Track 2: Early Release — Cloud Cost Estimate (2x Safety Buffer)¶
Phase: Production (1000 users) Region: ap-southeast-1 (Singapore) AI: Gemini Pro (primary) + DeepSeek (fallback) Note: All estimates include 2x safety margin for budget planning.
Monthly Cost at 1000 Users¶
| Component | Monthly | Notes |
|---|---|---|
| ECS Fargate (2 tasks) | $36-56 | Scale to 2 tasks for 1000 users |
| RDS PostgreSQL (db.t3.small) | $60-120 | Upgrade from micro for query perf. ~20GB storage |
| Gemini Pro (AI analytics) | $160-400 | ~10K analytics/insights requests/mo |
| DeepSeek API | $20-80 | Fallback + bulk analytics. $0.14/M input, $1.10/M output |
| Food Photo Analysis (Gemini Vision) | $1-30 | ~2K photos/mo. Flash = $0.80, Pro = $30 |
| Lambda (all functions) | $6-16 | Higher volume |
| API Gateway | $6-10 | ~5M requests/mo |
| S3 (uploads + food photos) | $10-30 | ~50GB (photos accumulate) |
| CloudWatch Logs | $10-20 | Production logging |
| ElevenLabs TTS | $44-198 | Creator $44 or Scale $198 |
| Amplify / S3+CF (web) | $10-30 | More traffic |
| Secrets + ECR | $6-10 | |
| Firebase FCM | $0 | Free |
| Total | $370-1,000/mo | 2x safety buffer applied |
| Per user/mo | $0.37-1.00 |
Cost Drivers¶
- Gemini Pro ($160-400) = 40-45% of total bill
- Food photos ($1-30) = negligible with Flash, moderate with Pro
- DeepSeek output pricing: $1.10/M tokens
- RDS upgrade to db.t3.small needed for analytics query performance
Optimization Strategies¶
- Use DeepSeek for bulk analytics (10-50x cheaper than Gemini Pro)
- Use Gemini Flash for food photos ($1/mo vs $30/mo for 2K photos)
- Cache AI results for 6 hours (T2-023) — reduces API calls 50%+
- Rate limit: 20 analytics/day per user (T2-024)
Production AI Provider Decision (TODO)¶
Gemini Pro is the current default, but consider DeepSeek or MiniMax as primary production AI to reduce the biggest cost driver ($160-400/mo → potentially much less).
| Provider | Input | Output | Vision | Notes |
|---|---|---|---|---|
| Gemini Pro | ~$1.25/M | ~$5.00/M | Yes | Current default. Expensive at scale. |
| DeepSeek (V3/R1) | $0.14/M | $1.10/M | Yes (VL) | 4-5x cheaper output. Already used as fallback. |
| MiniMax (M2.5) | TBD | TBD | TBD | Already paying $80/mo for dev tooling. Evaluate for production AI. |
| Gemini Flash | ~$0.075/M | ~$0.30/M | Yes | 15x cheaper than Pro. May be sufficient for most analytics. |
Decision needed before production: Benchmark DeepSeek and MiniMax against Gemini Pro for: - Health analytics quality (accuracy of insights, recommendations) - Food photo analysis accuracy - CornerMan AI coaching response quality - Latency (response time for user-facing features)
If DeepSeek or MiniMax quality is comparable, switching could drop the AI bill from $160-400/mo to $30-100/mo at 1000 users.