## Money the site itself is always free to host for unlimited visitors monthly costs scale only with how many voice notes are sent in own GPU = lowest cost but more work Cloudflare = simpler but about 2× the cost either way, global intake and processing only costs a few hundred dollars/month at peak volume pilot ~$1.1k domain snag ~$8/month hosting WhatsApp intake free voice notes cost fractions of a cent scale own GPU path → 100k notes/month ≈ $225 Cloudflare path → 100k notes/month ≈ $425 infra glue (workers, queues, R2) → ~$50/month | Cost | Pilot | At Scale (100k notes/mo) | |----------------------------------|------------------------------|---------------------------------------------| | Domain | ~$1,054 one-off | ~$25/year renewal | | Hosting | ~$8/month (Obsidian Publish) | Free (GitHub Pages + Cloudflare) | | WhatsApp number | ~$3/month | ~$3/month | | Voice processing | ~$0.003/note | Own GPU: ~$225/mo; Cloudflare: ~$425/mo | | Infra glue (workers, queues, R2) | ~$5/month | ~$50/month | | Storage (summaries only) | ~$0.01/month | ~$0.06/month | | **Monthly total** | **~$16/mo** (excl. domain) | **Own GPU: ~$280/mo; Cloudflare: ~$480/mo** | ### Baseline site costs domain GoDaddy snag (one-off) → ~$1,054 renewal → ~$25/year hosting phase one → Obsidian Publish (~$8/month) phase two → GitHub Pages + Svelte + Cloudflare (essentially free) intake WhatsApp user-initiated → free number rental only (~$3/month) ### Transcribing and summarising voice notes assumptions voice notes are capped at 10 minutes average length is ~6.5 minutes costs below assume typical notes of around 6.5 minutes option A – own GPU we rent a pool of GPUs (graphics cards) to do both transcription and summarising per note cost → ~$0.0025 (fractions of a cent) processing time → ~4 minutes for a 6.5 min note upgrade path → if load grows, one pool can handle transcription and another can handle summaries storage → nothing kept beyond the short summary text option B – Cloudflare Workers AI we send the note to Cloudflare’s servers, which run speech-to-text and then a small model to summarise per note cost → ~$0.0035 (still fractions of a cent) processing time → close to real-time, depending on demand overhead → a few dollars/month for the serverless “glue” (queues, workers) storage → same as Option A, only short summaries kept cost comparison | Volume | Hours audio (avg 6.5 min each) | Own GPU (T4 cluster) | Cloudflare Workers AI | |---------------|--------------------------------|----------------------|-----------------------| | 1,000 notes | ~108 hrs | ~$2.50 | ~$3.50 | | 100,000 notes | ~10,833 hrs | ~$225 | ~$425 | notes own GPU = cheapest but needs more setup and maintenance Cloudflare = about 2× the price, but simpler and always available we'd prefer the cloudflare option (option B) and we'd retire it as the thing grew larger #### Voice infrastructure could use whatsapp transport messages arrive through Meta’s Cloud API and go straight to a Cloudflare Worker replies are free as long as we answer within 24 hours of the user’s last message every time the user sends something, the 24-hour clock resets if we want to message them after that window, Meta charges per message (new rules from Jul 2025) storage raw audio is deleted immediately transcripts are also deleted immediately only summaries are stored → even at 10M notes total, that’s just ~4 GB compressed storage cost ≈ **~$0.06/month** bandwidth traffic between WhatsApp and Cloudflare is free site visits are also free, no matter how many people come