Cheapest VPS with dedicated CPU cores — real options starting at $2/month

The "4 vCPU" problem

When Contabo sells you a VPS with "4 vCPU" for $6.60/month, and Hetzner sells you a CCX with "2 dedicated vCPU" for EUR 42.99/month, those are not the same kind of product wearing different price tags. They're structurally different things.

Contabo's 4 vCPU means you have access to four virtual CPU threads on a physical host that's running many other tenants' VPS instances. When demand is low, you might get close to four full cores of performance. When the host is busy — and budget hosts are often busy — the hypervisor divides the physical CPU time among all tenants, and your "4 vCPU" delivers whatever's left. Community reports on LowEndTalk document 20-50% CPU steal on busy Contabo nodes. Your 4 vCPU is effectively 2-3 vCPU some of the time.

Hetzner's 2 dedicated vCPU means two physical AMD EPYC cores are assigned exclusively to your instance. No other tenant can use them. Your 2 cores deliver 2 cores of performance at 3am and at 3pm. The number is smaller, but it's a number you can trust.

The budget VPS market rarely makes this distinction visible. Most comparison sites list "vCPU" as a flat number, as if a shared Contabo vCPU and a dedicated Hetzner vCPU were the same unit. They're not, and the gap between them matters for any workload that sustains CPU load: compilation, video encoding, database queries, game servers, anything that doesn't sit idle waiting for HTTP requests.

CPU steal: what it is and how to check yours

Open a terminal on your VPS. Run top. Look at the %st column in the CPU line. That number is steal time — the percentage of CPU cycles your virtual machine requested but didn't get because the hypervisor gave them to another tenant.

• 0-2% steal: Normal. The host isn't overloaded. This is what dedicated-core and no-overselling providers typically show.
• 5-10% steal: The host is moderately loaded. Acceptable for web servers and light workloads. Noticeable for CPU-bound tasks.
• 20%+ steal: The host is oversold enough that you're losing a measurable fraction of the CPU you're paying for. Contabo users report seeing this range on busy nodes.
• 50%+ steal: You're getting roughly half your nominally allocated CPU. Time to migrate.

The test: run top during peak hours (UTC afternoons are usually worst on European hosts) and watch %st for 10 minutes. If it stays below 5%, your provider either has dedicated cores or an honest overselling ratio. If it spikes above 20%, you're on an oversold node regardless of what the plan page says.

One catch: CPU steal is invisible at purchase time. No provider advertises their overselling ratio. You can only measure it after you've bought the VPS. This is why the distinction between "dedicated core" (contractual guarantee) and "shared vCPU" (provider's word) matters — the dedicated label is the only pre-purchase signal.

The dedicated-core tier ($2-$11/month)

Two providers in the budget tier still offer genuinely pinned dedicated CPU cores — your vCPU maps to a physical core that no other tenant can touch. Until June 2026 there were three; Hetzner's CCX line was the third, and its exit is covered below because it changes the whole map.

DartNode VDS Slices — from $2/month

The cheapest dedicated-core product in the market. A Slice allocates a physical CPU core and a dedicated RAM segment exclusively to your instance. The entry plan ($2/month) gives you 1 dedicated core, 1GB RAM, 15GB NVMe on hardware in DartNode's own Texas datacenter. Larger Slices scale up from there.

The catch is twofold. First, Slices are frequently sold out — dedicated resources on owned hardware means finite capacity, and DartNode doesn't oversell. Second, DartNode's 2025 operational track record included servers offline for weeks, backup failures, and 7-day support wait times. The company has acknowledged the problems and added staff in 2026, but the history is short enough that trust is still being rebuilt. For a $2/month dev box or side project where you keep off-server backups, the value math works. For production, the reliability question hasn't been fully answered yet.

Netcup Root Server — from ~$11/month

Netcup explicitly separates its product line into VPS (shared resources, cheaper) and Root Server (dedicated resources, more). The RS 1000 — roughly €10.36/month after May 2026 pricing — gives you dedicated AMD EPYC 9645 cores and DDR5 RAM in netcup's own infrastructure in Nuremberg or Vienna. Unlike a VPS where "4 vCPU" means "shared access to 4 threads," a Root Server's core count is a floor, not a ceiling.

Netcup is the most mature option in this tier. German company, connected to the Anexia Backbone Europe with direct peering at DE-CIX, Dell and HP server hardware, and replacement parts stocked on-site for same-day swaps. The trade-off is location: Nuremberg, Vienna, Manassas (VA), or Singapore. If you need a dedicated-core VPS in, say, Los Angeles or Tokyo, netcup can't help.

Hetzner CCX — the exit that redrew this list (now EUR 42.99/month)

For years the answer to "cheapest trustworthy dedicated cores" was Hetzner's CCX line: dedicated AMD EPYC vCPUs in Hetzner's own datacenter parks, with the CCX13 (2 dedicated vCPU, 8GB RAM, 80GB NVMe, 20TB traffic) sitting around EUR 16. That era ended on June 15, 2026, when Hetzner repriced the dedicated line: the CCX13 now costs EUR 42.99/month ($50.49), verified against Hetzner's own price API on 2026-07-30.

Hetzner is still the provider other providers get benchmarked against: three owned datacenter parks (Nuremberg, Falkenstein, Helsinki), ISO 27001 certified, a track record measured in decades. But the budget math is gone. At $50.49/month the CCX13 costs nearly eight times a Contabo entry VPS, and its discount against DigitalOcean's $63 dedicated Droplet has shrunk from 5x to about 20%. If you need guaranteed CPU from the lowest-risk provider and the budget ceiling is not $15, the CCX is still excellent. It just no longer belongs in this article's price bracket, and pretending otherwise would misprice the whole comparison.

Hetzner and DigitalOcean are included for context. Before June 2026, the gap between the CCX13 (then ~$16) and DigitalOcean ($63) was where the "dedicated CPU = expensive" myth came from. After the June repricing, Hetzner joined the expensive club, and the true budget dedicated-core story now belongs to DartNode and Netcup.

The low-steal tier (no-overselling policies)

Between "dedicated core" and "oversold shared" sits a middle ground: providers that don't contractually guarantee dedicated cores but operate policies that keep CPU steal near zero.

BuyVM / FranTech — from $3.50/month

BuyVM runs a deliberate no-overselling policy. They limit the number of VPS instances per physical node, which is why the service is perpetually sold out — when a node is full, they don't add more tenants; they wait for capacity. The result is that BuyVM users consistently report near-zero CPU steal on AMD Ryzen 3900X hardware, despite the CPU being technically "shared."

The distinction from true dedicated cores matters contractually but less in practice: BuyVM's no-overselling stance is a core brand identity, not a promotional bullet point they could quietly drop. At $3.50/month for 1 vCPU / 1GB RAM with unmetered 1 Gbps bandwidth, it's the closest thing to dedicated-core pricing without the dedicated-core label. The catch, as always with BuyVM: good luck buying one when they're in stock.

HostHatch NVMe — from $4/month

HostHatch's NVMe line runs on AMD EPYC 7R13 processors in facilities like Equinix AM3 and Interxion ZUR1. VPSBenchmarks recorded Geekbench 6 single-core scores above 1,300 on HostHatch's NVMe 32GB plan — a number that implies low CPU contention, because Geekbench scores drop visibly on oversold nodes. HostHatch hasn't publicly stated a no-overselling policy, but the hardware quality and benchmark consistency suggest conservative allocation. The support problems and promotional-tier deprioritization are the trade-off.

SpeedyPage — from $4.29/month

SpeedyPage runs AMD Ryzen 9950X and 7900 hardware — current-generation desktop CPUs with high single-thread performance. Community users on LowEndTalk have reported "CPU steal is excellent, only Singapore occasionally sees slightly higher rates." SpeedyPage hasn't claimed dedicated cores or a formal no-overselling policy, but the benchmark results and community reports are consistent with honest allocation. At $4.29/month with DDR5 RAM and Gen4 NVMe, it occupies the space between "budget VPS" and "cloud platform" that DartNode aims for with Slices, but with a better operational track record.

The shared tier (and when it's fine)

Not every workload needs dedicated cores. Most web servers, small applications, and personal projects spend most of their time idle, waiting for HTTP requests. For these workloads, shared vCPU is the right trade-off: you pay less because you're sharing resources you weren't going to use anyway.

Contabo — from $6.60/month (4 vCPU / 8GB)

Contabo is the extreme case of shared-CPU optimization. You get 4 vCPU / 8GB RAM / 100GB SSD for $6.60/month on the current Cloud VPS 4 — specs that no dedicated provider can touch at that price. The CPU overselling is real (20-50% steal documented on busy nodes), and the IOPS throttle on the older SSD line is a separate issue. But if your workload is RAM-bound (a database that fits in memory and rarely queries), storage-bound (file serving, backups), or idle most of the time (a personal site that gets 100 visits/day), Contabo's nominal 4 vCPU is more than enough for the spikes, and the 8GB RAM is the real value. Dedicated cores would be wasted.

RackNerd — from ~$1.83/month

RackNerd's flash-sale VPS plans are shared CPU on upstream colocation hardware. CPU steal varies by location and node — you're at the mercy of whoever's upstream and how aggressively they've packed the host. For the price ($10-25/year on sales), the expectation should be calibrated: this is a disposable VPS for testing, VPN endpoints, or throwaway projects, not a platform for sustained compute. Within that frame, RackNerd delivers.

Which tier is right for your workload

You need dedicated cores if:

• Your workload sustains CPU load for extended periods — compilation, rendering, encoding, CI/CD runners
• You're running latency-sensitive services where a CPU steal spike means a visible user-facing delay — game servers, real-time APIs, trading bots
• You need predictable performance, not just average performance — SLA-driven production systems

A no-overselling provider is enough if:

• You want low CPU steal without paying the dedicated-core premium
• Your CPU usage is bursty (spikes and idle periods) rather than sustained
• You're willing to accept that the "no overselling" guarantee is a policy, not a contract — and trust the provider's track record

Shared CPU is fine if:

• Your workload is RAM- or storage-bound, with CPU usage under 20% most of the time
• You're running a website, blog, or personal app that serves requests in bursts
• Price is the primary constraint and you'd rather have 8GB RAM at $6.60/month than 1GB at $2/month with dedicated cores

FAQ