I spun up identical VPS instances across seven providers this month and ran them through the same suite of tests. Same plan tier, same regions where possible, same workload scripts. The goal was simple: find out which platforms actually deliver what they promise when you push them.
Most VPS marketing pages list vCPUs and RAM but skip the details that matter—CPU steal, disk latency under load, network consistency. I wanted numbers I could trust when recommending hosts to clients.
Test methodology
Each provider got a mid-tier plan—typically 4 vCPUs, 8GB RAM, 160GB SSD. I chose instances in US East or Western Europe datacenters to keep latency comparable. Testing ran over 72 hours to catch performance variation across different times of day.
The benchmark suite included:
- CPU: Sysbench multi-threaded prime calculation, single and multi-core tests
- Disk I/O: FIO sequential and random read/write with varying block sizes
- Network: iperf3 to multiple geographic endpoints, sustained transfer tests
- Real-world load: NGINX serving static files, WordPress with WooCommerce under siege load testing
I also monitored CPU steal time—the percentage of time your VM wanted CPU but the hypervisor gave it to another tenant. High steal means you're fighting for resources.
Provider performance breakdown
DigitalOcean
DigitalOcean's Droplets showed consistent performance across all tests. CPU steal stayed under 2% even during peak hours. Single-core performance landed in the middle of the pack, but multi-threaded workloads scaled well across all four vCPUs.
Disk I/O surprised me. Random read IOPS hit solid numbers with low latency variance. Sequential writes occasionally showed small stutters but nothing that would affect typical web workloads. Network throughput to East Coast endpoints maxed out the allocated bandwidth consistently.
The control panel is still the cleanest in the industry. Spin up a Droplet, get a predictable experience.
Linode (Akamai)
Linode instances delivered the highest single-core CPU scores in testing. If your application is single-threaded or has one hot path, that matters. Multi-core scaling was equally good. CPU steal stayed below 1% for the entire test period.
Disk performance showed excellent random I/O characteristics—important for databases. Sequential throughput met expectations. Network tests revealed Akamai's CDN backbone working in the background; transfers to European endpoints were faster than the instance's geographic location suggested.
One quirk: the first boot after deployment took longer than other providers. After that, rock solid.
Vultr
Vultr's high-frequency compute instances use newer CPUs and you feel it. Single-core bench scores competed with Linode. Multi-core performance scaled cleanly. CPU steal ranged from 1-3%, acceptable for a public cloud.
Disk I/O numbers were middle of the road. Not bad, not exceptional. Random reads handled typical database query patterns fine. Network performance to US endpoints was excellent; international routes showed more variation.
Vultr's snapshot and backup pricing is aggressive compared to competitors—factor that into total cost if you automate backups.
Hetzner Cloud
Hetzner delivered the best price-to-performance ratio, but there's a catch. Their European datacenters performed beautifully—low CPU steal, excellent disk I/O, fast network within Europe. If your audience is European, this is hard to beat.
US deployments showed higher latency and less consistent network performance to American endpoints. The infrastructure is optimized for their home region. Single and multi-core CPU scores were competitive. Disk IOPS matched or exceeded more expensive providers.
API and control panel are straightforward. Documentation assumes you know what you're doing.
OVHcloud
OVHcloud's VPS lineup is confusing—multiple product lines with overlapping specs. I tested their "Comfort" tier. CPU performance was adequate but showed occasional steal spikes up to 8% during evening hours European time. That's noticeable if you're running latency-sensitive apps.
Disk I/O started strong but degraded under sustained heavy writes. Random read performance stayed consistent. Network throughput within OVH's network was fast; external routing took longer hops than necessary in some tests.
Pricing is competitive if you prepay annually. Month-to-month is less attractive. The dashboard needs work—too many clicks to reach common settings.
Contabo
Contabo markets on raw specs—huge amounts of RAM and storage for the price. Performance tells a different story. CPU steal regularly hit 10-15% during testing. Your mileage will vary wildly depending on your noisy neighbors.
Single-core CPU scores were the lowest in the group. Multi-core scaling existed but started from a lower baseline. Disk I/O showed high variance—sometimes fast, sometimes sluggish. Network tests revealed congestion during peak hours.
If you're running a development environment or a forgiving workload where cost matters more than consistency, Contabo works. For production where users notice lag, I'd look elsewhere.
Kamatera
Kamatera offers granular customization—pick exact CPU, RAM, and storage combinations. I built a config matching the other tests. Performance sat in the middle tier for most metrics. CPU steal averaged 3-5%, higher than the top performers but manageable.
Disk I/O was the standout—very low latency on random operations. Sequential throughput met spec. Network performance varied by endpoint; some routes were fast, others took unexpected paths.
The interface feels dated but functional. Billing is transparent—you pay for exactly what you configure. Support responded quickly during my tests.
What the numbers actually mean
Raw benchmark scores don't tell the full story. Here's what mattered in real-world testing:
CPU steal under 3% kept applications responsive during traffic spikes. Above 5%, you notice lag in PHP-FPM workers, slower database queries, delayed cron jobs. The difference between 1% and 8% steal is the difference between smooth scaling and mysterious slow requests.
Disk I/O latency matters more than throughput for most hosting workloads. A database doing thousands of small random reads cares about milliseconds per operation. Web servers serving files want throughput. Match the metric to your use case.
Network consistency beat peak speed. A host delivering 800 Mbps reliably outperformed one that hit 950 Mbps sometimes and 400 Mbps other times. Users don't average their experience.
Choosing based on your workload
For general web hosting—WordPress, small apps, APIs—DigitalOcean or Linode give you the most predictable experience. You pay a bit more but save hours troubleshooting weird performance issues.
Database-heavy applications benefit from Linode's strong random I/O or Hetzner's European performance. Low latency on small operations adds up when you're doing millions of them.
If you're serving European users exclusively, Hetzner offers exceptional value. Just don't expect the same performance from their US locations.
Development and staging environments can use Contabo or OVHcloud to save money. The performance variance won't affect your workflow much and the cost savings are real.
CPU-bound workloads like video encoding, image processing, or compilation want Linode or Vultr's high-frequency plans. Single-core speed makes a measurable difference.
CPU steal and the oversubscription problem
Every provider oversubscribes their physical hosts to some degree—it's how cloud economics work. The question is how much.
CPU steal is the best indicator of oversubscription you can measure from inside your VM. Run this to check your current steal:
mpstat 1 60 | grep -A 60 'CPU' | awk '{print $1,$3,$4,$10}'
Watch the steal column. Under 2% is excellent. 2-5% is acceptable. Above 5% means you're competing for CPU time more than you should be. Above 10% is a problem—either migrate to a different host or switch providers.
Some hosts let you request instance migration if performance is poor. Others expect you to rebuild on a new instance. DigitalOcean and Linode have usually moved instances for me when I've provided steal metrics.
Network performance patterns
I ran iperf3 tests to endpoints in five regions—US East, US West, Europe, Asia, and South America. The results showed three patterns:
First-tier providers (DigitalOcean, Linode, Vultr) had direct peering with major networks. Traffic took efficient paths. Latency matched geographic distance predictably.
Second-tier providers (OVHcloud, Kamatera) had good regional performance but took extra hops for intercontinental traffic. Not terrible, but noticeable if you check traceroutes.
Budget providers (Contabo) showed congestion during peak hours and sometimes routed traffic through unexpected transit providers. Your mileage varies.
For most web hosting, network performance within the provider's primary region matters more than cross-continental speed. CDN offloads the majority of static content anyway.
Disk I/O and storage types
All tested providers advertised SSD storage but implementation varies. Some use local NVMe drives, others use network-attached SAN with SSD backing. The difference shows up in latency.
Test your disk with:
fio --name=random-read --ioengine=libaio --rw=randread --bs=4k --size=1G --numjobs=1 --runtime=60 --time_based --direct=1
Random read latency under 1ms is excellent—local NVMe. 1-3ms is good—fast SAN. Above 5ms, you'll notice it in database performance.
Sequential throughput matters less unless you're moving large files regularly. Most web applications make thousands of small requests, not giant sequential reads.
Price and hidden costs
Sticker price per month doesn't tell the whole story. Factor in:
- Snapshot costs (some providers charge per GB per month)
- Backup fees (can double your bill)
- Network transfer overages (rate varies wildly)
- DDoS protection add-ons
- Support tier costs
DigitalOcean and Vultr charge for snapshots but include generous transfer. Hetzner includes backups but charges more for extra transfer. Linode bundles most features but costs more upfront.
Run the calculation with your actual usage patterns. A cheaper base price with expensive snapshots can cost more than a pricier host with free backups if you snapshot daily.
What to test before committing
Don't trust provider marketing or my benchmarks alone. Spin up a test instance and run your own workload.
Install your actual application stack. Run your database queries. Push realistic traffic through your web server. Monitor CPU steal, disk latency, and network consistency for at least 48 hours across different times of day.
Most providers offer hourly billing or credits for new accounts—test costs a few dollars and saves you from migrating later when performance issues appear under real load.