Dropbox moved file storage off Amazon S3 onto its own infrastructure - Magic Pocket - completing the migration in 2016. What did that decision buy and what did it pay, and which facts about your own situation would have to be true before the same move is defensible?
Show the full answer Hide the answer
What is gained
The gain is unit cost at a scale where storage is the dominant line item. Dropbox's product is storage, so the cost of a stored byte is not an infrastructure expense sitting beside the business — it is the business's cost of goods sold. At exabyte scale, the margin between a hyperscaler's price and the amortised cost of disks, racks, power and the engineers who run them is large enough to move gross margin by points, and points of gross margin at that revenue are worth more than most product work.
The second gain is design freedom: a storage layer built for one known access pattern can make choices a general-purpose service cannot, in erasure coding, placement and compaction.
What is paid
- A permanent engineering organisation. Not a project. Durability at that scale is a standing team with an on-call rotation, a hardware supply chain, and a decade of maintenance ahead.
- Durability risk transferred onto you. You are now the party that loses customer data if the erasure coding, the repair pipeline or the capacity forecast is wrong. The cloud provider's durability guarantee was a product you were buying; you have stopped buying it.
- Capital and forecasting. Hardware is bought ahead of demand. Over-forecast and the capital is idle; under-forecast and you cannot serve growth, on a lead time measured in months rather than an API call.
- Opportunity cost. The strongest infrastructure engineers in the company are now working on storage.
The crossover conditions
Four things had to be true, and all four have to be true for you:
| Condition | Why it is load-bearing |
|---|---|
| The workload is the product's core cost driver | Otherwise you are optimising a line item that does not move the P&L |
| The scale is large enough that a few percent of unit cost exceeds the cost of the team | Below that, the team is more expensive than the saving |
| The access pattern is narrow and stable | A general-purpose service wins whenever the pattern is still changing |
| You can fund a decade of operation, not a migration | Durability obligations do not end when the migration does |
The flip is sharp. Move the same reasoning to a company where storage is 4% of spend and the arithmetic inverts: a team of fifteen costs more than the entire saving, forever.
When this is the wrong answer
Almost always, and specifically: when the workload is not the core cost driver, when you are reasoning from list price rather than your negotiated price, and when the motivation is lock-in anxiety rather than a number. "We are worried about lock-in" is not a crossover condition — exit cost is real, but it is paid once, and the standing cost of self-operation is paid every year.
Common weak answers
- "They saved money by leaving the cloud." True and useless without the scale and the product fit that made it true. Copying the conclusion without the conditions is how a Series B company ends up running disks.
- "Cloud is more expensive." Per unit at scale, sometimes. Per unit including the people, the capital and the risk, usually not until you are very large and very specialised.