPU, 2 GB, 40 GB NVMe at C$11.99 on 3 years, regular C$19.99Self-managed root by default; optional management; snapshots and dedicated IPUnlimited transfer, but port is unpublished. Backup wording varies across configurations and locales

What “unlimited” transfer should mean during procurement

“Unlimited” normally means there is no metered monthly byte allowance for ordinary permitted use. It does not mean infinite capacity or immunity from:

  • the plan’s port-speed ceiling;
  • network congestion;
  • DDoS mitigation;
  • abuse controls;
  • content and acceptable-use rules;
  • sustained-use review;
  • a restriction on resale, proxies, scraping, scanning, or public mirrors;
  • a provider’s right to protect other customers.

Ask a provider with an unmetered plan:

  1. Is the stated port rate dedicated, shared, typical, or maximum?
  2. Can the port sustain that rate continuously?
  3. Is traffic shaped after a threshold?
  4. Is inbound and outbound treated identically?
  5. Are backups, private traffic, mitigation traffic, and cross-region traffic included?
  6. What happens before suspension—alert, throttle, invoice, or manual review?
  7. Are media streaming, CDN origin, VPN, public mirror, and file-distribution workloads permitted?

An honest capped product can be safer than an “unlimited” product whose operational threshold is unpublished.

IPv4 is now a real line item

Public IPv4 scarcity has changed cloud bills. Providers take several approaches:

  • one IPv4 included in the bundle;
  • IPv6-only plan with IPv4 as an add-on;
  • hourly or monthly fee for every public IPv4;
  • charge only for idle or unattached addresses;
  • NAT or load balancer required for private-only instances.

For a single VM the difference may be small. For 50 tiny workers, the address bill can exceed compute savings. Prefer private networking and IPv6 where the application permits it, but include the real cost of NAT gateways, managed load balancers, egress paths, and operational complexity.

Powered off is not deleted

The console power button is not a billing contract.

Common behaviors:

  • Full compute keeps billing: the VM still reserves its allocation until deleted.
  • Compute stops, storage continues: deallocation releases CPU/RAM but attached disks and IPs remain chargeable.
  • Archive mode: the provider converts the VM into a storage-only state. Clouding.io documents this explicitly.
  • Per-active-minute compute: some products stop compute billing in a documented stopped state, but volumes, snapshots, backups, licenses, or reserved IPs remain.

Test the exact operation:

  1. Shut down inside the guest.
  2. Stop from the control plane.
  3. Deallocate or archive, if available.
  4. Read the invoice meter.
  5. Delete only after exporting data and confirming dependent resources.

Guest shutdown alone often leaves the instance allocated and billable.

DDoS protection is not application security

Infrastructure DDoS mitigation can absorb or filter volumetric network attacks. It does not automatically prevent:

  • WordPress/plugin exploitation;
  • credential stuffing;
  • SQL injection;
  • vulnerable admin panels;
  • malicious API requests that look valid;
  • data exfiltration;
  • ransomware through stolen SSH keys.

A provider firewall is not a web application firewall. A WAF is not patch management. Backups are not incident response. Define each layer separately: network filtering, host firewall, SSH policy, OS updates, secrets, application security, logs, alerts, and recovery.

Managed versus unmanaged VPS

An unmanaged provider normally keeps the physical host, virtualization layer, and network operating. You own the guest:

  • OS patching;
  • SSH hardening;
  • firewall;
  • web/database stack;
  • backups and restore testing;
  • monitoring;
  • malware response;
  • application troubleshooting.

A managed VPS may add a control panel, migrations, monitoring, OS updates, service recovery, backups, security tools, and application support. Read exclusions. Some vendors manage only supported control-panel software. Others will advise but not modify custom services. Root access can reduce what the provider will guarantee.

Ask for a responsibility matrix before buying:

IncidentProviderCustomerShared
Physical host failure
Guest OS security update
Broken web-server configuration
Database corruption
Compromised WordPress plugin
Restore from backup
Performance diagnosis
DDoS attack

Blank cells force the sales promise to become an operational answer.

Recommendations by workload

Small website or API

Start with 1–2 vCPU and 1–4 GB RAM on Hetzner, DigitalOcean, Akamai Cloud/Linode, Vultr, OVHcloud, Clouding.io, or Civo. Choose the nearest credible region, enable backups, and monitor memory, CPU steal, disk latency, and outbound transfer.

Do not overbuy 16 vCPU for a cacheable site. A smaller server with a CDN, correct caching, tuned database, and lightweight application can outperform a badly configured larger VM.

WordPress and WooCommerce

An unmanaged VPS gives control, not automatic WordPress expertise. Budget for patching, backups, mail delivery, caching, malware response, staging, and monitoring. If the team does not want that work, compare managed WordPress hosting before choosing a VM.

For a custom stack, 2–4 vCPU and 4–8 GB is a sensible test range for a modest site. WooCommerce checkout, search, imports, webhooks, scheduled actions, and logged-in traffic must be tested uncached.

VPN, proxy, or egress-heavy service

Network policy is the product. Shortlist OVHcloud, Civo, Gcore, Hetzner, and region-appropriate specialists only after confirming that the intended use is allowed. Check:

  • port cap;
  • included transfer;
  • overage or throttling;
  • DDoS behavior;
  • IP reputation;
  • traffic accounting direction;
  • abuse response;
  • legal jurisdiction.

Database server

Prioritize predictable CPU, memory headroom, storage latency, durability, and recovery over headline disk capacity. Use dedicated CPU or a measured general-purpose family. Record p95/p99 disk latency under load, not only sequential throughput. Keep an independent backup and rehearse point-in-time recovery.

Build runner or batch worker

Burstable/shared CPU can be poor value when compilation is continuous. Compare dedicated/high-frequency families from Akamai Cloud/Linode, Vultr, UpCloud, Gcore, and hyperscalers. If workers are disposable, stopped/deallocated billing and image startup time may matter more than monthly bundle price.

Game server

Single-thread performance, jitter, DDoS handling, route quality, and region matter more than aggregate vCPU count. Test from real player networks at peak hours. Dedicated CPU is usually safer for a stable tick rate.

Storage, backups, or public mirrors

Do not assume a cheap large-disk VPS is a backup product. Verify disk durability, backup isolation, restore throughput, traffic rules, checksum workflow, and provider policy. Object storage may be operationally safer even when the per-GB price is higher.

Enterprise application inside a cloud ecosystem

Choose AWS EC2, Google Compute Engine, Azure VMs, Oracle OCI, IBM Cloud VPC, Alibaba ECS, or Tencent CVM because of the surrounding architecture—identity, databases, private networking, compliance, managed services, procurement—not because a single VM card looks cheap.

Model every network path. A multi-zone design can add traffic charges between the load balancer, application, database, NAT gateway, logs, and object storage.

A benchmark protocol that produces useful evidence

Run the same image, region class, and instance size for long enough to observe variance.

Before the test

  1. Record provider, region, zone, plan ID, CPU model, vCPU class, RAM, disk type, kernel, and test time.
  2. Confirm whether the plan is shared, burstable, or dedicated.
  3. Disable unattended jobs that would distort results, but keep production-like monitoring.
  4. Use at least three instances where noisy-neighbor risk matters.

CPU

  • Single-thread and all-core work.
  • Ten-minute burst and one-hour sustained run.
  • Record throughput plus run-to-run variance.
  • Watch CPU steal time, throttling, and credits.

Memory

  • Measure available memory after base services start.
  • Test bandwidth and latency only as supporting data.
  • Trigger realistic application pressure; do not use swap as a substitute for sizing.

Storage

  • Random 4K read/write at realistic queue depths.
  • Sequential read/write.
  • fsync latency for database-style durability.
  • Sustained test long enough to exhaust burst caches.
  • Repeat while snapshots/backups run.

Network

  • Same-region private path.
  • Public internet to at least three independent networks.
  • Multiple geographic destinations.
  • Single-flow and multi-flow throughput.
  • Latency, jitter, packet loss, and retransmissions.
  • Repeat at regional peak and off-peak hours.

Application

Synthetic infrastructure tests do not replace the real workload:

  • cache hit versus uncached request;
  • database read/write mix;
  • queue and cron behavior;
  • login, search, checkout, webhook, upload, and backup;
  • failover and restoration.

Keep the test scripts and raw results. A provider can change host hardware without changing the plan name.

Buying checklist

  • Exact region, currency, tax treatment, and term?
  • Cash due today and price after promotion?
  • Shared, burstable, or dedicated vCPU?
  • Processor generation fixed, selectable, or variable?
  • Local or network disk? Persistent or ephemeral?
  • Published IOPS/throughput baseline, burst, and cap?
  • One public IPv4 included? What does an extra or idle address cost?
  • Inbound, outbound, or both directions counted?
  • Decimal GB/TB or binary GiB/TiB?
  • Transfer allowance pooled, prorated, and reset how?
  • Overage fee, throttle, suspension, or fair-use review?
  • Port speed maximum, expected, or guaranteed?
  • Private, cross-zone, and cross-region traffic cost?
  • NAT gateway and load-balancer hourly plus per-GB charges?
  • Powered-off, stopped, deallocated, archived, and deleted billing?
  • Backup schedule, retention, location, encryption, and restore process?
  • Snapshot price and whether snapshots survive instance deletion?
  • DDoS scope? Host firewall? WAF? Malware response?
  • Managed support scope, exclusions, response target, and escalation?
  • Export path for disks, images, IPs, DNS, logs, and backups?

FAQ

What is the best VPS provider in 2026?

For low-cost European general use, Hetzner is the strongest mainstream default. For explicit unmetered transfer, compare OVHcloud. For approachable global developer clouds, shortlist DigitalOcean, Akamai Cloud/Linode, and Vultr. For no separately billed VM egress, examine Civo and Gcore. The best answer changes with region, sustained CPU, storage latency, traffic, and support.

What is the cheapest VPS?

There is no durable global winner. Clouding.io publishes a tiny configurable entry, while Hetzner, netcup, Contabo, Time4VPS, Hostinger, RackNerd, and low-stock specialists compete aggressively. Compare upfront cash, renewal, CPU class, IPv4, backup, and transfer before calling a plan cheap.

Which VPS has unlimited bandwidth?

OVHcloud VPS currently publishes unlimited traffic with plan-specific public bandwidth caps. Other providers use large included allowances, no separate VM egress fee, or fair-use language. “Unlimited” never removes the physical port rate, provider policy, or end-to-end network constraints.

Which VPS is best for 10 TB of monthly egress?

Start with OVHcloud, Hetzner EU, Civo, Gcore, and Oracle OCI. Compare location, port rate, allowance, policy, and the application’s use. Ordinary hyperscaler internet egress can cost hundreds or more than a thousand dollars at this volume before compute and other network services.

Is a vCPU a physical core?

Not necessarily. It can be a scheduled share, a hardware thread, a burstable allocation, or a dedicated virtual thread. Read the provider’s CPU-class documentation. “Dedicated vCPU” still does not automatically mean an entire physical core.

Does turning off a VPS stop billing?

Often no. Guest shutdown usually leaves resources allocated. Some clouds require deallocation; some providers bill until deletion; Clouding.io offers an archive state that stops CPU/RAM billing while storage and backups continue. Read the exact lifecycle rules.

Are snapshots backups?

Snapshots are useful recovery points but may share the same account, credentials, region, and failure domain as the VM. Important data needs an independent, encrypted, versioned copy plus a tested restore procedure.

Is a 1 Gbit/s VPS guaranteed to deliver 1 Gbit/s?

No. It may describe a port ceiling. CPU, virtualization, route, remote endpoint, protocol, distance, congestion, and policy all affect delivered throughput. Look for a contractual guarantee or SLO if minimum throughput matters.

Should I choose AWS Lightsail or AWS EC2?

Choose AWS Lightsail for a simpler fixed bundle with included transfer. Choose AWS EC2 when you need the wider AWS instance catalog, networking, IAM, autoscaling, managed services, and architecture controls. Model egress and cross-zone paths for both.

Is managed VPS worth the price?

Yes when the included patching, migration, monitoring, panel, recovery, or troubleshooting replaces work the team would otherwise perform badly or expensively. No when the application needs a custom stack outside the provider’s management boundary and the team already operates it well.

How much RAM does a small VPS need?

Use 1 GB only for tightly controlled single-purpose services. Start most small web applications at 2–4 GB, and test 4–8 GB for databases, WordPress, stores, or multi-container stacks. Measure actual peak usage and leave recovery headroom.

Which provider is best in Spain?

Clouding.io is the clearest Spain-based option here: it operates in Barcelona, assigns Spanish-geolocated fixed IPs, offers custom sizing, includes 4 TB monthly transfer, and supports archive mode. OVHcloud, AWS, Azure, Google Cloud, and others also offer Spanish or nearby regions/products; compare the exact service, latency, and data-location requirements.

Can I host WordPress on a VPS?

Yes, but you become responsible for more of the operating system, web stack, mail, backups, security, and incident response. Compare that burden with the 30-provider managed WordPress guide and the focused WordPress hosting with New Relic comparison.

Conclusion

The VPS market is not one ladder from cheap to expensive. It is several billing and responsibility models sharing the same label.

  • Buy a value VPS when the workload is understood and the team can operate Linux.
  • Buy a developer cloud when API quality, regions, images, and predictable self-service matter.
  • Buy a high-transfer plan when network economics dominate.
  • Buy a hyperscaler VM when the surrounding cloud architecture creates the value.
  • Buy managed VPS when operational responsibility—not raw RAM—is the constraint.

Then test the exact region and plan. Verify the invoice with real traffic. Restore a backup. Measure sustained CPU, disk latency, and network variance. The best provider is the one whose limits remain acceptable after the marketing card ends.

Official sources checked

00
Previous

Previous article

Best app deployment platforms in 2026: 87 compared

Next article

Best WordPress hosting in 2026: 30 hosts compared

Next

Explore the tools or browse interactive maps for more experiments.

Back to Blog Posts