Two providers will sell you an eight-core, sixteen-thread dedicated server this month. On AMD’s own specification pages, one of those CPUs runs a 3.6 GHz base clock with 16 MB of L3 cache and a 35 W TDP; the other runs 4.7 GHz with 96 MB of L3 and 120 W. Six times the cache and 3.4 times the power envelope, at the same core count — and neither hosting page shows any of it. This guide shows how to compare the best dedicated server providers when the deciding number is missing from the comparison.
Why dedicated server providers can’t be ranked on published specs
Most “best dedicated server providers” pages produce an ordered list. The inputs for such a list are rarely published consistently enough to order. When we compare dedicated servers across vendors, the first problem is usually not the price: each vendor publishes a different subset of the hardware specification, so the fields you need are present for some rows and missing for others.
Benchmarking settled this long ago. SPEC’s run rules for CPU 2017 require, for results used in public, “a full disclosure of results and configuration details sufficient to independently reproduce the results,” because “a published SPEC CPU 2017 result is a declaration that the observed level of performance can be obtained by others.” A hosting product card is marketing, not a benchmark submission — but a ranking built from product cards is a ranking of marketing copy.
One variable reorders any dedicated-server comparison: which resource your workload saturates first, and whether the provider publishes that resource’s number. A single-threaded game loop saturates one core’s clock. A video packager saturates memory bandwidth and cache. A storage node saturates the uplink. A CI fleet saturates nothing for twenty hours a day and everything for four. Pick the wrong spec and you pay for cores you cannot use; pick a provider that does not publish the number your workload depends on and you cannot tell in advance.
What each provider actually publishes
| Provider | CPU model named | Base / boost clock | TDP | L3 cache | Setup fee stated | Notice period stated |
|---|---|---|---|---|---|---|
| Hetzner (AX / EX) | Yes | No | No | No | Yes, per model | Yes on AX42: “Cancellation period: immediately” |
| OVHcloud (Advance / Scale / Game) | Yes | Yes, as base / boost | No | No | Yes, per model | Not on the product pages |
| Hivelocity | Yes, in the model string | Yes, in the model string | No | No | Yes: no setup fee | No |
| phoenixNAP Bare Metal Cloud | Yes on current generations | Partly | No | No | No setup fee; billing-model page instead | Billing terms rather than a notice period |
Read from each vendor’s own public pages, November 2026.
OVHcloud is the most complete of the four on CPU disclosure. Its listings carry the model with base and boost frequency — “AMD EPYC 4464P 12 c / 24 t 3.7 GHz / 5.4 GHz”, “AMD RYZEN 7 9800X3D 8 c / 16 t 4.7 GHz / 5.2 GHz” — and its game servers even publish a per-SKU “CPU score”, though without any stated methodology, so it is not a number you can hold anyone to. Hivelocity puts the clock into the model string (“E-2236 3.4GHz”, “7965WX 5.4Ghz Threadripper”). Hetzner’s AX comparison publishes model names, architecture and core counts, but no clock frequency for any model.
Recovering the number the host leaves out
Naming the exact CPU model is the disclosure that matters most, because it makes every other field recoverable from the manufacturer. Hetzner’s AX42 page names an “AMD Ryzen 7 PRO 8700GE” and publishes no clock; AMD’s page for that part lists a 3.6 GHz base clock, up to 5.1 GHz boost, 16 MB of L3 and a 35 W default TDP.
| CPU | Where it is sold | Cores / threads | Base clock | L3 cache | TDP |
|---|---|---|---|---|---|
| AMD Ryzen 7 PRO 8700GE | Hetzner AX42 | 8 / 16 | 3.6 GHz | 16 MB | 35 W |
| AMD Ryzen 7 9800X3D | OVHcloud GAME-1 2026 | 8 / 16 | 4.7 GHz | 96 MB | 120 W |
| Intel Xeon 6527P | OVHcloud Advance-5 2026 | 24 / 48 | 3.0 GHz | 144 MB | 255 W |
Sources: AMD and Intel specification pages.
The first two rows are both “8 cores, 16 threads” on a hosting card. One has 30.6% more base clock, six times the L3 cache and 3.4 times the thermal envelope. None of those numbers appears on the hosting pages that sell them.
Two consequences follow. A provider that names the exact CPU model gives you more than one that quotes “8 cores / 32 GB”, even if its page looks thinner. And a low-power part such as the 8700GE can be a good buy for bursty web workloads and a poor one for sustained all-core compression or transcoding — the manufacturer tells you which, the host does not.
The setup fee, as a share of your tenancy
The one-time fee is the second place a provider comparison reorders, and the one buyers model worst. The comparable figure is the entry premium: setup fee ÷ monthly rate ÷ the months you will actually keep the server. It is a ratio, so it survives currency and hardware differences. The method is worked through in Hetzner vs OVH.
| Provider | Setup fee as a share of one month | Over 12 months | Notes |
|---|---|---|---|
| Hetzner, standard AX / EX range | about 50% | about 4.2% | Fee listed per model on the 15 June 2026 price-adjustment page |
| Hetzner, -LTD range | 0% to about 50% | 0% to about 4.2% | Mostly a flat €39; €0 on two models, €49–59 on the largest |
| OVHcloud, most Advance and Game listings | 100% | 8.3%, or 0% with a 12-month commitment | Installation fee equals one month; the US Advance page says it is included on commitments of 12 months or longer |
| Hivelocity | 0% | 0% | “Hivelocity does not charge a setup fee for dedicated servers” (custom builds may differ) |
| phoenixNAP Bare Metal Cloud | 0% | 0% | “No contractual obligations or pre-payments are required” |
| INXY.hosting | 0% | 0% | No setup fee and no minimum term, EU and US locations |
Hetzner’s half-month ratio is stable across its AX and EX models, in both the euro and dollar columns of its price-adjustment page. The -LTD range inverts the shape: because most of its fees are a flat €39, the AX42-1-LTD pays about 50% of a month while the AX162-1-LTD pays about 12% — on a flat-fee range, the entry premium falls as the server gets bigger.
On OVHcloud, the waiver lives in the FAQ rather than on the price card: the card prints the installation fee with no condition, while the FAQ states that “the setup fees for your dedicated server are included for any commitment period of 12 months or longer.” Model a twelve-month tenancy from the price card alone and you overstate the entry cost; model a three-month proof of concept from the FAQ and you understate it.
One more caution: the monthly rate you divide by depends on which page you read. Hetzner’s AX comparison shows the AX42 “starting from” €79.00 with a second type at €99.00; the product page shows €99.00; the price-adjustment table lists AX42-1 at €97.30 excluding IPv4. A quote should cite its page, its date and whether an IPv4 address is included.
Nobody prices the exit
On a no-commitment provider, the risk is not committing — it is forgetting. phoenixNAP’s billing documentation is unusually clear: “The minimum billing period for a reservation is one month.” “Deleting a server does not cancel the billing for your reservation.” And if a reservation is not renewed or cancelled before it expires, “your billing changes to the hourly pricing model,” and “you will be charged by the hour until you delete your server.” Auto-renewal is on for one-month reservations and off for 12, 24 and 36-month ones — so the longest commitments are the ones that fall back to hourly rates on a server that is still running. Public IPs are billed separately, at $0.0035 per hour each — about $2.56 a month, $30.66 a year.
Hetzner sits at the other end: the AX42 page states “No minimum contract term” and “Cancellation period: immediately,” while its cancellations overview notes that “You can cancel some products immediately, while others have fixed periods such as ‘30 days to the end of the month’” and that “these periods depend on the product and the interface.” The notice period is a per-product fact that is not on the price card — and it is what prices your rollback if you ever have to leave. See the datacenter migration plan.
The 2026 refresh: more money, not always more clock
OVHcloud’s US price page lists its 2024 and 2026 Advance generations side by side. Across five listings built on the same AMD EPYC 4000-series family, the 2026 generation is priced 17.8% to 45.3% above its 2024 namesake, while the published base clock moves between −8.1% and +2.6%:
| Model | 2024 price | 2026 price | Price change | Base clock 2024 → 2026 |
|---|---|---|---|---|
| Advance-1 | $147 | $204 | +38.8% | 3.8 → 3.9 GHz |
| Advance-2 | $198 | $255 | +28.8% | 3.8 → 3.8 GHz |
| Advance-3 | $255 | $332 | +30.2% | 3.7 → 3.4 GHz |
| Advance-4 | $325 | $383 | +17.8% | 4.2 → 4.3 GHz |
| Advance-STOR | $338 | $491 | +45.3% | 3.8 → 3.8 GHz |
Advance-3 is the row to look at: 30.2% more money for 8.1% less base clock, with boost and core count unchanged. These are different CPUs with different memory and storage, so this is not a price rise on the same product — but the generation label does not tell you which way the number your workload cares about moved. For more on 2026 pricing, see the dedicated server pricing guide.
What breaks in production
Buying core count for a single-threaded workload. A 16-core server with a lower base clock can lose to an 8-core one with a higher clock on every tick. The arithmetic is in our guide to dedicated servers for game hosting.
Treating a 35 W part like a 120 W one. Low-power CPUs are fine for bursty work and weak for sustained all-core load.
Assuming the port speed is the traffic allowance. “1 GBit/s port”, “3Gbps unmetered and guaranteed public bandwidth” and a monthly transfer allowance are three different promises — see metered vs unmetered bandwidth.
Letting a long reservation expire. On phoenixNAP, auto-renewal is off for 12, 24 and 36-month reservations, and the fallback is hourly billing on a running server.
Quoting a “starting from” price. On a configurable range, a floor is not the price of the server you will get.
Comparing a model name across generations. OVHcloud’s Advance-5 is an AMD EPYC 8224P at $504 in the 2024 listing and an Intel Xeon 6527P at $731 in the 2026 one — different vendor, same name.
Which provider, for which workload
| Workload | Strongest fit from published pages | Why | Where it loses |
|---|---|---|---|
| Single-thread-bound (game loops, older PHP monoliths) | OVHcloud Game range | Base and boost clocks published on every listing; 4.7 GHz base on GAME-1 | Installation fee equal to a month unless you commit for a year; 1 Gbps public bandwidth |
| Steady web and app fleets, cost-led | Hetzner AX / EX | About half a month’s setup fee, immediate cancellation on many products, unlimited traffic on 1 Gbit uplinks | No clock, TDP or cache published; EU locations only |
| US-located, clock-visible | Hivelocity | Clock in every model string; no setup fee | No contract terms or overage rates on the listing page |
| Burst and short-lived capacity (CI, batch, load tests) | phoenixNAP Bare Metal Cloud | Hourly billing, no contractual obligations | Lapsed reservations fall back to hourly billing silently; public IPs billed separately |
| Uncertain tenancy, EU or US | INXY.hosting | No setup fee, no minimum term, SLA; multiple providers on one contract | Configurations are quoted to your workload rather than listed on a public price card |
How to choose, in seven steps
Name the resource your workload saturates first: one core’s clock, memory bandwidth, IOPS, uplink — or nothing most of the time. One answer, not three.
Measure it on what you run now, at p99, not the mean. A four-hour peak is the specification.
Filter the shortlist to providers that publish that number, or that name the exact CPU model so you can recover it.
Recover the missing fields from the manufacturer: base clock, boost, cache and TDP from AMD’s or Intel’s own page, not a review site.
Compute the entry premium for your real tenancy — at three months and at twelve. The ranking often changes.
Get the exit terms in writing per product: notice period, what happens when a reservation expires, and whether deleting the server stops the bill.
Run the shortlist in parallel for a month before moving production. Two servers for a month cost less than one wrong server for a year.
The decision
Is your workload bound by a number any provider publishes? If yes, the shortlist is whoever publishes it. If not — if you are bound by memory bandwidth, cache or sustained thermals — the shortlist is whoever names the exact CPU model.
How long will you keep the server? Under six months, the setup fee dominates and no-fee providers win regardless of hardware. Over a year, it is worth roughly 4–8% and the hardware decides.
Who pays if you are wrong? That is the notice period and the reservation-expiry behaviour — the only part of this comparison that appears on no price card.
INXY.hosting buys dedicated capacity across multiple providers rather than selling one vendor’s range, so there is no single ranking to defend. Send us the p99 figure for the resource from step 1 and the tenancy you are willing to commit to, and we will build the shortlist against your workload — see dedicated and bare-metal hosting.
FAQ
What makes one dedicated server provider better than another?
Not its position on a listicle. A provider is better for your workload when it publishes the specification your workload is bound by, its setup fee is low relative to your tenancy, and its notice period is clear per product. Those three facts are knowable before you buy.
Why don’t providers publish CPU clock speeds?
Some do — OVHcloud and Hivelocity show base or boost clocks on their listings. Others publish only the CPU model. Where the model is named exactly, clock, cache and TDP can be recovered from AMD’s or Intel’s specification pages, which is why an exact model name is worth more than a spec summary.
Is a higher core count always better on a dedicated server?
No. Workloads bound by a single thread — many game loops, older application monoliths, some database write paths — are limited by one core’s clock. Two “8-core” servers on the market today differ by 30% in base clock, and an 8-core CPU can out-clock a 16-core one from the same family.
How much does a setup fee really cost?
Divide it by the monthly rate, then by the months you will keep the server. Hetzner’s standard range runs about half a month, roughly 4% over a year. An OVHcloud installation fee equals a full month, 8.3% over a year, unless a 12-month commitment waives it. Hivelocity, phoenixNAP and INXY.hosting charge none.
Can I compare dedicated server providers on price per core?
Only within one provider and one generation. Across providers, two “8-core” servers can differ six-fold in L3 cache and 3.4-fold in power envelope, so price per core compares different products. Price per unit of the resource you actually saturate is the comparable figure.
