For a decade, the rule was "never build your own data centers — the cloud is cheaper." Dropbox got big enough to discover the rule has an expiration date, and walking away from Amazon S3 saved them tens of millions.
Storing customer files on S3 was the right call when Dropbox was small. At exabyte scale, two problems emerged: the bill was enormous, and renting storage meant they couldn't tune the full stack for their very specific, very large workload. So they built Magic Pocket, their own storage system.
The cloud is like renting an apartment: zero upfront cost, someone else handles maintenance, perfect when you're starting out. But if you're going to occupy an entire city's worth of space forever, at some point buying — building your own — becomes cheaper than renting, and gives you control to customise everything.
This crossover only happens at genuinely massive scale. For almost everyone, the cloud stays cheaper. Dropbox crossed it because storage is their product, at a size where every efficiency multiplies across exabytes.
Magic Pocket is a custom immutable block store. Files are split into blocks of up to 4MB, each addressed by its SHA-256 hash (so identical blocks dedupe automatically and content can't change under a given address), then aggregated into 1GB buckets for efficient handling. The clever economics live in how durability is achieved at different ages of data.
Fresh data uses n-way replication — just keep several full copies. Simple and fast to write, but storage-hungry. As data ages and access cools, it's converted to erasure coding, which achieves the same durability using far less raw storage (think: store math that can reconstruct lost pieces, instead of whole duplicates). They run on HDDs and, from 2018, the first petabyte-scale deployment of SMR drives — another density-for-cost trade.
Worth knowing
Be precise about that dollar figure: the ~$74.6M two-year operating-cost saving comes from Dropbox's SEC S-1 filing and press analysis, not the engineering blog — cite it as such. The durable engineering lesson stands on its own: owning the stack let them tier durability by data temperature, a knob the cloud abstraction simply didn't expose.
The gap it reveals
The industry mantra is 'cloud is always cheaper, never build your own.' The senior realisation is that this is a scale-dependent claim with a crossover point — and that at extreme scale, owning infrastructure unlocks workload-specific optimisations (age-based replication vs. erasure coding, custom drives) that a rented abstraction hides. Knowing the rule and its expiration date is the judgment.
In the interview room
If a design hinges on cost at huge scale, the impressive note is nuance about build-vs-buy: "cloud storage is right until volume makes the crossover to owned infrastructure worthwhile — then you can tier durability by data age, replicate hot data and erasure-code cold." Most candidates treat 'use the cloud' as axiomatic; showing you know when it stops being true signals real cost reasoning.
The reframe
Best practices are usually scale-dependent claims with the scale left unstated. 'Don't build your own data center' is excellent advice — until you're storing exabytes, when it quietly inverts. The skill isn't memorising the rule; it's knowing which variable (here, total volume) decides when the rule flips, and recognising when you've crossed it.
The cloud is cheaper — until it isn't. Know where your crossover lives.
Primary source →
dropbox.tech — Magic Pocket Infrastructure