VPS CPU Benchmark Comparison 2026 — 20+ Providers Tested

Methodology: 72 Hours, Not 60 Seconds

A single benchmark run is a screenshot, not a measurement. Here is what we did on each of the 22 servers:

• Plan: Entry-level shared vCPU from each provider, US datacenter (east coast where available), provisioned mid-week
• OS: Ubuntu 24.04 LTS, fresh install, no additional software beyond monitoring tools
• Sysbench loop: sysbench cpu run --threads=1 --time=30 executed every 15 minutes for 72 hours — 288 samples per provider
• Geekbench 6: Single-core run at 4 AM, 12 PM, and 8 PM local datacenter time — 9 samples over 3 days
• CPU steal: mpstat 10 running continuously, sampled every 10 seconds for the full 72 hours — 25,920 steal-time readings per provider
• Period: February 24 through March 14, 2026 — staggered starts to avoid testing all providers during the same global traffic pattern

From the 288 sysbench samples per provider, we computed: median (typical), p99 (the floor), and CV% (coefficient of variation — how much it wobbles). CV below 3% is rock-solid. Above 8% is CPU roulette. We ranked by p99, not median, because your users experience the tail, not the average.

The Consistency Ranking — All 22 Providers

Sorted by p99 sysbench — the floor your CPU maintains 99% of the time. CV% = coefficient of variation (lower = more stable). Steal (p95) = 95th percentile CPU steal over 72 hours.

Read carefully. Hostinger posts the highest median (4,410) but drops to third on p99 (3,960). Hetzner barely budges under pressure — p99 of 4,240 is just 1.6% below median. Kamatera's median of 4,260 looks spectacular for $4/mo, but its p99 falls 12% to 3,740. The 5.2% CV and 2.8% steal explain why: tight packing on entry-tier instances.

CPU Steal Time — The Hidden Tax

CPU steal time is the metric that benchmark articles almost never mention, and it is the single best predictor of inconsistent performance. When the hypervisor gives your scheduled CPU time to another tenant, Linux logs it as "steal." You can watch it in real time with mpstat 1 or top (the "%st" column).

Think of it as a tax. If steal is 5%, every operation takes 5% longer. And it spikes during peak hours — exactly when you need CPU most.

Our 72-hour p95 steal observations fell into four tiers:

• Under 1% (premium): Hetzner (0.3%), AWS Lightsail (0.4%), Vultr (0.5%), Google Cloud (0.6%), Azure (0.7%), Linode (0.8%). Your CPU is your CPU, almost all the time.
• 1-3% (solid): BuyVM (1.0%), Hostwinds (1.1%), Hostinger (1.2%), DigitalOcean (1.4%), UpCloud (1.9%), IONOS (2.2%), A2 Hosting (2.5%). Noticeable in benchmarks, unlikely to affect real applications. Hostinger landing here explains its median-to-p99 gap.
• 3-5% (oversubscribed): Kamatera (2.8%), OVHcloud (2.9%), ScalaHosting (3.1%), InterServer (3.4%), Cloudways (3.8%). Expect occasional latency spikes, longer cron jobs, intermittent database timeouts.
• 5%+ (aggressive density): Contabo (6.2%), RackNerd (7.8%), GreenCloudVPS (8.3%), SpartanHost (9.1%). Meaningful CPU loss to neighbors. An 80ms PHP render becomes 95ms regularly, 130ms occasionally.

This explains a pattern from community forums: "RackNerd was fast when I signed up but got slower over months." New servers land on fresh hosts with low density. As the provider sells more instances on that hardware, steal climbs. The initial benchmark was real — just not sustainable.

Peak vs. Floor: Why Rankings Flip

To make the ranking divergence concrete, here is how the top 10 reorder when you switch from median to p99:

The pattern: low steal and tight CV climb the p99 ranking. Flashy peaks with oversubscription drop. Vultr jumps from 5th to 2nd on its 2.1% CV. Kamatera falls from 3rd to 6th because 12.2% median-to-p99 drop means the headline number misrepresents the daily experience.

This is not about Kamatera being bad — at $4/mo with a p99 of 3,740, it beats most providers at any price. But if you chose it over Hetzner because the median was higher, you made the wrong call. Hetzner's p99 of 4,240 beats Kamatera's median of 4,260 in terms of what you can depend on.

Geekbench vs. Sysbench — Which to Trust

Sysbench tests a single operation (prime calculation) isolating clock speed and IPC. Geekbench 6 runs 20+ workloads — AES-XTS, JPEG compression, HTML5 parsing, SQLite — exercising different CPU subsystems.

Sysbench is the better ranking tool for VPS because its simplicity produces less variance from non-CPU factors. Geekbench scores on VPS are more volatile — cache-heavy subtests (SQLite) are more affected by neighbor contention than compute-only ones (AES with AES-NI).

That said, GB6 single-core correlated with sysbench at r=0.94 across our 22 providers. Notable divergences:

• Hetzner and Hostinger scored proportionally higher on Geekbench, suggesting their EPYC hardware benefits from wider instruction set tests (AVX-512, AES-NI)
• BuyVM and OVHcloud scored proportionally lower, consistent with older hardware lacking modern instruction extensions
• AWS Lightsail showed up to 15% Geekbench variance between runs, likely from T3 burst credit throttling during extended test suites

Use sysbench for provider comparisons. Use Geekbench if your workload is encryption-heavy or involves image/video processing.

AMD EPYC vs. Intel Xeon in 2026

We identified hardware via /proc/cpuinfo and lscpu on each server. The 2026 VPS landscape runs on two processor families:

• AMD EPYC 9004 (Genoa/Bergamo): Hetzner, Vultr (HF nodes), Hostinger, DigitalOcean (Premium). Scored 8-12% higher on single-thread sysbench than Intel equivalents. The larger L3 cache (up to 384MB shared) reduces memory latency during hypervisor context switches, improving p99 consistency.
• Intel Xeon Sapphire/Emerald Rapids: AWS Lightsail, Google Cloud, Azure, Linode. Solid single-thread performance, slight advantage in AVX-512 heavy workloads (encryption, scientific computing).
• Previous generation: Contabo (EPYC 7002), RackNerd (EPYC 7002 / Cascade Lake mix), BuyVM (Xeon E-2300), GreenCloudVPS (EPYC 7002), SpartanHost (Cascade Lake SP). The 15-30% generational gap compounds with higher oversubscription on budget platforms.

Practical takeaway: pick AMD EPYC 9004 nodes when available. But the provider's oversubscription policy matters more than the chip brand. A conservatively loaded Intel box beats a packed EPYC box every time.

Value Analysis — Consistent Points per Dollar

Traditional value rankings divide peak score by price, rewarding oversubscribed hardware. We divided p99 score by price instead:

Kamatera and Hetzner nearly tie on p99 value, but Kamatera's 5.2% CV means cheap and wobbly while Hetzner's 1.8% CV means cheap and stable. For anything latency-sensitive, Hetzner's extra $0.59/mo is the best infrastructure money you will spend.

Contabo at rank 20 is the biggest revelation. Its raw specs (8GB RAM, 200GB SSD for $6.99) look unbeatable on paper. But a p99 of 2,640 and 8.7% CV mean the CPU is worse than providers at half the price. You are buying RAM and storage at Contabo, not CPU. Memory-bound workloads? Fine. CPU-bound? Look elsewhere.

Who Should Use What

The recommendations sort by what you actually need:

Best Overall: Hetzner ($4.59/mo)

Highest p99. Lowest CV. Lowest steal among non-hyperscalers. 2 vCPU cores, 20TB bandwidth. Limited US datacenters (Ashburn, Hillsboro) and ticket-only support are the only trade-offs. Full Hetzner review.

Best Multi-Region Consistency: Vultr ($5.00/mo)

2.1% CV and 0.5% steal across 17 datacenter locations. Hourly billing, good API. Does not win on peak numbers but delivers what it promises, every hour of every day. Full Vultr review.

Best Peak Performance: Hostinger ($6.49/mo)

Highest median at 4,410. The p99 drop to 3,960 matters less if your CPU usage is bursty — build pipelines, image processing, periodic batch jobs. Fast for 30 seconds, not 30 hours. Full Hostinger review.

Best Budget Consistency: BuyVM ($3.50/mo)

Low raw scores (p99 of 2,980), but 3.3% CV and 1.0% steal beat providers at twice the price. Runs its own hardware, does not oversell. Ideal for light workloads — personal blogs, small APIs, Discord bots. Full BuyVM review.

Best Hyperscaler: AWS Lightsail ($5.00/mo)

2.4% CV and 0.4% steal reflect AWS infrastructure discipline. Burstable T3 means throttling after sustained use — which is actually a consistency feature. Full Lightsail review.

Avoid for CPU Workloads: Contabo, RackNerd, GreenCloudVPS, SpartanHost

8-11% CV and 6-9% steal times. If you see a benchmark article ranking these highly, check whether they ran a single test or a sustained measurement. The single test is real — just not representative.

Frequently Asked Questions

What is p99 CPU consistency and why does it matter more than average benchmark scores?

p99 CPU consistency measures the performance level that 99% of your benchmark samples meet or exceed. If a provider's p99 sysbench score is 3,800 but their average is 4,200, that means 1% of the time your CPU is dropping to 3,800 or below — a 10% performance cliff during the worst moments. For web servers, the worst moments are exactly when you care most: traffic spikes, Black Friday, product launches. Average scores hide these drops. A provider with a 4,000 average and tight p99 of 3,950 will feel faster in production than one averaging 4,300 but cratering to 3,200 during neighbor contention.

How did you measure CPU steal time across providers?

We used mpstat and sar running continuously for 72 hours on each VPS, sampling every 10 seconds. CPU steal time (%steal in mpstat output) shows the percentage of time your vCPU was waiting because the hypervisor allocated physical CPU to another tenant. We recorded the median steal, p95 steal, and maximum steal observed. A median steal above 5% means the provider is oversubscribing CPU heavily. Some budget providers hit 15-20% steal during peak hours, which directly reduces your application performance by that same percentage.

Why do Geekbench scores differ from sysbench scores in VPS environments?

Geekbench 6 runs a battery of real-world workloads (AES encryption, image compression, ray tracing, HTML parsing) that stress different CPU features: AVX-512 instructions, branch prediction, cache hierarchy. Sysbench CPU does a single operation — prime number calculation via trial division — that mostly tests raw clock speed and IPC. On VPS instances, Geekbench scores are more volatile because the diverse workloads trigger different levels of hypervisor overhead. We include both because sysbench gives a clean comparison metric while Geekbench better reflects real application diversity. Neither alone tells the full story.

Do dedicated vCPU plans actually eliminate noisy neighbor problems?

Mostly yes, but not entirely. Dedicated vCPU plans (like DigitalOcean Premium, Vultr High Frequency, or Hetzner Dedicated) pin your VM to specific physical cores that are not shared with other tenants. This eliminates CPU steal time in our measurements — we consistently saw 0% steal on dedicated plans. However, shared resources like memory bandwidth, L3 cache, and the hypervisor itself still create some variance. In our tests, dedicated plans showed 1-3% score variance across 72 hours versus 8-22% on shared plans. The premium is worth it for latency-sensitive workloads.

Which CPU architecture performs better for VPS — AMD EPYC or Intel Xeon?

In our 2026 tests, AMD EPYC 9004 series (Genoa/Bergamo) consistently outperformed Intel Xeon Sapphire Rapids on single-thread sysbench by 8-12%. More importantly for VPS workloads, EPYC's larger L3 cache (up to 384MB shared) reduces memory latency when the hypervisor is context-switching between tenants. Providers running EPYC hardware (Hetzner, Vultr High Frequency, Hostinger) showed tighter p99 consistency. That said, the provider's oversubscription ratio matters more than the chip brand — a lightly loaded Intel box will beat a heavily oversubscribed EPYC box every time.

How often should I re-benchmark my VPS?

Run a quick sysbench test once a quarter, or immediately after any maintenance notification from your provider. Providers migrate VMs between physical hosts during maintenance windows, and your new host may have different hardware or different neighbor density. We have seen providers quietly move accounts to older hardware racks after initial provisioning. If your p99 scores drop more than 10% from your baseline, open a support ticket — you may have been migrated to a more contended host, and most providers will relocate you if you ask.

Is a VPS with lower peak CPU but better consistency actually faster for web applications?

For most web applications, yes. Web servers process requests in a queue. When CPU performance is inconsistent, some requests get fast service and others hit a slow window, creating unpredictable response time spikes. At 50+ concurrent users, these slow requests block the queue and create cascading delays. A server averaging 4,000 with a p99 of 3,900 will maintain smooth response times under load. A server averaging 4,300 but dropping to 3,000 at p99 will show periodic latency spikes that users perceive as the site being "sometimes slow." Tail latency, not average throughput, determines user experience.

Can I improve CPU consistency on a shared VPS without upgrading plans?

Three things help. First, provision your VPS during off-peak hours (Tuesday-Thursday, early morning UTC) — providers often place new VMs on the least loaded host at creation time, and you keep that placement until maintenance. Second, use CPU pinning if your provider exposes it (Kamatera and some OpenStack-based providers do). Third, schedule your heaviest cron jobs during your provider's off-peak hours (typically 2-6 AM in the datacenter's timezone). None of these replace a properly provisioned dedicated-CPU plan, but they can tighten your p99 by 5-10% on shared instances.