Skip to main content

​
Cernio — Production Simulation

Version: 1.0 Date: 2026-03-25 Source: Hetzner VPS Plan, Strategy Docs (02, 04), Handbook (Ch. 189-210) Purpose: Simulate infrastructure, AI costs, and operational load at 10, 50, 200, and 1,000 users

​
1. Infrastructure Baseline

​
Current Architecture (Post-Migration to Hetzner)

ComponentServerSpecsMonthly Cost
Galata (Production)Hetzner CX438 vCPU, 16GB RAM, 160GB NVMe€12.80 (~$13.90)
Kadikoy (Worker/AI)Hetzner CX334 vCPU, 8GB RAM, 80GB NVMe€7.40 (~$8.03)
SnapshotsBoth serversWeekly automated~€1.72 ($1.87)
CloudflareCDN + R2Free tier$0
Total infrastructure~$23.80/mo
Prices reflect Hetzner post-April 2026 increase (+30-37%).

​
What Runs Where

Galata (Production):
  • Supabase (PostgreSQL + GoTrue Auth + PostgREST + Kong Gateway)
  • dose.com.tr (corporate site)
  • admin.dose.com.tr (admin panel)
  • Coolify orchestrator + Traefik
  • WireGuard VPN, Uptime Kuma, Vaultwarden
Kadikoy (Worker/AI):
  • Cernio Next.js application (SSR)
  • Python scraper
  • 16 API routes (including AI-calling routes)
  • Batch scripts (finder, cleaner, bulk_fixer)

​
Scaling Decision: CPU vs Network Bound

WorkloadBound ByImpact
AI discovery API callsNetwork (waiting for LLM response)Low CPU, high latency
AI batch classificationNetwork (25 companies per call)Low CPU, moderate latency
Web scraping (Perplexity/Google)Network (search API response)Low CPU, high latency
Python PDF/Excel parsingCPU (text extraction)High CPU, brief duration (max 60s)
Next.js SSRCPU (React rendering)Moderate CPU, scales with concurrent users
PostgreSQL queriesI/O (disk reads)Low CPU, depends on query complexity
Key insight: Most workloads are network-bound (waiting for AI APIs). CPU is rarely the bottleneck. This means CX (shared CPU) servers are appropriate even at moderate scale.

​
2. Simulation: 10 Users (Beta — Month 0-2)

​
User Behavior Profile

MetricValue
Total registered users10
Daily active users4-6
Concurrent users (peak)2-3
Discovery searches/day5-8
Contact reveals/day2-4
Page views/day100-200

​
Resource Consumption

ResourceUsageCapacityUtilization
Kadikoy CPU~5% avg, 15% peak4 vCPULow
Kadikoy RAM~2.5GB (Next.js + Node)8GB31%
Galata CPU~8% avg (Supabase + other sites)8 vCPULow
Galata RAM~10GB (Supabase + all services)16GB62%
Database size~500MB160GB<1%
API calls/day5-10 AI callsRate limit: 10/minMinimal
Bandwidth~2GB/day20TB/month<1%

​
Monthly Costs

ItemCostNotes
Infrastructure$23.80Fixed
AI API (240 searches/mo)$14.40240 × $0.06
AI API (90 reveals/mo)$2.7090 × $0.03
Total$40.90
Revenue$0Beta — free access
Net-$40.90Acceptable beta cost

​
Scaling Triggers: None

Everything comfortably within capacity. No action needed.

​
3. Simulation: 50 Users (Early Paid — Month 4-6)

​
User Behavior Profile

MetricValue
Total registered users50 (35 free, 12 Pro, 3 Team)
Daily active users15-25
Concurrent users (peak)5-8
Discovery searches/day30-50
Contact reveals/day20-40
Batch operations/week2-3 (Team users)
Page views/day500-1,000

​
Resource Consumption

ResourceUsageCapacityUtilization
Kadikoy CPU~12% avg, 35% peak4 vCPUComfortable
Kadikoy RAM~3.5GB8GB44%
Galata CPU~12% avg8 vCPULow
Galata RAM~11GB16GB69%
Database size~2GB160GB1.3%
API calls/day30-60 AI callsRate limit coversModerate
Bandwidth~8GB/day20TB/month1.2%

​
Monthly Costs

ItemCostNotes
Infrastructure$23.80Fixed
AI API (1,200 searches/mo)$72.001,200 × $0.06
AI API (900 reveals/mo)$27.00900 × $0.03
Stripe fees$22.00~15 transactions
Total costs$144.80
Revenue$1,15512 Pro × 59+3Team×59 + 3 Team × 149
Net profit+$1,01087.5% operating margin

​
Scaling Triggers: None Yet

All resources below 50% utilization on average. Database well within capacity. No upgrade needed.

​
4. Simulation: 200 Users (Growth — Month 9-12)

​
User Behavior Profile

MetricValue
Total registered users200 (120 free, 65 Pro, 14 Team, 1 Enterprise)
Daily active users60-90
Concurrent users (peak)15-25
Discovery searches/day120-200
Contact reveals/day80-150
Batch operations/week8-12
Page views/day3,000-5,000

​
Resource Consumption

ResourceUsageCapacityUtilizationStatus
Kadikoy CPU~25% avg, 60% peak4 vCPUModerateWatch
Kadikoy RAM~5GB8GB62%Watch
Galata CPU~18% avg8 vCPULowOK
Galata RAM~12GB16GB75%Watch
Database size~8GB160GB5%OK
Database connections~30 activeDefault pool: 10030%OK
API calls/day120-250 AI callsRate limit: 600/hrModerateOK
Bandwidth~25GB/day20TB/month3.8%OK

​
Monthly Costs

ItemCostNotes
Infrastructure$23.80Fixed — still on original servers
AI API (4,800 searches/mo)$288.00Growing
AI API (3,450 reveals/mo)$103.50Growing
AI API (batch, 500 ops/mo)$15.00Team/Enterprise users
Stripe fees$190.00~80 transactions
Total costs$620.30
Revenue$5,78565 Pro + 14 Team + 1 Enterprise
Net profit+$5,16589.3% operating margin

​
Scaling Triggers

TriggerThresholdActionCost Impact
Kadikoy CPU avg >50%ApproachingRescale CX33 → CX43 (8 vCPU, 16GB)+€5.40/mo (+$5.86)
Galata RAM >80%ApproachingMonitor — Supabase is the heavy consumer—
DB connections >60Not yetEnable PgBouncer connection pooling$0 (config change)
Rescale Kadikoy from CX33 to CX43 (~30/mototalinfrainsteadof30/mo total infra instead of 24/mo). This is a Hetzner “Rescale” operation — one-click, zero-downtime (virtual hardware upgrade).

​
5. Simulation: 1,000 Users (Scale — Month 24-36)

​
User Behavior Profile

MetricValue
Total registered users1,000 (520 free, 350 Pro, 110 Team, 20 Enterprise)
Daily active users300-450
Concurrent users (peak)50-80
Discovery searches/day500-800
Contact reveals/day300-500
Batch operations/day50-100
Page views/day15,000-25,000

​
Resource Consumption

ResourceUsageCapacity (after upgrade)UtilizationStatus
Kadikoy CPU~45% avg, 85% peak8 vCPU (CX43)HighUpgrade needed
Kadikoy RAM~10GB16GB (CX43)62%OK
Galata CPU~30% avg8 vCPU (CX43)ModerateOK
Galata RAM~14GB16GB87%Upgrade needed
Database size~40GB160GB25%OK
Database connections~80-120 activePool: 100 (needs PgBouncer)HighConfigure
API calls/day600-1,200 AI callsRate limit: 1,200/hrModerateOK
Bandwidth~80GB/day20TB/month12%OK

​
Monthly Costs

ItemCostNotes
Infrastructure (upgraded)$60.00Galata CX53 + Kadikoy CX53 (see below)
AI API (20,000 searches/mo)$1,200.00Largest cost center
AI API (12,000 reveals/mo)$360.00
AI API (batch, 2,000 ops/mo)$60.00Enterprise + Team
AI API (deep analysis, 500/mo)$50.00Enterprise users
Stripe fees$1,100.00~480 transactions
Monitoring tools (PostHog, Sentry)$50.00May exceed free tiers
Total costs$2,880.00
Revenue$35,200350 Pro + 110 Team + 20 Enterprise
Net profit+$32,32091.8% operating margin

​
Required Infrastructure Upgrades

ServerCurrentUpgrade ToNew SpecsNew Cost
GalataCX43CX5316 vCPU, 32GB RAM, 320GB~€22/mo
KadikoyCX43CX5316 vCPU, 32GB RAM, 320GB~€22/mo
Total infra~€50/mo ($54)

​
Additional Infrastructure Needs at 1,000 Users

NeedSolutionCostPriority
Connection poolingPgBouncer (already in Supabase stack)$0High
CDN for static assetsCloudflare (already configured)$0Already done
Background job queueBullMQ + Redis on Kadikoy$0 (on existing server)High
Search result cachingRedis on Kadikoy$0 (on existing server)Medium
Monitoring upgradePostHog Cloud (0−50/mo)+Sentry(0-50/mo) + Sentry (0-26/mo)$0-76/moMedium
Backup verificationAutomated restore tests (monthly)$0 (script)High

​
6. Scaling Roadmap

Users    Infra Action                                    Monthly Infra Cost
──────   ────────────────────────────────────────────   ──────────────────
0-50     Original config (CX43 + CX33)                  $24
50-200   Rescale Kadikoy CX33 → CX43                    $30
200-500  Rescale Galata CX43 → CX53                     $44
500-1K   Rescale both to CX53, add PgBouncer + Redis     $54
1K-5K    Add VPS3 (dedicated worker), separate DB        $80-120
5K+      Consider managed DB or Supabase Cloud           $200+

         $200 ┤                                              ████
              │                                         ████ ████
         $120 ┤                                    ████ ████ ████
              │                               ████ ████ ████ ████
          $54 ┤                          ████ ████ ████ ████ ████
          $44 ┤                     ████ ████ ████ ████ ████ ████
          $30 ┤                ████ ████ ████ ████ ████ ████ ████
          $24 ┤  ████ ████ ████ ████ ████ ████ ████ ████ ████ ████
            0 ┤
              └──10───50──100──200──500──1K───2K───5K───10K──20K──
                                Users

​
7. AI Cost Simulation

​
Cost by Provider Strategy

StrategySearches/moCost/SearchMonthly AI CostNotes
All Gemini Flash20,000$0.03$600Cheapest but no web search
Gemini + Perplexity20,000$0.06$1,200Current strategy (balanced)
All Perplexity Sonar20,000$0.10$2,000Best search quality, most expensive
Self-hosted OSS20,000$0.01$200Future: GPU VPS (~$50/mo)

​
Cost Reduction Strategies

StrategySavingsImplementation
Search cache hit rate (30-day TTL)20-40%Already built (export_ai_search_history)
Smart provider routing30%Use Gemini for classification, Perplexity only for web search
Batch classification (25/call)80% vs individualAlready built (AI batch pipeline)
Result deduplication10-15%Skip companies already enriched recently
Off-peak processing0% (no time pricing)N/A for API-based
Self-hosted models (future)60-80%Requires GPU VPS ($50-100/mo) at 5K+ users

​
Projected AI Cost at Scale

UsersSearches/moCache HitNet SearchesAI Cost/moAs % of Revenue
103005%285$17N/A (beta)
501,50015%1,275$776.7%
2006,00025%4,500$2704.7%
1,00030,00030%21,000$1,2603.6%
5,000150,00035%97,500$5,8502.8%
AI cost as % of revenue decreases at scale because subscription revenue scales linearly while AI cost benefits from caching and batch efficiencies.

​
8. Database Growth Simulation

​
Storage Projection

UsersCompanies in DBContactsSearchesEst. DB Size% of 160GB
102,5005003,000500MB<1%
5012,0003,00015,0002GB1.3%
20050,00015,00060,0008GB5%
1,000250,00080,000300,00040GB25%
5,0001,000,000350,0001,500,000150GB94%

​
Scaling Actions

ThresholdAction
DB > 100GBUpgrade to CX53 (320GB disk) or add volume
DB > 250GBConsider dedicated DB server or Supabase Cloud
1M+ companiesImplement data archiving for stale records
10M+ search_historyAggregate and purge raw records older than 6 months

​
9. Concurrent User Simulation

​
Next.js SSR Performance

Concurrent UsersAvg Response TimeServer Load (Kadikoy CX43)Status
1-10<100ms<10% CPUExcellent
10-30100-200ms15-30% CPUGood
30-60200-400ms30-50% CPUAcceptable
60-100400-800ms50-70% CPUUpgrade soon
100+>800ms>70% CPUUpgrade required

​
API Route Performance (AI-Calling)

RouteAvg LatencyBottleneckConcurrency Limit
POST /api/discover8-15sLLM response time10 req/60s per user
POST /api/headhunt5-10sWeb search API10 req/60s per user
POST /api/ai/classify3-8sBatch LLM call10 req/60s per user
POST /api/ai/complete2-5sLLM response20 req/60s per user
GET /api/companies<200msPostgreSQL queryNo limit
GET /api/contacts<200msPostgreSQL queryNo limit
AI routes are latency-bound by external APIs (8-15s), not server resources. Even at 1,000 users, only 50-80 are concurrent, and only a fraction are running AI queries simultaneously. The rate limiter prevents abuse.

​
10. Breakpoint Analysis

​
When Does Each Resource Hit Its Limit?

ResourceLimitBreakpoint (Users)ActionCost to Fix
Kadikoy CPU4 vCPU sustained >60%~150-200Rescale to CX43+$6/mo
Kadikoy RAM8GB >80%~300-400Rescale to CX43 (16GB)Included above
Galata RAM16GB >85%~500-800Rescale to CX53 (32GB)+$13/mo
DB storage160GB >80%~4,000-5,000Upgrade disk or CX53+$13/mo
DB connections100 concurrent~500-700PgBouncer pooling$0
Bandwidth (20TB)>60% utilization~10,000+Unlikely to hit$0
Stripe free tierN/A (pay per txn)NeverAlways pay-per-use% of revenue
PostHog free tier1M events/month~2,000-3,000 usersUpgrade to $0-50 tier$0-50/mo

​
Infrastructure Cost vs Revenue at Each Breakpoint

Revenue/Costs
     │
$35K ┤                                              ●  Revenue
     │                                         ●
$25K ┤                                    ●
     │                               ●
$15K ┤                          ●
     │                     ●
$10K ┤                ●
     │           ●
 $5K ┤      ●
     │  ●                                              ● Costs (incl. AI)
 $1K ┤● ─ ─ ● ─ ─ ● ─ ─ ● ─ ─ ● ─ ─ ● ─ ─ ● ─ ─ ● ─ ─ ●
   0 ┤
     └──10───50──100──200──500──1K───2K───3K───5K──
                        Users
Revenue grows much faster than costs at every scale. The gap widens because fixed costs are minimal and AI costs benefit from caching.

​
11. Disaster Scenarios

​
What If Traffic Spikes 10x Overnight?

ScenarioImpactRecovery
Viral LinkedIn post → 500 signups in a daySSR slows to 2-3s, some 503sHetzner Rescale (15 min) — bump Kadikoy to CX53
Trade fair demo → 100 discovery searches in 1 hourRate limiter kicks in (10/min/user), queue formsTemporarily increase rate limit, AI APIs handle it
Bot attack → 10,000 requests/minCloudflare WAF blocks, rate limiter rejectsAlready protected at network + app level
Database corruptionWAL-PITR restores to any point in timeMax 5 min data loss (WAL shipping interval)
Hetzner datacenter outageFull downtimeNo HA — accept ~4h/year downtime for cost savings

​
Acceptable Downtime Target

TierSLAMonthly Downtime BudgetAppropriate At
Current99.5%3.6 hours0-1,000 users
Growth99.9%43 minutes1,000-10,000 users (add HA)
Scale99.95%22 minutes10,000+ users (multi-region)

​
12. Cost Summary — All Scales

ScaleUsersInfra/moAI/moStripe/moTotal Cost/moRevenue/moMargin
Beta10$24$17$0$41$0-100%
Early50$24$99$22$145$1,15587%
Growth200$30$407$190$627$5,78589%
Scale1,000$54$1,670$1,100$2,824$35,20092%
Big5,000$120$6,850$5,000$11,970$175,00093%
Operating margin improves with scale. This is the hallmark of a well-architected SaaS: fixed costs become negligible, AI costs benefit from caching, and revenue scales linearly with users.

​
Summary

QuestionAnswer
Can current infra handle 200 users?Yes — one Kadikoy upgrade (~$6/mo more)
When is the first real upgrade needed?~200 users (Kadikoy rescale)
What’s the biggest cost at scale?AI API calls (~48% of costs at 1K users)
Is the cost model sustainable?Yes — margin improves from 87% to 93% at scale
When might we need a 3rd server?~2,000-5,000 users (dedicated worker VPS)
Can we handle a viral spike?Yes — Hetzner Rescale (15 min) + rate limiting
Next: INDEX.md — Document index and PM Knowledge Base guide