All systems operational 12 elite locations · Europe & Asia Pay with crypto · 0 KYC
Home / Guides / Self-Hosting Nextcloud on a VPS: Size It Right
Guide

Nextcloud on a no-KYC VPS — files on your own disk

Nextcloud VPS on VPSbit - self-host files on a KVM box

A two-person instance fits 4 GB; preview generation wants 32 GB. Official install on your own NVMe, Titan at $19.20/mo annual-eff, paid in XMR.

Short answer

A two-person instance fits 4 GB; preview generation wants 32 GB. Official install on your own NVMe, Titan at $19.20/mo annual-eff, paid in XMR.

Self-hosting swaps a subscription for a job title

The pitch for self-hosting usually implies the monthly bill disappears; it does not, it changes shape. You trade a per-seat subscription for a rented slice and a title nobody warned you about: administrator. The platform's side of the invoice is the KVM machine, the unmetered port, and the IP — installs, updates, backups, and any mail the domain attracts sit on your side of the line, and Nextcloud's own admin manual is written for exactly that reader.

What you host is the upstream product, unmodified: files, calendars, contacts, share links, and optional Talk for calls, behaving as the project ships them. Teams point a domain at the slice so data lands on a disk they control, with previews, versioning, and federation working as documented, and the data directory wherever you pointed it — /var/www/nextcloud/data by default. Billing clears in any of 19 coins, and there is no managed-Nextcloud tier anywhere in the catalog pretending otherwise: root is the product.

The honest ledger has two columns. Self-hosting buys: no per-user fee that grows seat by seat, no scan of your documents by an advertising company, no migration hostage situation when a provider changes its terms. Self-hosting costs: patch nights, capacity planning, and an abuse mailbox that now has your domain on it. Families and small teams usually find the trade favorable precisely because the costs are scheduled work rather than surprises — but schedule them, because the box will not raise its hand when something is due.

The classic mistake is importing consumer-cloud expectations along with the files. A rented slice does not silently fix your schema, rename your photos, or email you a monthly summary of what it did for you. Decide from day one that updates live on your calendar, restore tests live on your calendar, and the snapshot button is your colleague rather than your janitor. The rest of this guide is that calendar, expanded into the order the work actually happens.

The first calendar entry is sizing, and sizing an instance is mostly a conversation about RAM rather than disk — which is why the plan ladder matters more than the storage headline. The next section leads with the memory numbers, because they are the ones that decide whether the instance you install is the one you keep.

VPS from $4.80* — Core on the annual term (list $6/mo), one invoice for 12 months, no auto-charges. Dedicated with IPMI from $39.20/mo. Crypto invoice, no KYC.

Launch now

RAM, not disk, decides which plan holds the instance

Storage is the number everyone multiplies and the number that matters least. A photo library of 200 GB sits quietly on any plan's NVMe; what hurts the instance is the moment several users upload at once and the preview generator starts turning JPEGs and video into thumbnails, which is a CPU-and-RAM problem wearing a storage costume. Size the plan for the thumbnail work first and the archive second, and the instance will feel fast long after the library stops fitting on a laptop.

The ladder, stated plainly: Core carries 2 vCPU, 4 GB of DDR5 ECC, and 80 GB of NVMe at $6/mo list, and it holds a calm two-person instance without complaint. Prime's 4 vCPU and 8 GB across 160 GB is the stated floor once previews fire for several users. Titan — 8 vCPU, 32 GB ECC, 320 GB NVMe at $24/mo list, $19.20/mo annual-eff — is the practical ceiling of the VPS line and the comfortable answer for a household or a small team. Start a two-person instance on Core at $6/mo list — $4.80/mo annual-eff — and move up only when the thumbnail queue tells you to, not when the marketing does.

Memory math that survives contact with production: the database wants a stable quarter of RAM, PHP workers eat a few dozen megabytes each under load, and preview generation for a burst of phone videos will spike hard until preview:generate-all drains the queue — put it on cron rather than on-demand so the spike lands where you scheduled it. Watch the truth with free -h during an upload party, not after it. Instances rarely die of disk; they stagger on swap and time out on PHP while the disk still has half its capacity showing.

The mistake is buying disk you cannot use because RAM starves the app: a 320 GB library on a 4 GB slice generates previews at the speed of apology and times out the web UI under the very upload burst the storage was bought for. When archives grow toward terabytes, that is dedicated-disk territory — 2 x 2 TB on Summit or 2 x 1 TB on Forge — while the app stays on a RAM-appropriate machine, or everything moves to metal together in one planned evening.

Why the virtualization floor matters here: preview workers, cron, and a database all assume the memory they are told is theirs, and full virtualization is what makes the 8 GB honest instead of a burst pool that evaporates on the first busy evening. The survival story comes next — snapshots versus backups — and it is yours alone to run, on any tier you pick.

Snapshots cover a bad upgrade; your backup covers a dead building

Two different disasters wear similar costumes. A migration gone wrong corrupts the instance but leaves the building standing — a snapshot taken an hour earlier restores it in minutes. A dead host, a failed array, or a deleted account takes the snapshots down with the box, and no amount of snapshot discipline helps. The second disaster is the one people insure against too late, usually exactly once.

Snapshots ride the VPS plans as a checkout add-on, and they are the right tool for the upgrade ladder: take one before every version bump, roll back in minutes if occ upgrade chokes on something the release notes did not predict. Pair them with an encrypted pull to a second machine on a different network — borg create or restic backup against the data directory and a database dump — and the building-fire scenario gets its own exit. The Borg and Restic documentation both assume exactly this shape: encrypted, deduplicated, scheduled, and off-site.

The pull direction matters as much as the tool. A backup that the Nextcloud server can write to is a backup ransomware can encrypt from inside; a pull from a second machine whose keys the slice never sees cannot be destroyed by anything happening on the box. Schedule the dump with the database's own tools, ship it in the same pull, and keep a couple of generations — yesterday's mistake and last week's calm both deserve a restore point you can actually reach.

The mistake that ends in grief is the untested archive: a Borg repo that has been succeeding for a year but has never been mounted, with a passphrase written in a file on the same disk it protects. Restore to a scratch machine twice a year, time the run, and write the passphrase down somewhere that is not the server. Restores you have tested are the only metric that counts; an untested archive is a hope with a filename.

With sizing and survival settled, the install itself is almost anticlimactic — the official manual covers it end to end, and the main discipline is refusing clever shortcuts. That is the next section: which install path to take, and which forum one-liner to leave alone, forever, no matter how persuasive the thread looks at midnight.

The official manual install ages better than a clever one-liner

Nextcloud offers two documented paths: the classic PHP-plus-database install, and the All-in-One container image that bundles the stack behind its own management tooling. Both are maintained by the project, both are covered step by step in the admin manual, and both end somewhere you can debug at 2 a.m. The scripts that promise the same result in one command are maintained by nobody you can file an issue against.

The classic path is well-trodden: a supported PHP version, a database, the web server of your choice, then occ maintenance:install to lay down the config and the schema, and TLS on your own domain before the first login. Check state any time with php occ status and the built-in overview page. Put the data directory and the database on the local NVMe from the start — a KVM slice's disk is the fastest storage in its building, and the database feels every millisecond of anything slower within the first week of real use.

The All-in-One image trades a little transparency for a lot of convenience: containers for the database, the app, and the background workers, with backups and updates orchestrated by the project's own tooling. It is a good default for operators who do not want to learn PHP tuning; the classic install is the better teacher and the lighter footprint on small plans. Both run fine on 2 vCPU and 4 GB for a small instance — the sizing rules from the earlier section apply to either path without modification.

The mistake is mixing the two: an All-in-One install then hand-edited into the classic layout, or a forum snippet run on top of either path because it seemed faster. Pick one road, follow its manual, and keep every deviation written down in a file that lives outside the web root — /var/www is a good test of that rule. Future-you during an upgrade is the real audience for that file, and future-you does not answer tickets or remember the thread you half-read.

Whichever path you choose, the upgrades that follow it are the part that decides whether the instance is still yours in three years. The project moves at a steady cadence, and the operators who keep up are the ones who turned upgrade day into a ritual — which is exactly what the next section describes, step by step, in the order that never fails.

Upgrade day is a snapshot day, by rule, not by mood

Nextcloud releases move at a steady cadence, and each major version expects to be upgraded in sequence rather than leapt across. Treat upgrade day as a scheduled ritual with a fixed order — snapshot, maintenance mode, upgrade, checks, restore service — and the ritual turns version anxiety into a 30-minute window you have done a dozen times. Skip the ritual and every upgrade becomes a first flight with passengers aboard.

The order in full: take the snapshot; enter occ maintenance:mode --on so no user writes land mid-migration; run the updater and then occ upgrade; apply the housekeeping commands the release notes name, such as occ db:add-missing-indices; log in as an ordinary user and click through files, shares, and previews; then leave maintenance mode and watch the log for an hour. The maintenance and backup chapters of the official manual document this loop; the snapshot is what makes it calm instead of courageous.

Rollback is where the snapshot earns its keep. A failed migration on a snapshotted box is a ten-minute rewind and a calm decision about what to do differently on the next attempt; the same failure without one is an archaeology project on a production disk with users refreshing a broken page. Snapshot before every upgrade even when the release notes are boring — especially when they are boring, because the boring ones are how complacency gets installed alongside the new version.

The mistake is upgrading during the evening peak because the notification annoyed you, or stacking a server OS upgrade and a Nextcloud major in the same window. One variable per window, notifications batched into a monthly review, and the calendar entry named honestly — upgrade window, not quick update — keeps the ritual intact for years. Users forgive a planned Tuesday morning; they remember a surprise Thursday evening for exactly as long as they remain users.

Upgrades and backups are the maintenance half of the job; the other half is judgment the manual cannot make for you — where the box should physically sit, and which legal regime should hold the disk. Geography is the next section, and it is chosen once, deliberately, instead of inherited from wherever the first checkout click landed.

Pick the flag for the process you actually fear

Location marketing sells latency; location reality sells law. Every flag on the map bills the same base list price, and the differences that matter are the peering under the floor and the legal regime over the disk. Choose by asking which process you are actually defending against — a civil demand, a dragnet, a latency complaint from users — because each answer points at a different city, and only one of the three is about milliseconds.

The working shortcuts: Zurich when judicial process is the requirement, with Swiss FADP and a non-EU posture at a +35 percent premium; Reykjavik when you want no mandatory data retention under a Tier IV roof, at +20 percent; Amsterdam when the users sit in western Europe and round-trips through AMS-IX matter more than anything legal. Norway matches Iceland's premium and posture; Luxembourg adds a data-trust regime at +15 percent. The other eight flags bill at base list, and several of them are excellent.

The coin is independent of the flag, entirely: XMR, BTC, or any USDT network clears the same invoice in any of the 12 cities, and the payment step is covered fully in the guide to paying with Monero. Latency arithmetic is the only universal: from central Europe, Amsterdam and Frankfurt answer in single-digit milliseconds; from East Asia, Tokyo, Hong Kong, and Taipei do the same. Users notice 30 milliseconds on every page load; nobody notices a jurisdiction until the day they need it.

The mistake is choosing by aesthetic — the Iceland box for a team in Singapore, the Tokyo box for a family in Berlin — and paying the latency tax forever on a distinction nobody ever audits. Write the threat model down in one sentence, match the city to it, and revisit only when the model changes. Migration between flags is possible but it is real work; choosing once, carefully, is cheaper than migrating twice, casually.

Wherever the disk lands, the domain pointed at it inherits obligations the platform does not absorb: the abuse mailbox, the AUP, and the categories that are out everywhere. Running a public instance makes you a publisher of sorts — the next section is about the mail that proves it, and the habit that keeps it boring.

You inherit the abuse mailbox with the domain

Point a domain at a slice and the internet starts writing to it: bounce-back spam reports for the mail you did not know was relaying, phishing claims when a user's password dies, copyright notices when someone shares what they should not. The platform delivers the mail; reading and answering it is part of operating the instance, and the skill transfers directly from the guide to abuse tickets and DDoS filtering.

What the policy allows is wider than most consumers assume: legal adult content is permitted on these boxes, privacy tooling is normal traffic, and the prohibited list is short and absolute — CSAM, phishing, spam, and attack traffic, on every plan, paid with any coin, in all 12 locations. The AUP does not bend for crypto billing any more than it bends for cards; the snapshot schedule and the Borg job remain yours, and so does staying on the right side of the short list.

For a Nextcloud operator the practical exposure is usually shares, not the app: a public link to a 4 GB folder gets pasted into a chat, the chat screenshots it, and eventually a report lands with a URL in it. Answer with data — the share ID, the UTC timestamp, the permission you revoked — and the case closes on facts within a day. Keeping the file ledger current with occ files:scan after big migrations makes those answers instant instead of archaeological.

The mistake is silence: the abuse mail forwarded to an inbox nobody owns, the report that ages past the point where a factual answer helps anyone. Abuse desks escalate unanswered cases on a schedule that has nothing to do with your vacation plans. Put the abuse@ contact on the same monitoring as your backups, answer with fields and fixes, and the mailbox becomes routine correspondence instead of a standing hazard to the whole domain.

Compliance handled, the last attack surface is one you configure yourself, usually by leaving defaults alone: share links that never expire, previews of everything for anyone, background jobs still running on page loads. The hardening pass is short because the controls already exist — the next section walks it in the order that actually reduces risk first.

Share links are the attack surface you configure by default

The memorable Nextcloud incidents are almost never kernel exploits; they are a share link with no expiry and no password, sitting in a public chat log for two years. Default links are convenient and permanent, which is the exact combination you do not want on a box you administer yourself. The hardening pass is short because the controls already exist — the work is deciding your defaults and enforcing them in config, not writing code.

Set the floor in configuration: enforce passwords on public links, cap link lifetimes, and disable public shares entirely for instances that only serve a team. Move background jobs from AJAX to occ background:cron with a system cron entry — the background-jobs chapter of the manual shows the exact line — so previews, cleanup, and activity mail run on a clock instead of on page loads. Then point the preview generator at a schedule too, because a burst of phone videos should queue quietly, not stampede the workers.

The archive question returns at the storage layer. Photo libraries and backups grow into terabytes no matter what the founding document said, and when the slice's NVMe fills, the honest options are pruning or metal. Move the library to the dedicated Forge at $49/mo list — $39.20/mo annual-eff — and keep the app where it is, or consolidate everything on the same box with 2 x 1 TB of NVMe behind IPMI. Mount the new volume, migrate the data, run occ files:scan to rebuild the ledger, and cut over in a single planned evening.

The mistake is hardening once and never again — the pass is a habit, not an event. Quarterly: re-read the sharing defaults after every major upgrade, because new features ship new defaults; re-run the public-link audit; restore-test the backup; and confirm that cron actually fired on schedule. Thirty minutes a quarter is the entire price of not becoming a cautionary tale in somebody else's group chat, and it is the cheapest line item in this guide.

Everything above repeats a pattern worth naming: the platform removes the hosting-class risks, and the judgment calls remain yours by design. The final section compresses the entire job — order, install, operate — into the order you actually run it in, from an empty slice to a steady instance with no gaps left open.

The Nextcloud install, in the order to run it

The whole job fits a sequence, and the sequence starts before the installer does. Order the slice with the sizing rules from the RAM section — two calm users fit the entry tier, previews push you up the ladder — and set the snapshot policy the same day, because a snapshot taken before the first upgrade is the one that saves the instance. Pick the location for the legal regime you actually fear, not the flag that photographs best.

Second, install from the official manual — PHP plus a database, or the project's All-in-One image — and put occ maintenance:install at the center of that evening rather than a script from a forum. TLS on your own domain before the first login, the data directory on local NVMe, and background jobs moved to occ background:cron with a system entry while nothing is yet at stake. Everything configured before users arrive is a decision; everything configured after is a risk with witnesses.

Third, set the schedules that make the instance boring in a good way: the preview generator on a queue rather than on demand, background jobs on a clock, and the sharing defaults — passwords on public links, capped lifetimes — written into config rather than left as whoever-installed-last's settings. Audit public links on the same day each month so the habit survives calendar churn and staff changes.

Fourth, build the backup that survives the building: an encrypted restic backup or Borg pull from a second machine on a different network, scheduled, keyed outside the slice, and restore-tested to a scratch box twice a year. Snapshots cover upgrade day; the pull covers everything else. Time the restore, write down the passphrase somewhere that is not the server, and the archive stops being a hope with a filename.

Only then does data move in: rsync the library, dump and restore the database, copy the config, and repoint the domain once the fresh box passes a real login test. Snapshot the imported instance before the first user logs in, so the rollback path exists from minute one of production life. After that, the steady state is the quarterly loop from the hardening section — thirty minutes, four checks — and the instance stays yours in every sense that the subscription version never was.

Is Nextcloud managed on these plans?

No. Root is the product: you install, update, back up, and operate the instance, and the platform's job ends at the KVM slice, the network, and the invoice. There is no managed-Nextcloud upsell anywhere in the catalog, and no support tier that will log into your instance on your behalf for any price.

Can I pay for the instance with Bitcoin or USDT?

Yes — BTC and 18 other coins, XMR included, clear the same invoice page with no card step. The exact-amount rule applies to every coin: one transfer, the printed figure, a single-use address, and the box images in about a minute after the confirmations land. Pick the network your sending wallet actually supports before the invoice prints.

Is Core enough for a family instance?

For two or three calm users, yes — 2 vCPU, 4 GB of ECC RAM, and 80 GB of NVMe handle a family library quietly. Preview generation and a growing photo archive are the two things that push an instance toward Prime's 8 GB, and a multi-user household with video previews is the textbook case for the next rung up.

Where do previews and small files actually fit?

On CPU and RAM rather than disk, which is why multi-user instances with preview generation enabled want more than the entry tier. Keep the generator on a schedule if uploads arrive in bursts, keep the database on local NVMe, and size the plan for the thumbnail work rather than for the library's total size in gigabytes.

Can I migrate an existing Nextcloud install here?

Yes: rsync the data directory, dump and restore the database, copy the config, and repoint the domain record once the fresh box passes a real login test. Snapshot the imported instance before the first user logs in, so the rollback path exists from minute one, and run the file scanner afterward to rebuild the internal ledger cleanly.

Primary sources

More ways to pay

Pay with Bitcoin · Pay with Monero · Pay with USDT (TRC-20/ERC-20/SPL) · Pay with Litecoin · Pay with Ethereum · Pay with Solana · Pay with x402 · Pay with Dogecoin · Pay with XRP · Pay with TRX · Monero on bare metal · Dedicated servers with crypto · Pick the jurisdiction first.

Ready to launch?

Build the box — VPS or bare metal — create the password, pay the invoice that follows.