The decision: buy composable control, not a universal “cheap VPS”
Vultr fits an operator who wants to choose the CPU family, region, storage mode, network topology and hourly lifecycle—and will operate the application and recovery path. Shared Cloud Compute covers modest services; dedicated and GPU are different products, not just larger lines in one VPS table. Vultr is not a blanket price, RAM, performance or support winner.
- Stopping does not stop the server bill. CPU, RAM, disk and IP remain reserved; ordinary hourly accrual continues. Destroy is what ends that server’s accrual, and it deletes its data.
- Backups are an input, not a hidden free safety net. The current Cloud Compute provisioning guide describes Automatic Backups as enabled by default, with an additional 20% of eligible base compute. Reconcile the live order summary and still make an independent, restore-tested copy.
New hosts start with setup. Existing data requires the billing/recovery gate: CONTROL WRITES / DEFINE CUTOFF → FINAL EXPORT → VERIFY FINAL-CUTOFF RESTORE → DESTROY → AUDIT. Evidence was checked 2026-11-05; catalog rows and marketing region lists do not prove checkout stock.
Choose a family by CPU and storage—not just RAM
Vultr “Regular Performance” (called Regular Cloud Compute in our tables) uses previous-generation Intel vCPU and SSD. High Performance AMD/NVMe or Intel variants are different; High Frequency advertises shared 3GHz+ Intel and local NVMe with another core/disk/transfer balance. Neither is a universal speed winner.
VX1 adds dedicated vCPU plus explicit High Performance Block Storage, local NVMe or mixed choices. This separates the storage decision from an ordinary instance-table disk label. High Performance Block Storage requires actual persistence/charging/region rehearsal; local NVMe use needs data recovered independent of the selected server’s lifecycle. Treat that local capacity as scratch unless independent application recovery proves otherwise.
| Family | Compute / storage | Billing and backup interpretation | Decision use |
|---|---|---|---|
| Regular Performance / “Regular” | Shared previous-gen Intel, SSD | Ordinary non-GPU documented cap; eligible Auto Backups | Cost-sensitive modest service where older-generation hardware is acceptable |
| High Performance | Shared AMD or Intel depending on plan/region, NVMe | Ordinary documented cap; eligible backups; exact variant important | Common comparison for n8n, WordPress and small CPU model experiments |
| High Frequency | Shared 3GHz+ Intel, local NVMe | Different disk/transfer profile, not a measured speed guarantee | Single-thread-sensitive apps after a resource comparison |
| Optimized Cloud Compute | Dedicated AMD; four workload variants | Dedicated product, not interchangeable with shared smallest plans | Sustained CPU or noisy-neighbor isolation requirement |
| VX1 | Dedicated; block, local NVMe or mixed storage | No current Auto Backups per VX1 guide; cap/storage snapshot/transfer checks required | Deliberate storage-flexible dedicated path with a non-Vultr-backed recovery plan |
| Cloud GPU / Bare Metal | GPU accelerator / dedicated physical hardware | GPU 730-hour documented cap and standard SLA exclusion; Bare Metal excluded from generic snapshot wording | Only after hardware, driver, persistence and cost acceptance |
Read the displayed price as a ledger input
These are displayed base prices, not an all-in invoice or universal rate card. Live product/region/backup/volume/IP/license/tax/egress determine cost. The pricing page and plans API inform—not replace—the order summary.
| Base plan example | Monthly display | RAM / storage / included transfer | Base arithmetic with eligible 20% backup |
|---|---|---|---|
| Regular 0.5 GB IPv6-only | $2.50 | 0.5 GB / 10 GB SSD / 0.5 TB | $3.00 |
| Regular 0.5 GB default IPv4 | $3.50 | 0.5 GB / 10 GB SSD / 0.5 TB | $4.20 |
| Regular 1 GB | $5 | 1 GB / 25 GB SSD / 1 TB | $6 |
| Regular 2 GB | $10 | 2 GB / 55 GB SSD / 2 TB | $12 |
| Regular 4 GB | $20 | 4 GB / 80 GB SSD / 3 TB | $24 |
| Regular 8 GB | $40 | 8 GB / 160 GB SSD / 4 TB | $48 |
| High Performance 2 GB | $12 | 2 GB / 50 GB NVMe / 3 TB | $14.40 |
| High Performance 4 GB | $24 | 4 GB / 100 GB NVMe / 5 TB | $28.80 |
| High Performance 8 GB | $48 | 8 GB / 180 GB NVMe / 6 TB | $57.60 |
| High Frequency 2 GB | $12 | 2 GB / 64 GB NVMe / 2 TB | $14.40 |
| High Frequency 4 GB | $24 | 4 GB / 128 GB NVMe / 3 TB | $28.80 |
| High Frequency 8 GB | $48 | 8 GB / 256 GB NVMe / 4 TB | $57.60 |
| Optimized General Purpose | $30 | 4 GB / 75 GB NVMe / 5 TB | $36 |
| Optimized CPU | $28 | 2 GB / 50 GB NVMe / 4 TB | $33.60 |
| Optimized Memory | $40 | 8 GB / 50 GB NVMe / 5 TB | $48 |
| Optimized Storage | $75 | 8 GB / 150 GB NVMe / 6 TB | $90 |
VX1 examples showed 2 vCPU/8 GB block storage at $0.060/hour and 2 vCPU/8 GB/120 GB local NVMe at $0.076/hour. API $43.80/$55.48 monthly fields equal simple 730-hour arithmetic—not proof of a cap or complete boot-volume total. Missing Auto Backups and unknown snapshot/transfer behavior need product-specific recovery.
Do not “correct” hourly prices with a calculator. For Regular 2 GB, the captured webpage displayed $0.015 while the API returned $0.014; neither establishes which field is rounded or controls the invoice. The $10 displayed monthly number is the stated base example; the selected hourly figure is separately attributed when needed. Rounding 672×$0.014 to an inferred invoice total would put an invented algorithm into a purchasing article.
The $10 Regular 2 GB + $2 backup example is not universal across regions. Reserved IPv4 and default/IPv6-only differ; panel/API/tax/credit terms need separate checkout checks.
Hourly does not mean stopped is free
Server billing documents ordinary Cloud Compute, Optimized and Bare Metal at a 672-hour monthly cap, Cloud GPU at 730 hours, hourly rounding and an invoice on the first of the month. A stopped server keeps billing because resources remain reserved; a temporary test needs recovery-safe destruction, not just Stop.
There is a genuine VX1 ambiguity. Current pricing text says actual VX1 hourly use is not capped at 672 hours, while general non-GPU server billing text says a 672-hour cap. $43.80 equals 730×$0.060, but that arithmetic cannot adjudicate the discrepancy. Read the exact SKU’s live order display and later invoice/support clarification; do not let either article or tool promise you 672 or 730 for VX1.
A global list is not a stock or latency guarantee
The Cloud Compute marketing page captured 34 regions, including Chennai; the public regions API, rechecked during drafting, returned 33 and still omitted Chennai. Keep both facts attached to their sources. It is not a license to say “33 global locations but not Chennai” without the marketing context, or “34 regions available to every VPS” as a checkout promise.
Plan footprints differ: captured Regular 2 GB listed 31 locations, VX1 9, and a selected L40S GPU row none. That is neither reserved capacity nor uniform backup/storage support. Verify the selected plan/region and its price rather than transfer a base budget unchanged.
The captured São Paulo location costs for Regular 1/2/4/8 GB showed 1.5× base monthly prices ($7.50/$15/$30/$60 rather than $5/$10/$20/$40). Similar API rows appeared for some HP/HF and Optimized plans at this capture, but neither makes 1.5 a worldwide multiplier or a permanent contractual quote. Compute your actual selected region’s total before comparing providers.
Transfer and IP decisions change the bill and connectivity
For documented Cloud Compute behavior, the account receives 2 TB free outbound bandwidth each calendar month, without rollover, and each running instance accrues its plan allowance in hourly shares, using a 672-hour month. Outbound excess costs $0.01/GB and can occur during a month once current outbound use exceeds allocated share. This is not “a generous flat 2 TB included forever,” “unlimited 2 TB every rolling 30 days,” or proof about VX1 or excluded traffic.
A plan listed as 2 TB, converted here to 2048 GB, contributes 2048 × documented_running_hours / 672 GB, plus the separate account pool. At 168 running hours, that share is 512 GB, not an unconditional first-hour 2 TB. Units/control-panel use determine actual calculations. Backup/snapshot, VPC/Object Storage, inbound, stopped/deleted and VX1 behavior remain unknown; see the transfer worksheet.
Ordinary default IPv4, smallest-plan IPv6-only and $3/month Reserved IPv4 differ. The $1 smallest-plan difference is no universal attached-IP fee; extra attached IPv4 cost is unverified. IPv6-only is acceptable only if registries, DNS, clients and webhooks support it; an open model port is not a connectivity repair.
Vultr Firewall and host UFW/nftables are separate layers, and Docker-published containers need an actual effective-exposure check. A database listening on 5432, n8n on 5678 or Ollama on 11434 does not become safe merely because both firewall menus exist. The operator decides whether those should exist on loopback, inside Compose/VPC, or through a deliberately protected front door. Host validation turns this into a checklist rather than a slogan.
Marketplace images shorten provisioning, not ownership
In the reviewed results and direct candidate URLs, no dedicated Vultr-labs n8n or Ollama application listing was identified; those candidates returned 404. That is a scoped catalog finding, not proof of universal absence. Docker Ubuntu supplies a runtime, not a complete pinned n8n+Postgres+HTTPS stack. The n8n operations guide supplies that custom reference deliberately.
Third-party N99panel supports N8N, advertising a 14-day trial then $4.99/month plus compute. It is neither Vultr-managed n8n nor free with compute. Confirm panel support/security/backup/license renewal and cancellation independently.
The Vultr-labs WordPress image advertises Ubuntu 24.04, Nginx/MariaDB, Cockpit:9080, PHPMyAdmin, Certbot, Wordfence and SMTP Mailer; OpenLiteSpeed is different. Discover current PHP/core/paths/admin access instead of trusting “Latest.” Maintaining neither those layers nor an exportable site makes managed WordPress a better fit.
OpenClaw here really is an agent with Code Server, Docker/Homebrew, skills/subagents and account connections—not a reused hosting-panel label. The listing mentions DeepSeek v3.2 compaction; local gateway hosting does not imply entirely local processing. ComfyUI is a separate GPU-dependent image workflow. See AI apps before granting privilege or budgeting VRAM.
Scenario → starting envelope → stop or exit signal
These are conditional editorial hypotheses, not tested minimum, benchmark or high-confidence capacity plan. They help you select the first bounded experiment and the evidence to collect. Production sizing is provisional until your software versions, workflow payloads, peak concurrency, context, disk growth and target service levels are known.
| Reader job | Conditional starting hypothesis | Observe / stop signal | Next branch if that job does not fit |
|---|---|---|---|
| Deliberately tiny hosted-model n8n test | Consider Regular 1 GB for one low-payload, manual-only workflow if measured OS/n8n fits; consider 2 GB if database/proxy headroom needs it | Swap persists, OOM, timeouts or updates overlap memory unprepared | Stop workloads, reduce concurrency or data volume; consider HP/HF 4 GB rather than assuming 1 GB production-ready |
| Small n8n + private Postgres + HTTPS | Consider HP/HF 4 GB, or acceptable Regular 4 GB if cost and older hardware tradeoffs are understood; 8 GB only for measured added heads | Database/execution growth, API retry loops, RAM/disk pressure or current restore process interrupts service | Bound pruning and production concurrency first, split DB/model or increase resources based on the actual constraint |
| Test/small self-managed WordPress | Consider Regular 2 GB if plugin/theme/cache and backup footprints fit; consider HP/HF 4 GB when measured PHP/DB/cache demand warrants | Slow admin/page operations caused by PHP/DB not remote API; full disk or backup unable to complete | Inspect actual bottleneck; managed WordPress may be better than merely adding another vCPU |
| One small CPU model experiment | Consider 4 GB for one carefully chosen small model, one short-context request and adequate reserve | Resident weights/KV/runtime exhaust RAM, repeat swap, process killed, unbounded queue or model download consumes recovery space | Smaller licensed model/tag, lower context/concurrency, separate model host, external model or proper GPU; 8 GB is only a possible modest quantized 7B-class hypothesis, not universal 7B support |
| Sustained dedicated CPU / larger memory | Optimized or VX1 after product/storage/cost/backup checks | Exact VM/storage/region invoice unknown, CPU generation/driver or recovery support cannot be confirmed | A different confirmed family/provider or managed service; VX1 is not an inexpensive shortcut around application recovery |
| Multi-model, image generation or VRAM requirement | Separate Cloud GPU budget, driver/runtime and source-specific VRAM worksheet | No matching GPU stock, workflow requires external model weights, models exceed VRAM/system RAM or credit expiry negates test goal | External API or managed image service; ordinary shared CPU cannot be upgraded into CUDA by adding RAM |
| Lawful email/automation use | Explicit destination consent, rate limits and least-privilege accounts | Any spam, unauthorized scanning, open proxy, crypto mining, bulk-abuse or excessive-resource pattern | Do not buy for that use. AUP restrictions are not routing or configuration puzzles. |
Compare cash shape, recovery and control—not winner rankings
| Provider lens | What the other cluster establishes | What changes at Vultr | Decision test |
|---|---|---|---|
| Contabo | Larger RAM-per-euro tiers, many regions, fair-use/unlimited-focus traffic and provider-side recovery alongside fair-use concerns | Vultr requires product/region CPU, 672/730-hour, pooled outbound and backup/destroy decisions rather than maximum RAM/$ | If RAM-per-dollar dominates, price Contabo honestly; if finer hourly/product/network selection dominates, compare its actual Vultr total |
| RackNerd | Annual KVM specials, manual lifecycle care and one-click live-control app installation | Vultr’s $10 Regular 2 GB is $120/year base if used all 12 months before add-ons/tax, versus a dated 2 GB special example $35.99/year—not identical storage/traffic/cash/ops | Prefer upfront annual-special economics only if turnover and support boundaries fit; do not compare annual cash to monthly multiply and assume all else equal |
| InterServer | $3 per 2 GB linear monthly Linux KVM slice, three US regions and manual KVM recovery/support route | Vultr offers broader regions/products/API/VPC and pooled outbound, with more plan/backup/storage edge cases | A three-US-region simple monthly KVM decision can be easier; global/cloud-composable requirements favor a separately verified Vultr offer |
Who fits—and who should choose another route
Before order, answer these short questions:
- Can every selected charge be named? Region, compute, volume, eligible backup, IP, panel/network/API and transfer; if no, is there evidence from the exact live plan?
- Will your workflow tolerate shared CPU and this model’s memory/disk/context? If no, what measured threshold sends it to a dedicated or hosted/GPU path?
- Can a user reach IPv4/IPv6 and admin/model/private services only as intended? If no, resolve that on provisioning rather than after exposing a credential.
- Are daily operations owner-defined? Updates, pruning, expiration, rate limits, legal destination use and revoke/kill process for an agent have one named owner.
- Could you destroy it safely today? Full cutoff is controlled, exact current data is independently restored/tested, target ID/region is named, and retained disks/addresses/licenses/invoices are known.
If those answers are yes, proceed to secure a fresh Vultr host. If the primary problem is actually selecting a model or agent topology, read AI apps. If “stop billing” is the immediate need, do not touch the server until final-cutoff restore and cancel-last is satisfied.
Selected official references, checked 2026-11-05: pricing, Cloud Compute, VX1 product and guide, plans/regions APIs, billing/stopped, transfer, backup/snapshot, provisioning, reserved IPv4, Marketplace, TOS/SLA/AUP. Current labels, exact selected plan price/stock and unresolved product-specific terms always need live confirmation.