# VPS vs Dedicated Server: When to Switch | FYRHost
> CPU steal, ECC RAM, NVMe queues, IPMI — four numbers that decide. Sentinel tops out at $26.80 annual-eff; Shield metal starts at $30.80. Same 19-coin invoice.
Source: https://fyrhost.com/guides/vps-vs-dedicated/
[Home](https://fyrhost.com/) / [Guides](https://fyrhost.com/guides/) / VPS vs Dedicated Server: When to Switch Guide
# VPS vs Dedicated — The Four Numbers That Decide the Switch, With Both Ladders on One Crypto Checkout
 CPU steal, ECC RAM, NVMe queues, IPMI — four numbers that decide. Sentinel tops out at $26.80 annual-eff; Shield metal starts at $30.80. Same 19-coin invoice.
Published 2026-08-30 · Updated 2026-10-09 · FYRHost Editorial · 18 min read
**Short answer** CPU steal, ECC RAM, NVMe queues, IPMI — four numbers that decide. Sentinel tops out at $26.80 annual-eff; Shield metal starts at $30.80. Same 19-coin invoice.
Xeon E-2388G → 2× AMD EPYC 9474F ECC DDR4/DDR5 NVMe RAID 0/1/5/10 IPMI on all 1–25 Gbps unmetered 19 coins + XMR
## Plan ledger
| Plan | CPU | RAM | NVMe | RAID | Port | IPMI | Price |
|---|---|---|---|---|---|---|---|
| Shield | 8c / 16t | 32 GB DDR4 ECC | 2 x 500 GB NVMe RAID 0/1 | 0/1 | 1 Gbps | IPMI + DDoS | $38.50 list · $30.80/mo annual |
| Vault | 16c / 32t | 64 GB DDR4 ECC-R | 2 x 1 TB NVMe RAID 0/1 | 0/1 | 1 Gbps | IPMI + DDoS | $68.50 list · $54.80/mo annual |
| Bastion | 32c / 64t | 128 GB DDR4 ECC-R | 2 x 2 TB NVMe RAID 0/1 | 0/1 | 10 Gbps | IPMI + DDoS | $108.50 list · $86.80/mo annual |
| Fortress | 32c / 64t | 256 GB DDR5 ECC-R | 4 x 2 TB NVMe RAID 0/1/5/10 | 0/1/5/10 | 10 Gbps | IPMI + DDoS | $173.50 list · $138.80/mo annual |
| Citadel | 96c / 192t dual | 512 GB DDR5 ECC-R | 4 x 2 TB NVMe Gen4 RAID 0/1/5/10 | 0/1/5/10 | 25 Gbps | L3/L7 scrubber | $298.50 list · $238.80/mo annual | Full ledger: [all dedicated SKUs](https://fyrhost.com/dedicated/).
FYRHost: VPS from $3/mo and dedicated from $30.80/mo annual. The map spans forty locations — Reykjavik (Iceland) and Zurich (Switzerland) among them, plus Singapore, Dubai, Cape Town and fourteen others. Signup is email plus a 12-character password, or token-only. No KYC. Checkout is crypto-only across BTC, XMR, ETH, BNB, SOL, LTC, TRX, XRP, DOGE, DASH, ADA, GRAM (TON), and USDT (TRC-20/ERC-20/BEP-20/SPL). No payload or guest traffic logs; connection metadata ≤24h. DDoS included. One SKN company, one crypto checkout. Deploy at https://fyrhost.com/deploy/.
## The short version
Most “we outgrew our VPS” stories turn out to be neighbour noise, and the honest fix is a resize or a quiet neighbor, not a new hardware class. Stay virtual while the workload fits a slice of a shared EPYC or Xeon host: web apps, WireGuard endpoints, CI runners, game instances, pruned chain nodes. Move to metal when one of four statements turns true — a whole socket is the workload, storage queues are the product, the kernel itself gets replaced, or out-of-band console access matters more than provisioning speed.
The decision is hardware math, not payment plumbing: both classes here bill through the same nineteen-coin invoice behind the same no-KYC signup, so nothing about the choice involves a different checkout. The catalog keeps the boundary narrow on purpose. Sentinel, the largest VPS, carries 12 vCPU, 32 GB DDR5 ECC and 640 GB NVMe on a 10 Gbps unmetered port for $33.50 list ($26.80 annual-eff). Shield, the smallest dedicated, carries an 8c/16t Xeon E-2388G, 32 GB ECC and 2×500 GB NVMe behind IPMI for $38.50 list ($30.80 annual-eff).
Where the reasoning breaks is motive. Operators order metal for prestige, for a billing quirk, or on the assumption that bare silicon is faster per core in some absolute sense. Then they discover that a busy VPS slice on EPYC 9454P silicon embarrasses an idle E-2388G core for core. Exclusivity pays when your load would actually use the whole socket; it refunds nothing when the socket idles at 4%.
The rest of this page turns the four statements into measurements you can run in an evening, then walks the exact price lines and the virtualization features that keep workloads virtual. It finishes with the switch procedure itself, including the rollback window that makes it reversible.
Keep one distinction in your pocket as you read: capacity is rentable in slices, exclusivity is not rentable at all. Everything else on this page — price gaps, port speeds, ECC classes — is arithmetic you can run from the price table in an evening. Only your own workload can answer the exclusivity question, and the four measurements exist to turn that answer from a feeling into numbers: `steal`, memory, `queue` depth, and whether `IPMI` would have saved your last incident. You can afford to measure first — the metal is not going anywhere, and neither is the $4.00 gap.
VPS from $3* (fund $50+), storage 1–5 TB NVMe from $24/mo annual-eff, dedicated with IPMI from $30.80/mo. Crypto invoice, no KYC.
[Launch now](https://fyrhost.com/deploy/?kind=dedicated&plan=shield&period=12/)
## Four numbers that decide it
CPU steal comes first: the share of your clock the hypervisor hands to someone else. Watch the `st` column in `vmstat 1` or the `%st` row in `top` across a week of real traffic. Persistently above a few percent means a neighbour is eating your clock, and no upgrade below metal fixes it, because sharing is structural. A spike during your own batch job is normal; a floor of 5% at 3 a.m. is the number that matters.
Resident memory is second. `free -m` settles it in one line: when the available column sits against plan RAM and swap-in rates climb, the NVMe is now pretending to be memory and every page fault is a disk round-trip. The virtual ladder gives you room first: 4 GB on Sentry, 16 GB DDR5 ECC on Warden, 32 GB on Sentinel. Resizing is live, so memory pressure alone is a resize event, not a class change.
Sustained write bandwidth is third, and databases feel it before dashboards do. `fio --name=w --rw=write --bs=1m --direct=1 --size=8G` prints what a shared queue really serves; when the figure collapses under your production write pattern, the queue has become the product. On metal the queue is yours alone — RAID 0/1 on Shield's 2×500 GB, up to RAID 0/1/5/10 across Fortress's four 2 TB drives — which is the difference between a queue you tune and a queue you share.
IPMI and ECC class come fourth, and they are binary: no VPS plan offers out-of-band management, and every dedicated tier ships it with registered ECC DDR4 or DDR5. The classic mistake here is buying metal to fix an application bug — N+1 queries migrate with you byte for byte. Measure the four numbers first; the class choice is the conclusion, never the premise, and the next section prices the conclusion.
Run the four measurements on a schedule, not in a crisis: a week of `vmstat` history on a quiet month outvalues an evening of graphs during an incident. Steal and queue numbers drift with neighbours — a tenant who launches a batch job bends your numbers for days — so treat single bad days as weather and persistent trends as climate. Your resize decision deserves the climate version, and since the panel resizes live, collecting the extra week of evidence costs you almost nothing but attention.
## Where the price lines cross
Run both sides at annual efficiency and the gap is $4.00 a month: Sentinel at $26.80 against Shield at $30.80. That is the entire premium for an 8c/16t socket instead of 12 shared vCPU, 2×500 GB of NVMe on your own RAID instead of a slice, 1 Gbps unmetered instead of a shared fabric, and IPMI at the back of the chassis. Very few computing decisions offer a cleaner exchange rate, which is why the crossover gets argued about so much — it is genuinely close.
Above Shield the ladder stops being close and starts being categorical. Vault adds 16c/32t and 64 GB at $54.80 annual-eff ($68.50 list); Bastion reaches 32c/64t and 128 GB at $86.80. Fortress carries 256 GB DDR5 across four drives at $138.80, and Citadel pairs dual EPYC 9474F — 96c/192t, 512 GB, 25 Gbps — with an L3/L7 scrubber at $238.80. When a workload needs a number from that column, no VPS config imitates it, and pretending otherwise just moves the same ceiling around.
Order Shield at $38.50 list ($30.80 annual-eff) when the gap measures smaller than an hour of your time, and take the [annual Shield term](https://fyrhost.com/deploy/?kind=dedicated&%3Bplan=shield&%3Bperiod=12&plan=shield&period=12/) if the workload is staying — the −20% is the same discount the virtual ladder uses. The full ledger of both classes sits on the [dedicated lineup page](https://fyrhost.com/dedicated/), and the [side-by-side comparison](https://fyrhost.com/compare/vps-vs-dedicated-bare-metal/) keeps the specs in one table for the argument you are having with yourself.
Location multiplies cleanly, which keeps the comparison honest at any flag: a `+20%` Reykjavik premium or a `+35%` Swiss one scales both classes the same way, so the $4.00 gap survives geography unchanged. Term works identically — the `-20%` annual discount applies to VPS and metal alike, which is why your comparison should always use same-term figures. Your decision table needs exactly two numbers per class per flag: annual-eff against annual-eff. Everything else on the invoice is the same nineteen coins, the same no-KYC account, the same single-use address.
The mistake to avoid is comparing list against annual across classes, which manufactures a 20% gap that does not exist. Compute both sides on the same term, at the same location premium — Iceland and Norway add 20%, Switzerland 35%, and thirty-four sites stay at base — and the crossover lands where this section says it does.
## What a VPS still does better
Features are not the reason to switch, because KVM virtualization already hands you the interesting ones. Nested virtualization, Docker, custom ISOs and raw sockets ship on every plan — the stack documented at the [KVM project](https://www.linux-kvm.org/page/Main_Page) is the same one under your slice. Tor relays, mail stacks, VPN endpoints and phone-size side projects never need metal, and operators who migrate them “because dedicated sounds serious” pay a latency of days for the privilege.
Operations favor the virtual side even harder. A rebuild is a panel click, a fresh box images about 60 seconds after an invoice credits, and CPU and RAM resize live — so an experiment that dies costs a minute, not a racking slot. Bare metal rebuilds through IPMI work, but they measure in tens of minutes, and the discipline of reproducible provisioning that a 60-second image cycle teaches is worth more than any single hardware spec.
Availability math favors slices too: six Sentry boxes at $6.00 annual-eff each spread a service across six hosts, six power domains and six uplinks, so one failure is a sixth of a capacity loss instead of a total one. The mistake is verticalizing by habit — one big box per service — when the workload's real requirement is redundancy rather than exclusivity.
Run the horizontal test before the vertical one: can your service live on two small boxes instead of one medium one? A pair of entry slices at `$6.00` annual-eff each buys rolling restarts, a blast radius of half, and a migration path that never goes dark — none of which any single chassis provides at any price. Your users experience uptime as a property of the design, not of the silicon, and a `KVM` rebuild that finishes in `60 seconds` is the cheapest design headroom on the menu.
So the honest statement of the trade is this: the VPS class wins on rebuild speed, resize elasticity, failure isolation across many boxes, and every hypervisor feature short of the bare silicon itself. Metal wins on the four numbers from earlier — and only on them. The next section lists what those exclusivities actually buy.
## What only metal gives you
IPMI is the headline exclusivity: an out-of-band console on the chassis that answers when the OS does not. `ipmitool sol activate` from your desk reaches a boot menu, a kernel panic or a filesystem check on a box 3,000 km away, and a wedged server reimages without a rescue guest to bootstrap from. Every dedicated tier here ships it — Shield through Citadel — and nothing in the virtual class approximates it, because the hypervisor is precisely the layer that has failed in this scenario.
The kernel and the stack below it come second. Metal lets you run your own build from [kernel.org](https://kernel.org/), replace the bootloader, install a hypervisor of your choosing, or put a BSD directly on the iron — none of which survive a shared host's feature policy. Add registered ECC across larger banks, a RAID policy you control from 0/1 to 0/1/5/10, and unmetered ports scaling 1 to 25 Gbps, and the class definition writes itself: you are buying the layers a hypervisor normally keeps.
Where it breaks is time. Those powers arrive after 2–4 hours of racking and hand-over, and they arrive with obligations — the storage controller, the firmware, the rescue procedure and the kernel are now yours, on a machine that cannot be cloned away from trouble in 60 seconds. Ordering metal for a project that gets rebuilt hourly is paying a chassis price for a workflow the class deliberately slows down.
Try the out-of-band drill on your own schedule, not the incident's: once the metal arrives, open the console, watch a reboot from the `IPMI` view, and type into the bootloader once on purpose. Fifteen minutes of practice converts the console from a last resort into a tool you reach for early, and your `ipmitool` muscle memory is worth more than its spec sheet. Your panic level during the first real hang is a direct function of whether that console is familiar — and familiarity costs exactly one quiet lunch break.
The mental model that survives contact with reality: the VPS class rents outcomes, the metal class rents components. The next two sections follow the two workloads most often cited when the component version is genuinely required — storage pressure and connection-heavy relays.
## The storage-pressure case
Chain nodes are the cleanest example because the numbers are public. A Bitcoin full node's chainstate grows every block, and the pruned mode documented by [Bitcoin Core](https://bitcoincore.org/en/download/) keeps recent blocks while discarding old ones — `prune=550` in `bitcoin.conf` caps the block files near 550 MB, and the block-chain mechanics behind that figure are laid out in the [developer documentation](https://developer.bitcoin.org/devguide/block_chain.html). A pruned node fits the smallest VPS disk; an archival node, indexing everything forever, does not fit anything shared.
Diagnose the pressure the same way regardless of workload: `iostat -x 1`, and read the `await` column while production writes run. When await climbs into double-digit milliseconds and `%util` pins near 100%, the queue has become the bottleneck, and adding CPU to a queue problem is the expensive version of ignoring it. On shared NVMe the queue is segmented between tenants; on dedicated RAID 0/1/5/10 it answers to your write pattern alone, which is the whole argument when storage *is* the product.
Bulk that is not latency-sensitive belongs off the compute tiers entirely: the [storage line](https://fyrhost.com/storage/) prices 1–5 TB NVMe buckets from $24/mo annual-eff, so an archive that would bloat a boot disk lives one protocol hop away instead. The mistake to avoid is the reverse trade — keeping an archival node on a 40 GB Scout and discovering the eviction storms at 3 a.m. rather than at sizing time.
Watch your own growth curve instead of the size chart: chainstate grows every block regardless of your plans, and a disk that fits this quarter can fail next quarter on arithmetic alone. `df -h` in a weekly cron with an alert at 70% turns that surprise into a ticket you send yourself. Prune settings, storage offload, and the class decision all get cheaper when the warning arrives months early — your future self files this under boring wins, which is where the best infrastructure stories live.
So the storage rule reads: pruned and cache-shaped workloads stay virtual and resize upward cheaply; archival and write-shaped workloads either move to the storage line or move the whole queue onto metal where the queue is exclusive. Connection-heavy relays make the same argument with a different resource, and that is next.
## The relay-and-connections case
Peer-to-peer workloads fail differently: not on bytes but on sockets. A healthy node keeps dozens of persistent peer connections alive, as the [P2P network documentation](https://developer.bitcoin.org/devguide/p2p_network.html) describes, and proxies, message queues and game masters multiply that by tenants. The knobs are operating-system ones — `ulimit -n`, `net.ipv4.ip_local_port_range`, `net.netfilter.nf_conntrack_max` — and they turn identically on both classes, which is why “needs many connections” alone is not a metal argument.
It becomes a metal argument when the connections are the revenue: a relay or tracker whose port count and pps rate are the product. Such a load runs against a 10 Gbps Sentinel at $26.80 annual-eff until saturation measures real, then moves onto Citadel's 25 Gbps uplink, where the port is exclusive and the scrubber rides along. Absorption and connection capacity scale together at the top of the ladder in a way no slice imitates, because the edge devices in front of a shared fabric must stay conservative for every tenant at once.
Where the case breaks is latency-shaped workloads: if the requirement is proximity to peers or users rather than exclusivity of fabric, region choice beats class choice every time. Forty locations price at one list — Iceland, Norway, Switzerland, Luxembourg, Cyprus and Panama carry published premiums and the other thirty-four do not — so a same-class move across the map costs a redeploy, not an upgrade. Check the map before the chassis; it is the cheaper experiment by weeks.
Count your connections the way the kernel does: `ss -s` gives the totals, `conntrack -C` gives the tracked state, and both belong in the same weekly glance as disk and memory. A relay that doubles its peer count in a month is a queue problem in preparation, whatever the port speed says today. Your graphs, not the brochure's, pick the rung — and the right rung is the one with headroom at your 2 a.m. peak, not your office-hours average.
With the two hard cases covered, what remains is procedure: how the switch actually executes, what the clocks on each side do, and how the rollback window keeps the decision cheap to reverse — which is the next section's whole subject.
## The switch itself: provisioning and rollback
The checkout is identical for both classes, which removes it from the decision entirely: pick plan and term, sign up with an email or an anonymous token plus a 12-character password, then pay a live-rate invoice in any of nineteen coins to a single-use address. The [payment-flow guide](https://fyrhost.com/guides/crypto-payment-flow/) documents the flow step by step. Nothing about choosing metal changes the payment rails, the identity posture, or the invoice math; the catalog, not the cashier, is where the classes differ.
The clocks do differ, and planning around them prevents most cutover pain. A VPS images roughly 60 seconds after the invoice credits; dedicated metal takes 2–4 hours for racking, imaging and IPMI hand-over, with the out-of-band credentials arriving last. Schedule the migration window for the morning you can watch the metal clock run, rsync the data across first — `rsync -aHAX --info=progress2 src/ dst/` — and let the final delta sync be small enough to survive the wait.
Cut over with a short `TTL` on the A record, watch both logs during the overlap, and keep the old VPS running for a full billing cycle as the rollback path. The reverse migration is the same rsync pointed backward, and a decision you can reverse is one you can make quickly. The mistake is decommissioning the old box on day one to “save the $6”: the first kernel panic on new hardware is exactly the moment that $6 rents you an escape hatch.
Rehearse the rollback before you need it: point a test hostname at the old box, confirm the backward `rsync` path runs, and time how long a full reverse sync takes on your data volume. The number is usually minutes, and knowing it converts the cutover from a leap into a step. Your database, if one exists, deserves its own dump-restore rehearsal — application files rsync, database state wants a proper dump and a restore window you have actually timed at least once.
Take Vault at $54.80 annual-eff ($68.50 list) when 32 GB of ECC stops being enough rather than when impatience does — the ladder exists so each step is small. With procedure covered, the closing section compresses the whole page into the order you actually run it on a random Tuesday.
## The switch decision, in the order to run it
Measure before anything else: a week of `vmstat 1` for steal, `free -m` against plan RAM, and `fio` plus `iostat -x 1` for the write pattern, sampled during real traffic rather than synthetic calm. Map the results to the virtual ladder first — entry rung through the top of the line — because a resize inside the class costs a rebootless minute while a class change costs a racking afternoon.
Then check the three metal-only triggers against the workload honestly: does the job want a whole socket, is the storage queue itself the product, does the kernel or the stack below the OS get replaced, and would IPMI have saved you hours in the last incident. Two yeses justify metal; one yes justifies waiting for a second; zero yeses close the question for another quarter. Run the same term comparison when you price it — $26.80 against $30.80 is the real gap, not $33.50 against $38.50.
Execute the switch in the order the constraints allow: order the metal in the morning, rsync the bulk across while it racks, pay from the same nineteen-coin checkout you already use, and cut DNS over with a short `TTL` once the IPMI hand-over mail lands. Verify the out-of-band console before you need it — `ipmitool sol activate` tested on a healthy box is a five-minute drill that pays for itself at 3 a.m. — and only then tune the RAID policy and the kernel you switched classes to reach.
Keep the old VPS billed through one full cycle as the rollback path, and decommission it only after the new box has survived its first patch cycle. File the four measurements from step one next to the new box's numbers, so the next decision starts from data. That is the entire procedure: measure, map, trigger-check, price on one term, migrate with overlap, verify out-of-band, retire late. The same invoice has covered both boxes the whole time, which is exactly why the decision should be argued in hardware terms and nothing else.
Write the outcome down when you are done: which measurement triggered the move, which tier answered it, and what the old box's numbers looked like on the way out. Six months later that paragraph is the cheapest consulting you will ever read — it tells you whether the exclusivity paid for itself, and it trains the next decision to start from evidence instead of itch. Your infrastructure memo archive, one paragraph per migration, outlives every dashboard you have ever pinned, and it is the only report your future self will actually trust.
Can I start on a VPS and move to dedicated later? Yes, and the move is a redeploy rather than an upgrade: rsync the data, cut the records over with a short TTL, and keep the old box billed for a cycle as the rollback path. No in-place path to exclusive metal exists, so plan the cutover window around the 2–4 hour racking clock.
Which class provisions faster? The VPS class, by hours: a box images about 60 seconds after the invoice credits. Dedicated metal needs 2–4 hours for racking and IPMI hand-over before your credentials arrive, with the out-of-band console landing last in the sequence.
Can I pay for a dedicated server with crypto? Yes — both classes draw on the same nineteen-coin checkout, XMR and BTC included, at live rates to single-use addresses. The invoice math is tier-independent: Shield at $38.50 list and Citadel at $298.50 list check out identically to a $5.49 Scout.
Do I lose DDoS filtering when I move to metal? No. L3/L4 filtering runs at the edge for dedicated tiers exactly as it does for VPS, and Citadel adds the L3/L7 scrubber on its 25 Gbps uplink. What metal adds on top is IPMI, so a wedged box reimages out-of-band instead of waiting for a rescue environment.
Is a dedicated server always faster per core? No. A busy slice on EPYC 9454P silicon outpaces an idle E-2388G core for core, and an app with N+1 queries runs the same bug on both classes. Exclusivity pays when the workload would use the whole socket, which is what the steal and utilization measurements exist to establish.
What is the one measurement that predicts the need for metal? CPU steal over a full week of production traffic — the st column in vmstat 1. A persistent floor above a few percent is structural to sharing and no VPS upgrade removes it; memory and queue pressure resize away, steal does not.
## Primary sources
- [KVM project documentation](https://www.linux-kvm.org/page/Main_Page)
- [Bitcoin Core: download and pruned node docs](https://bitcoincore.org/en/download/)
- [Bitcoin developer guide: the block chain](https://developer.bitcoin.org/devguide/block_chain.html)
## Related guides
- [Dedicated servers — full ledger](https://fyrhost.com/dedicated/)
- [Best no-KYC dedicated servers](https://fyrhost.com/best-no-kyc-dedicated-servers/)
- [Cheapest no-KYC dedicated](https://fyrhost.com/cheapest-no-kyc-dedicated/)
- [Monero bare metal server](https://fyrhost.com/bare-metal-monero/)
- [Buy dedicated with crypto](https://fyrhost.com/buy-dedicated-server-with-crypto/)
- [Offshore dedicated server](https://fyrhost.com/offshore-dedicated-server/)
- [No-KYC dedicated worldwide](https://fyrhost.com/no-kyc-dedicated-worldwide/)
- [VPS vs dedicated: when to switch](https://fyrhost.com/compare/vps-vs-dedicated-bare-metal/)
## More ways to pay
[Pay with Bitcoin](https://fyrhost.com/bitcoin-vps/) · [Pay with Monero](https://fyrhost.com/monero-vps/) · [Pay with USDT (TRC-20/ERC-20/SPL)](https://fyrhost.com/usdt-vps/) · [Pay with XRP](https://fyrhost.com/xrp-vps/) · [Pay with TRX](https://fyrhost.com/tron-vps/) · [Pay with Dogecoin](https://fyrhost.com/dogecoin-vps/) · [Pick the jurisdiction first](https://fyrhost.com/jurisdiction/).
## Checkout facts
- **Price** VPS from $3/mo. Dedicated from $30.80/mo.
- **Identity** No KYC. Token-only (no email) or email plus a 12-character password. No passport, phone, or card.
- **Payment** Nineteen ways to settle, none of them a card: XMR and BTC lead, ETH, SOL, LTC, BNB and TRX follow, then XRP, DOGE, DASH and ADA, GRAM (TON) for Telegram-adjacent rails, and USDT on seven networks — TRC-20, ERC-20, BEP-20, SPL, Polygon, Arbitrum, Optimism. Live-rate invoice.
- **Logs** No payload / no guest traffic logs. Billing and panel actions retained. Connection metadata ≤24h.
- **Platform** KVM on ECC and local NVMe. DDoS filtering included. Facilities in 40 cities across four continents — the locations table lists each one's tier, uplink and governing law. Need Iceland or Switzerland plus no KYC? Those two carry a published premium (+20% / +35%); every other flag stays at base list. Pay BTC, USDT, XMR, BNB, XRP, GRAM (TON) or 10 more. Open deploy to pick a plan.
## Ready to launch?
Build the box — VPS, storage or bare metal — create the password, pay the invoice that follows.
[Launch now](https://fyrhost.com/deploy/?kind=dedicated&plan=shield&period=12/)[Storage](https://fyrhost.com/storage/)