Assumptions (conservative, operator-realistic):
- S3 effective cost (storage + req + modest egress): $35/TB/year
- On-prem all-in amortized cost: $18/TB/year (after people + power)
- 4-year hardware amortization
- Stable, non-bursty workload
< 300 TB
| Metric | S3 | On-prem |
|---|---|---|
| Annual storage cost | ~$10k | ~$5k |
| Infra headcount | 0 | 0.25–0.5 FTE |
| People cost | $0 | $50k+ |
| Net economics | S3 wins | Loses |
Why:
You’re optimizing the wrong line item. One infra engineer costs more than the entire storage bill.
Operator note:
If you move off S3 here, it’s ideology, not economics.
300–500 TB
| Metric | S3 | On-prem |
|---|---|---|
| Annual storage cost | $15k–18k | $9k |
| Total effective cost | ~$20k | ~$60k |
| Savings | ❌ | ❌ |
Why:
You can run on-prem, but it’s a rounding error compared to staffing + risk.
Signal:
If storage outages won’t kill your business, you’re still too small.
~1 PB (the gray zone)
| Metric | S3 | On-prem |
|---|---|---|
| Annual storage | ~$35k | ~$18k |
| Egress (10%) | ~$110k | ~$5k |
| People + ops | $0 | ~$120k |
| Total | ~$145k | ~$143k |
Why it “depends”:
- With low egress → S3 still fine
- With steady egress / analytics → on-prem starts to win slightly
Operator truth:
This is where finance says “maybe” and infra says “are you sure?”
🔸 3–5 PB (clear crossover)
Example: 5 PB
| Metric | S3 | On-prem |
|---|---|---|
| Storage | ~$175k | ~$90k |
| Egress | ~$500k | ~$25k |
| Infra team | $0 | ~$250k |
| Annual total | ~$675k | ~$365k |
| Delta | ~$310k saved/year |
Why on-prem wins:
- Fixed costs amortize
- Egress dominates cloud bills
- Marginal cost per TB collapses
This is the “Jeremy zone.”
🔹 10+ PB (no contest)
Example: 10 PB
| Metric | S3 | On-prem |
|---|---|---|
| Storage | ~$350k | ~$160k |
| Egress | ~$1M+ | ~$50k |
| Infra team | $0 | ~$350k |
| Annual total | ~$1.35M+ | ~$560k |
| Savings | $750k–900k/year |
Why S3 breaks here:
- You’re paying a cloud tax on your core asset
- Cost scales linearly forever
- No negotiating leverage
On-prem at this scale becomes a moat.
🧠 Mental model to remember
S3 is O(n).
On-prem is O(1) + O(people).
Once people cost flattens, every extra PB massively favors on-prem.
⚠️ Final operator warning
On-prem only wins if:
- Storage is mission-critical
- You run infra like a product
- Downtime is unacceptable
- You think in 5-year horizons
Otherwise:
S3 is expensive — but organizational failure is worse.