The $15/Month Stack
Running 6 products on $15/month sounds like clickbait. It is not. Here is every service I use, what it costs, and — honestly — what breaks.
The Stack
Cloudflare Workers + Pages — $0/mo
What it does: Hosts static sites (Astro, Next.js export), runs edge functions, handles DNS, provides CDN.
Free tier: 100K requests/day, 10ms CPU time per invocation, 100 custom domains.
What breaks: The 10ms CPU limit is real. You cannot do heavy computation at the edge. Cold starts are rare but exist. Workers KV has a 1K write/second limit per namespace.
When to upgrade: If you need longer execution times or more than 100K daily requests. The $5/mo plan is generous.
Vercel — $0/mo
What it does: Hosts Next.js apps (InkCloak frontend) with serverless functions.
Free tier: 100GB bandwidth, 100 hours serverless execution, automatic HTTPS.
What breaks: Serverless cold starts on free tier can hit 2-3 seconds. No persistent connections. Build times are limited to 45 minutes.
When to upgrade: If you exceed bandwidth or need server-side rendering at scale. Pro is $20/mo.
Supabase — $0/mo
What it does: PostgreSQL database, authentication (Google OAuth, email), row-level security.
Free tier: 500MB database, 50K monthly active users, 1GB file storage.
What breaks: The 500MB database limit is tight if you store embeddings or large JSON blobs. Free projects pause after 1 week of inactivity.
When to upgrade: When your database exceeds 500MB or you need always-on availability. Pro is $25/mo.
Groq — $0/mo
What it does: Ultra-fast LLM inference. I use Llama 3.3 70B for chat and Qwen 2.5 for code generation.
Free tier: 14,400 tokens/minute, 30 requests/minute, no daily limit.
What breaks: Rate limits hit hard during burst traffic. The 30 req/min limit means you need queuing for concurrent users. Model availability can be spotty — sometimes a model is “unavailable” for hours.
Workaround: I use a Cloudflare Worker as a proxy (Groq blocks Russian IPs) and implement automatic fallback to OpenRouter free models.
When to upgrade: The $0.27/M token paid tier removes rate limits. Worth it above 50 concurrent users.
PostHog — $0/mo
What it does: Product analytics, session replay, feature flags, A/B testing.
Free tier: 1M events/month, 15K session replays, unlimited feature flags.
What breaks: Honestly, nothing yet. The free tier is absurdly generous for a solo maker. The main limitation is data retention (1 year on free).
When to upgrade: Above 1M events/month. But for 6 small products, I am nowhere near that.
Loops.so — $0/mo
What it does: Email marketing. Drip campaigns, transactional emails, subscriber management.
Free tier: 1,000 contacts, basic automation.
What breaks: The 1K contact limit is the real constraint. No custom domains on free tier (emails come from loops.so). Limited API access.
When to upgrade: When your list exceeds 1K contacts. Starter is $49/mo — steep for a solo maker.
Sentry — $0/mo
What it does: Error monitoring and performance tracking across all products.
Free tier: 5K errors/month, 10K performance transactions.
What breaks: The free tier lacks good alerting rules. You get email notifications but not Slack/Telegram integration.
When to upgrade: The $26/mo team plan adds better alerting and more volume.
RunPod — pay per use
What it does: GPU cloud for ML training and inference. Used for fine-tuning DeBERTa models.
Cost: $0.49 for my last training session (spot A100, 20 minutes).
What breaks: Spot instances can be preempted mid-training. You need checkpointing. The UI is rough but functional.
NVIDIA P100 — $150 one-time
What it does: Local GPU inference for development and light production workloads.
What breaks: It is an old card (Pascal architecture). No bfloat16 support, 16GB VRAM limits model size. Power consumption is high for 24/7 operation.
What I Saved
If I used typical paid alternatives:
| Service | Paid Alternative | Cost |
|---|---|---|
| Hosting | AWS EC2 + CloudFront | $200/mo |
| Database | RDS PostgreSQL | $50/mo |
| LLM API | OpenAI GPT-4 | $100/mo |
| Analytics | Mixpanel | $100/mo |
| Mailchimp | $30/mo | |
| Monitoring | Datadog | $50/mo |
| GPU cloud | Lambda Labs (always-on) | $300/mo |
| Total | $830/mo |
That is $830/month I am not spending. Over a year, that is almost $10K saved.
The Honest Truth
Free tiers have real limitations:
-
Rate limits force you to think about architecture. You cannot just throw requests at an API. You need caching, queuing, and graceful degradation.
-
Cold starts affect user experience. First-time visitors to serverless functions wait 2-3 seconds. Returning visitors get sub-second responses.
-
You are one growth spike away from hitting limits. If a product goes viral, you need to upgrade within hours. Have your credit card ready.
-
Debugging is harder. Free tier logging is minimal. When something breaks at 3am, you are reading Cloudflare Workers logs with no tracing context.
-
Some free tiers are loss leaders. They want you to upgrade. Product quality on free tiers can degrade over time as companies focus on paid customers.
When to Stop Being Cheap
Upgrade when:
- Your product has paying customers (use their money, not yours)
- Rate limits are causing user-visible errors
- You are spending more time working around limits than building features
- Uptime matters (free tiers have no SLA)
Stay free when:
- You are validating an idea
- Traffic is under 1K daily users
- You are the only user who notices downtime
- The product is not generating revenue
The $15/month stack is not a permanent solution. It is a launch strategy. Build cheap, validate fast, upgrade when the money flows.