Analysis of the 4 servers.md Analysis of the 4 servers (all reached successfully via the provided SSH commands/aliases). All four servers use Caddy as the web server/reverse proxy (no nginx). Docker is present on all, but usage varies significantly. 1. Contabo (149.102.150.185) — key: contabo2_new1 Specs: Ubuntu 22.04, 8 vCPU (AMD EPYC), 24 GiB RAM, 388 GB disk. Resources (current): • Disk: 56 GB used / 388 GB (15%) • RAM: 2.2 GiB used / 23 GiB total (17 GiB free, ~20 GiB available). Swap 4 GiB (unused). • CPU: 8 cores, load avg ~1.3–1.8, ~13% user + ~4% system (mostly idle). Proxy: Caddy (80/443 + admin on 2019). Domains (from Caddyfile): • auto.master1.vip, zeroclaw.addict.best → reverse proxy to localhost:18789 (openclaw/zeroclaw gateway) • banks.master1.vip → static files (/root/banks-master1) Containers (11 running): • outline + outline-db + outline-redis (Outline wiki) • umami-app + umami-db (analytics) • root-activepieces-1 (Activepieces) • listmonk-app + listmonk-db (newsletter) • root-postgres-1, root-redis-1 • agent-browserless (browserless/chrome) • Several internal-only services on 127.0.0.1 Apps / Services: • PM2 (4 apps): dentist-ide, master1-telegram, opencode-proxy, zeroclaw-gateway • Pocketbase instances (main + dentist) • openclaw (heavy Node gateway on 18789) • Systemd: biology-exams.service, master1-share.service, LXD, fail2ban, caddy, Docker • Multiple Node backends + custom tools Notes: Heaviest/most feature-rich server. Mix of self-hosted SaaS tools (Outline, Umami, Activepieces, Listmonk) + custom PM2 + Claw-related gateways. Plenty of RAM/CPU headroom. ─── 2. Hetzner Dokploy (46.62.228.173) — key: hetzner_dokploy Specs: Debian 12, 8 vCPU (AMD EPYC-Genoa), 16 GiB RAM, 301 GB disk. Resources (current): • Disk: 46 GB used / 301 GB (16%) • RAM: 5.0 GiB used / 15 GiB total (only ~704 MiB free, ~10 GiB available). Swap 4 GiB (1.8 GiB used). • CPU: 8 cores, load avg ~0.1–0.4 (very idle). Proxy: Caddy (sophisticated config with heavy PocketBase routing). Domains (from Caddyfile): • ielts.fast (main) — complex routing: • PocketBase admin/API (/pb/*, /_/* etc. → 127.0.0.1:8090) • TTS/phonetics API (/api/engines, /api/speak etc. → port 5000) • Telegram notify → port 3005 • Multiple promedic1.com sub-apps (coach, diet, ielts) via PocketBase systemd services Containers (only 3 running): • webapp_ielts (ielts-fast) • the-ielts • uptime-kuma Apps / Services: • Pocketbase (systemd): coach-pocketbase (coach.promedic1.com), diet-pocketbase (diet.promedic1.com), ielts-pocketbase • ielts-analysis.service, ielts-tts.service (phonetics) • Long-running agy process • Node server • Docker projects present: /opt/chatwoot, /opt/OpenHands, /root/ielts-fast, /root/female-promedic1 • Caddy, Docker, fail2ban, autoheal, qemu-guest-agent Notes: Key named "hetzner_dokploy" but no Dokploy panel detected (/etc/dokploy, etc. not present). It is a custom self-managed setup focused on IELTS + promedic1.com apps with heavy PocketBase usage. Memory is the tightest of the four (swap in use). ─── 3. AI Developer / Kuwait (213.199.36.17) — key: ai_developer_key Specs: Ubuntu 22.04, 6 vCPU (AMD EPYC), 12 GiB RAM, 97 GB disk. Resources (current): • Disk: 24 GB used / 97 GB (25%) • RAM: 1.2 GiB used / 11–12 GiB total (~554 MiB free, ~10 GiB available). No swap. • CPU: 6 cores, load avg ~0.1–0.6 (extremely idle, ~99% idle in snapshot). Proxy: Caddy. Domains (from Caddyfile): • prodemy.promedic1.com — SPLE Pro (Saudi Pharmacist Licensure Exam preparation app). Static files from /root/sple-exam-app with strong security headers + caching. Containers (1 running): • smm_automation (very low resource usage) Apps / Services: • MySQL (mysqld — notable memory user) • Redis • openclaw (Node gateway on 18789) • Caddy, Docker, fail2ban, php8.1-fpm, containerd • One docker-compose project: smm_automation Notes: Lightest load of the larger servers. Focused on the prodemy.promedic1.com SPLE app + supporting services (MySQL/Redis + openclaw). Very healthy resources. ─── 4. Hostinger (31.97.122.87) — alias hostinger (uses ~/.ssh/id_ed25519) Specs: Ubuntu 24.04, 2 vCPU (AMD EPYC), 8 GiB RAM, ~96 GB disk. Resources (current): • Disk: 24 GB used / 96 GB (25%) • RAM: 658–693 MiB used / 7.8 GiB total (~900+ MiB free, ~7.1 GiB available). No swap. • CPU: 2 cores, load avg ~0.1–0.3 (very idle). Proxy: Caddy. Domains (from Caddyfile): • diaa.master1.vip → static /var/www/taibaat • taibaat.master1.vip → static /var/www/diaa-taibaat • fast.promedic1.com → static SPA /var/www/fast-diet/current (with try_files for client-side routing) Containers: 0 running (Docker installed but no active containers in docker ps). Apps / Services: • Caddy + Docker/containerd • Compose files present but inactive: • /docker/n8n/docker-compose.yml (n8n) • /root/picoclaw (Go-based "Claw" project with Dockerfile + compose) • Standard Ubuntu services + qemu-guest-agent Notes: Smallest/cheapest server. Purely static sites + proxy right now. Docker/n8n/picoclaw projects are prepared but not running. Lowest resources (only 2 vCPU). ─── Summary Comparison ┌───────────┬──────┬────────┬────────┬──────────┬────────────┬───────┬──────────┬─────────────────────────────────┐ │ Server │ vCPU │ RAM │ Disk │ Used RAM │ Containers │ Proxy │ Load │ Main Focus │ ├───────────┼──────┼────────┼────────┼──────────┼────────────┼───────┼──────────┼─────────────────────────────────┤ │ Contabo │ 8 │ 24 GiB │ 388 GB │ 2.2 GiB │ 11 │ Caddy │ Low-Med │ Self-hosted tools + custom apps │ ├───────────┼──────┼────────┼────────┼──────────┼────────────┼───────┼──────────┼─────────────────────────────────┤ │ Hetzner │ 8 │ 16 GiB │ 301 GB │ 5.0 GiB │ 3 │ Caddy │ Very Low │ IELTS + promedic1.com + PB │ ├───────────┼──────┼────────┼────────┼──────────┼────────────┼───────┼──────────┼─────────────────────────────────┤ │ AI Dev │ 6 │ 12 GiB │ 97 GB │ 1.2 GiB │ 1 │ Caddy │ Very Low │ SPLE (prodemy.promedic1.com) │ ├───────────┼──────┼────────┼────────┼──────────┼────────────┼───────┼──────────┼─────────────────────────────────┤ │ Hostinger │ 2 │ 8 GiB │ 96 GB │ ~0.7 GiB │ 0 │ Caddy │ Very Low │ Static sites + prepared n8n │ └───────────┴──────┴────────┴────────┴──────────┴────────────┴───────┴──────────┴─────────────────────────────────┘ All servers are in good health resource-wise except Hetzner (higher RAM usage + swap). Let me know if you want deeper dives (e.g. full Caddyfile, specific app logs, docker-compose contents, or recommendations for optimization). ======================================================================================= Prioritized ranking for hosting SILMA (Voice Clone + Voice Generation) Best → Worst: ┌──────┬──────────────┬─────────────────┬───────┬──────────────────────────┐ │ Rank │ Server │ IP │ Score │ Recommendation │ ├──────┼──────────────┼─────────────────┼───────┼──────────────────────────┤ │ 1 │ AI Developer │ 213.199.36.17 │ Best │ Strongly Recommended │ ├──────┼──────────────┼─────────────────┼───────┼──────────────────────────┤ │ 2 │ Contabo │ 149.102.150.185 │ Good │ Acceptable (with limits) │ ├──────┼──────────────┼─────────────────┼───────┼──────────────────────────┤ │ 3 │ Hetzner │ 46.62.228.173 │ Fair │ Risky │ ├──────┼──────────────┼─────────────────┼───────┼──────────────────────────┤ │ 4 │ Hostinger │ 31.97.122.87 │ Poor │ Not recommended │ └──────┴──────────────┴─────────────────┴───────┴──────────────────────────┘ Detailed Reasoning 1. AI Developer (213.199.36.17) — Best Choice • Very clean server (only 1 light container running: smm_automation) • Extremely low resource usage: ~1.2 GB RAM used out of 12 GB (≈10 GB available) • 6 vCPU with almost zero load • No swap • Plenty of headroom to load voice models without fighting other processes • Good balance for CPU-based voice inference Best for SILMA because voice cloning/generation is sensitive to consistent CPU + RAM availability. This server gives the most predictable performance right now. 2. Contabo (149.102.150.185) — Second Best • Highest specs: 8 vCPU + 24 GB RAM + 388 GB disk (best raw power + storage) • Currently only using ~2.2 GB RAM (lots of headroom on paper) • 11 running containers + PM2 processes + openclaw + multiple databases Downside: Noisy server. Many services are already running (Outline, Umami, Activepieces, Listmonk, Pocketbase, LXD, etc.). Voice generation can be bursty and needs stable resources — contention is a real risk here. Use only if you set strict Docker resource limits (CPU + memory) on the SILMA container. 3. Hetzner (46.62.228.173) — Third • 8 vCPU is good • 16 GB RAM but already using 5 GB + 1.8 GB swap → memory pressure exists • Only 3 containers (relatively clean) • Low CPU load Problem: Swap is already active. Voice models (especially cloning) are memory hungry. When they load big checkpoints, this server can start swapping heavily, causing very slow generation or OOMs. 4. Hostinger (31.97.122.87) — Worst • Only 2 vCPU — this is the dealbreaker for voice AI. • 8 GB RAM is marginal even for small models. • Cleanest server (0 containers), but fundamentally underpowered. Voice cloning (especially any modern model like CosyVoice, Fish Speech, RVC, XTTS, etc.) will be painfully slow or may not even load properly on 2 cores. Important Notes for SILMA (Voice Clone + Generation) • All 4 servers are CPU-only (no NVIDIA GPU detected on any of them). • Voice quality and speed will be significantly worse than on a GPU server. • Some models will be very slow (especially zero-shot cloning or long-form generation). • RAM is usually more important than CPU for voice models. Prioritize servers with the most available RAM. • Recommended model types for these servers (CPU): • Lighter/faster: Piper, Coqui TTS (some versions), quantized models, smaller RVC models • Heavier: Will struggle (especially on Hostinger) My Recommendation ┌────────────────────────────────────────────┬──────────────────────────────┬────────────────────────────────────────────┐ │ Goal │ Best Server │ Action │ ├────────────────────────────────────────────┼──────────────────────────────┼────────────────────────────────────────────┤ │ Best performance + isolation │ AI Developer (213.199.36.17) │ Use this first │ ├────────────────────────────────────────────┼──────────────────────────────┼────────────────────────────────────────────┤ │ Maximum power + willing to limit resources │ Contabo │ Use with Docker --cpus and --memory limits │ ├────────────────────────────────────────────┼──────────────────────────────┼────────────────────────────────────────────┤ │ Only if the above two are unavailable │ Hetzner │ Risky — monitor swap closely │ ├────────────────────────────────────────────┼──────────────────────────────┼────────────────────────────────────────────┤ │ Avoid │ Hostinger │ Too weak for voice work │ └────────────────────────────────────────────┴──────────────────────────────┴────────────────────────────────────────────┘ Would you like me to: 1. Give you a ready-to-use Docker Compose example with resource limits for the best server? 2. Check what voice models would realistically run well on the top 2 servers? 3. Suggest commands to reserve resources cleanly on Contabo? ==================================================================================== ======================================================================