VDS と VPS ホスティング – 違いはなんですか? - JA

同じコア数の VPS と VDS の違いは 1 つあります。: 他に誰が CPU 時間を使用できるか. VDS ではユーザーのために予約されています, while a VPS shares it with other customers on the same host.

Memory isn’t always the dividing line. DigitalOcean’s docs say RAM, disk storage and network bandwidth are “always dedicated” on its shared and dedicated plans alike. そこには, the premium buys CPU time and nothing else.

None of the top-ranking guides puts a number on that premium, and it varies a lot. We priced the closest same-core pair at five hosts that sell both kinds on 19 10月 2026. The dedicated version costs 1.50x to 1.75x as much at Akamai, Vultr and DigitalOcean, 2.21x at Hetzner and 6.53x at Contabo.

And a dedicated vCPU (仮想CPU, the unit hosts sell processor power in) usually isn’t a whole core. Hetzner and DigitalOcean both define it as one hardware thread, one of the two that a modern core runs at once.

One naming warning: a few sellers use “VDS” for an entire physical machine. 私たちの 専用サーバーと VDS の比較 covers that other meaning.

最終レビュー済み: 10月 2026. 価格と機能を検証しました.

Every price below comes from the provider’s own pricing page or docs. We opened no accounts and bought nothing. We ran the steal-time commands on 18 10月 2026 (UTC) in a local Docker container, using Ubuntu 24.04.5 with procps-ng 4.0.4 and sysstat 12.6.1.

What a Shared vCPU Lets You Use

Every plan in this guide is a virtual machine (VM) on a hypervisor, the software layer that splits one physical server between customers. The two plan types differ only in how that layer hands out processor time.

On a shared plan, your vCPUs get slices of time on cores that other VMs also use. When the host is quiet, you can run flat out. When neighbors get busy, you wait your turn. DigitalOcean is blunt about it: a shared Droplet “could receive fractions of hyper-threads” when the Droplets around it are under load.

Each seller words its rules differently, so here they are side by side, quoted from their own pages:

Akamai’s 80% is the only hard limit any of these five publishes. It’s a fair rule for all of them: shared vCPUs are rented for bursts, not for a processor pinned at 100% all afternoon.

Lightsail shows what happens when you ignore that. Its docs warn about a server that stays in the “burstable zone” 長すぎる. これ “will eventually use up all the CPU capacity until it operates in the sustainable zone again”. On the USD 7 Linux plan, that zone starts at 10% of the CPU. ヘッツナーの “ベースライン” wording points at the same model, though we found no figure in its docs.

Memory is where the sellers part ways. Contabo sells “Dedicated CPU and RAM” on its VDS line. Its VDS page also says a VPS “shares the physical server’s CPU, メモリー, 保管所, およびその他のリソース”. So DigitalOcean’s reserved RAM holds for DigitalOcean. Don’t assume it everywhere.

Two older ideas don’t fit these plans either. Some guides tie VPS to OpenVZ containers and VDS to KVM virtual machines. At DigitalOcean, ヘッツナー, Akamai and Vultr, the providers’ own pages call both lines virtual machines, so the split is CPU scheduling, not the virtualization type.

“管理” is a separate axis too. It decides who patches and monitors the operating system, not how the CPU is shared. DigitalOcean, 例えば, says its Droplets leave “managing the operating system, アプリケーション, およびデータ” to you on every plan type.

Does a vCPU Equal a Core?

通常はそうではありません. A modern server core runs two instruction streams at once (Intel calls this Hyper-Threading, AMD calls it simultaneous multithreading, or SMT). Most hosts count each stream as one vCPU. Two vCPUs can therefore sit on one physical core.

Hetzner says so outright in its FAQ: 専用プランで “one vCPU equals one thread of a physical CPU core”. DigitalOcean defines a vCPU as “a single hyper-thread on a processor core”. Its dedicated plans guarantee “the full hyper-thread”, not the full core.

Contabo’s own page confirms the pattern from the other side. The Cloud VDS S card lists “6 Virtual Cores” AMD EPYC 上 7282. The structured data in the same page’s code still describes the plan as “3 Physical CPU Cores, 24 GB RAM, 180 GB NVMe SSD”. 7月に 2026 the visible card said 3 physical cores too. Same RAM, same disk, same port: six threads on three reserved cores.

Akamai is the outlier in wording. Its plan guide says Dedicated CPU Linodes “reserve full physical CPU cores”. We can’t check that from outside, and the guide doesn’t say how a vCPU maps onto a core’s two threads. Vultr’s pricing page says “fully dedicated” vCPUs and leaves the unit undefined.

なぜこれが重要なのか? なぜなら “6 専用コア” on one sales page and “6 専用の vCPU” on another can mean three physical cores or six. Ask before you compare.

The Price of Isolation: Same Core Count, Two Prices

Here’s the number the ranking pages leave out. For each host that sells both kinds, we took the closest pair with the same vCPU count and divided each price by its vCPUs.

The per-vCPU columns are our calculation (plan price divided by vCPUs). DigitalOcean, Vultr and Akamai prices are as listed, 税金明細が表示されていない場合. Hetzner charges EUR 0.50 a month for an IPv4 address on top of both plans, which nudges its multiplier to 2.18x. Contabo’s month-to-month prices (ユーロ 7.50 およびユーロ 49.00) give the same 6.53x.

Three patterns stand out. The US clouds sit in a tight band. All three charge USD 24.00 for the shared plan and USD 6.00 米ドルへ 9.00 more per vCPU for dedicated. Akamai’s G6 pair is the cleanest test, because RAM, disk and transfer are identical. 紙の上, only the CPU type changes.

Hetzner and Contabo give the dedicated plan twice the RAM, so part of their gap pays for memory, not isolation. That extra RAM is also why Hetzner’s EUR 11.75 premium per vCPU is the largest step in the table.

Contabo is the odd one. Its multiplier is the biggest, yet its dedicated vCPU is the cheapest in the table, because its shared vCPU costs just EUR 1.00. 小さな文字を読む, けれど. That rate needs a 24-月前払い, the card says it holds “初めて 24 月”, and the chip is a 2019-generation EPYC 7282.

Contabo also prices the trade at one exact figure. ユーロ 39.20 on the 24-month term buys either Cloud VPS 18 (18 共有vCPU, 96 GB) or Cloud VDS S (6 専用の vCPU, 24 GB). That tie is new, けれど, 次のセクションで示すように.

Hardware isn’t always equal inside a pair either. DigitalOcean’s docs place CPU-Optimized on “Intel’s Ice Lake and older processors”. Regular Basic runs on “a mix of first generation or older Xeon Scalable processors and AMD EPYC processors”. Choosing between DigitalOcean, Vultr and Akamai on more than the CPU? 私たちの クラウドホスティングのまとめ compares the three on far more than this one trade.

二 2026 Price Changes That Widened the Gap

の上 15 六月 2026, Hetzner raised the monthly price of its dedicated CCX13 from EUR 15.99 ユーロです 42.99. The shared CPX22 went from EUR 7.99 ユーロです 19.49, so the same-core premium grew from 2.00x to 2.21x. Hetzner’s docs say the new prices applied to new orders and rescales from 8 AM CEST that day.

Contabo’s change was quieter. 間 31 August and 16 9月 2026, Contabo raised Cloud VDS S from EUR 39.00 ユーロです 49.00 月, 除く. バット. それは 25.6% 上昇, and the 24-month rate moved from EUR 31.20 ユーロです 39.20.

The shared side didn’t move. クラウドVPS 6 showed EUR 7.50 (ユーロ 6.00 の上 24 月) in every archived copy of Contabo’s VPS page from 29 8月～ 11 10月. So Contabo’s same-core premium went from 5.20x to 6.53x, all of it from the dedicated side. The dates come from Wayback Machine copies of the page and our own record of it on 16 9月.

Only the dedicated side of Contabo’s pair moved, so its premium grew far more than Hetzner’s. ソース: Hetzner Docs price-adjustment page; Contabo pricing pages and Wayback Machine copies, 読んだ 19 10月 2026.

That rise also created the tie from the last section. の上 31 August the six-thread VDS cost EUR 31.20 の上 24 月, less than the EUR 39.20 クラウドVPS 18. Any guide still quoting the old VDS price understates today’s gap by about a fifth.

How to Check CPU Steal Time on Your Server

Steal time is the share of time a vCPU was ready to run while the hypervisor gave the physical CPU to another VM. The Linux kernel counts it, and the standard monitoring tools show it. If you’re on a shared plan and your site slows down, this is the first number to read.

We ran every command below in a fresh Ubuntu 24.04 container to capture real output. A container on a laptop has no paying neighbors, so every steal figure reads zero. The output shows you where to look, not what your server will say.

ステップ 1: Read the st value in top

top -bn1 | head -5

top - 23:21:36 up 6 days, 23:57,  0 user,  load average: 1.06, 1.01, 0.97
Tasks:   4 total,   1 running,   3 sleeping,   0 stopped,   0 zombie
%Cpu(s):  0.9 us,  0.9 sy,  0.0 ni, 97.2 id,  0.0 wa,  0.0 hi,  0.9 si,  0.0 st
MiB Mem :   7836.4 total,    507.1 free,   2576.4 used,   4952.1 buff/cache
MiB Swap:   1024.0 total,   1024.0 free,      0.0 used.   5260.0 avail Mem 

The last value on the %Cpu(s) ライン, st, is steal. Two neighbors on the same line matter as well: us is your own programs, と wa is time spent waiting for the disk. 追加 -1 (top -1 -bn1) to get one line per vCPU instead of an average. The top manual defines st なので “time stolen from this vm by the hypervisor”.

ステップ 2: Watch it over time with vmstat

vmstat 1 5

procs -----------memory---------- ---swap-- -----io---- -system-- -------cpu-------
 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st gu
 1  0      0 519040 424136 4646784    0    0     2    21  732    0  0  0 100  0  0  0
 0  0      0 519040 424136 4646784    0    0     0     0 1078 1382  0  0 99  0  0  0
 0  0      0 519040 424136 4646784    0    0     0     0  975 1314  0  0 100  0  0  0
 0  0      0 519040 424136 4646784    0    0     0     0 1284 1773  0  0 99  0  0  0
 0  0      0 519368 424136 4646784    0    0     0     0  480  497  0  0 100  0  0  0

Those two numbers mean “毎 1 2番目, 5 時間”. Steal is the st 桁. Skip the first line, なぜなら、 vmstat manual says it “gives averages since the last reboot”. The procps-ng 4.0.4 we ran adds a gu 桁 (guest time) 後 st, so don’t take the last column for steal.

ステップ 3: Get a cleaner percentage with mpstat and sar

Both come in the sysstat package. On Debian or Ubuntu, install it first (prefix the command with sudo if you aren’t root):

apt-get install -y sysstat

mpstat 1 3

Linux 6.12.76-linuxkit (c8bc221c6626) 	10/18/26 	_aarch64_	(10 CPU)

23:21:40     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
23:21:41     all    0.10    0.00    0.00    0.00    0.00    0.10    0.00    0.00    0.00   99.80
23:21:42     all    0.10    0.00    0.20    0.00    0.00    0.00    0.00    0.00    0.00   99.70
23:21:43     all    0.40    0.00    0.00    0.00    0.00    0.20    0.00    0.00    0.00   99.40
Average:     all    0.20    0.00    0.07    0.00    0.00    0.10    0.00    0.00    0.00   99.63

The mpstat manual describes %steal なので “involuntary wait by the virtual CPU or CPUs while the hypervisor was servicing another virtual processor”. 走る mpstat -P ALL 1 1 to see each vCPU separately, which catches one starved vCPU hiding inside a healthy average.

sar -u 1 3

Linux 6.12.76-linuxkit (c8bc221c6626) 	10/18/26 	_aarch64_	(10 CPU)

23:22:03        CPU     %user     %nice   %system   %iowait    %steal     %idle
23:22:04        all      0.00      0.00      0.00      0.00      0.00    100.00
23:22:05        all      0.20      0.00      0.30      0.00      0.00     99.51
23:22:06        all      0.10      0.00      0.10      0.00      0.00     99.80
Average:        all      0.10      0.00      0.13      0.00      0.00     99.77

For a long record, sar is the handiest layout. Raise the two numbers (interval in seconds, then count) and leave it running in a terminal across your busiest hour.

ステップ 4: Read the raw counter

grep '^cpu ' /proc/stat

cpu  1363003 0 644389 601499642 17486 0 435816 0 0 0

Steal is the eighth number, counted in clock ticks (USER_HZ) since boot (proc_stat manual). Ours is 0 after almost seven days of uptime, which is what a machine with no neighbors looks like. On a shared host, read it twice a few minutes apart: a number that keeps climbing means steal is happening now.

What the numbers mean

Contabo’s own benchmarking guide (23 7月 2026) gives usable bands. It calls 0-1% “typical of a genuinely dedicated allocation” と 1-5% “in occasional spikes” usually fine. It calls 5-10% sustained “measurable contention worth investigating”, 以上 10% sustained a sign that “ひたむきな” CPU “isn’t behaving like one”. Contabo sells both lines, so read that as a seller’s yardstick, but it’s a sensible one.

Measure at your busy hour. A single quiet sample at 3 午前. proves nothing either way.

Steal isn’t the only reason a VPS feels slow, and buying dedicated vCPUs for the wrong cause wastes money. In one Reddit thread from May 2026, a user on a roughly EUR 5 Contabo plan reported steal at 0 全体を通して. The slowdowns came from the disk: “iowait around 50%” and a load average near 16 の上 4 vCPU. A reply in the thread says moving to another host fixed it. Match the symptom to the cause:

What Is a VDS Used For? The Thresholds That Decide It

A VDS earns its premium when CPU demand is steady and delay costs you something. Short of that, a shared plan does the same work for less. These thresholds draw on the providers’ own rules above.

Shared vCPUs are enough when:

• Your CPU graph averages well under Akamai’s 80% ライン, and peaks last minutes, 数時間ではない.
• The work is the kind every seller lists for shared plans: ブログ, a CMS, トラフィックの少ないサイト, 小規模なデータベース, staging and dev servers.
• Steal stays at or under 5% during your busy hour.
• Your slowdowns show up as wa (ディスク) それよりも st (steal).

Dedicated vCPUs are worth the money when:

• A job holds the CPU near 100% for long stretches. Akamai names CI/CD builds (automated code builds and tests) and video transcoding, and Vultr lists video encoding and batch processing for its CPU Optimized line.
• Latency is the product: ゲームサーバー (on Hetzner’s and Vultr’s dedicated lists), チェックアウト, or a real-time API.
• Steal sits above 5% at peak, or spikes past 10%, and the slow moments line up with it.
• Average CPU use runs above 80% シェアプランで, which breaks Akamai’s stated rule and lives on borrowed burst time elsewhere.

選び方: Shared or Dedicated vCPU

Start from your workload and budget, then check one number on your current server. Five common cases:

• バジェット: a WordPress site or small store under USD 25/month, CPU averaging under 50% -> a USD 24.00 shared plan at Akamai, Vultr または DigitalOcean. 同じ 2 vCPUと 4 GB on a dedicated plan cost USD 36.00 米ドルへ 42.00, and steal won’t be your bottleneck.
• ワークロード: CI runners or video encoding that hold 100% 為に 20 minutes or more a day -> Akamai G6 Dedicated 4 GB（USD） 36.00. DigitalOcean’s same-spec CPU-Optimized plan costs USD 42.00 and comes with 25 GiB of disk instead of 80 GB.
• レイテンシー: a game server or real-time API with steal above 5% at the evening peak -> a dedicated plan near your users. In Europe that’s Hetzner CCX13 at EUR 42.99 と 8 GB; 他の場所で, Vultr’s CPU Optimized 2 vCPU plan at USD 40.00.
• Memory first: ネクストクラウド, a big cache or dozens of small sites on about EUR 40/month -> Contabo Cloud VPS 18. それは与える 18 shared vCPUs and 96 GB = ユーロ 39.20 の上 24 月. Cloud VDS S costs the same but gives 6 専用の vCPU と 24 GB, so it only wins when the CPU is the bottleneck.
• Symptom: a slow server where st is near 0 と wa is high -> don’t buy dedicated vCPUs. Ask for a move to another host or upgrade the storage first.

Planning to start shared and upgrade later? Check the move first. Akamai lets you switch plan type when resizing, and Hetzner’s FAQ says you can rescale to and from shared plans. DigitalOcean only resizes to plans with equal or larger disk. これは 80 GiB Basic Droplet can’t move to the 25 GiB CPU-Optimized plan with the same cores.

よくある質問

Is a VDS faster than a VPS?

Not on a quiet host. A dedicated vCPU runs on the same kind of silicon, and what you buy is the guarantee when neighbors get busy. Sometimes the dedicated line is even the older hardware: Contabo’s VDS runs an EPYC 7282, while its shared Performance VPS line uses newer EPYC chips.

Do I need a VDS for a WordPress site?

めったに. DigitalOcean, Akamai and Vultr all list blogs and low-traffic sites as shared-plan work, and page caching keeps WordPress CPU use low. The exception is a busy WooCommerce checkout where top shows steal above 5% at peak. そこには, 米ドル 6.00 米ドルへ 9.00 premium per vCPU at the US clouds is worth testing.

は 2 dedicated vCPUs better than 4 shared ones?

For steady heavy load, 通常ははい; for bursty load, usually no. At DigitalOcean, 米ドル 42.00 買う 2 dedicated vCPUs with 4 GiB, while USD 48.00 買う 4 shared vCPUs with 8 GiB. Short spikes favor the four shared vCPUs and double RAM. Hours at 100% favor the two guaranteed threads.

Why does top show steal on my dedicated vCPU plan?

Small spikes can appear even on reserved CPUs. Contabo’s guide calls occasional 1-5% “usually fine”, and sustained steal above 10% a CPU that isn’t behaving like a dedicated one. If you see that, log mpstat -P ALL across a busy hour and send it to support. アカマイ, 一つのために, says its dedicated plans “remove CPU contention entirely”.

結論

Buy a shared-vCPU VPS for blogs, stores with modest traffic, staging and anything that bursts. で アカマイ, Vultr と DigitalOcean the same USD 24.00 あなたを手に入れます 2 vCPUと 4 GB, and steal is unlikely to be your limit.

Pay for dedicated vCPUs when the CPU sits high for long stretches or steal tops 5% at your peak. Akamai’s G6 pair is the cheapest isolation among the US clouds at 1.50x. ヘッツナー costs 2.21x since June but adds RAM.

コンタボ is a special case. Its dedicated threads are the cheapest per vCPU here. But that price needs a 24-month prepay, runs on older chips and follows a 25.6% rise that hit the VDS line alone.

Whichever way you lean, 走る vmstat 1 5 on your current server at its busiest hour before you spend anything. Weighing a whole physical machine instead? The dedicated server vs VDS comparison linked at the top covers that step.