HomeResourcesCase study
Case study

When Dropbox Left the Cloud

By The SDL team·4 min read·Updated Sep 16, 2026

"Cloud is always cheaper" is a scale-dependent claim with an expiration date. At exabyte scale, Dropbox built its own and saved millions.

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.

Plain English

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.

Now the engineering

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.

When you’re big enough, renting becomes more expensive than owning store on Amazon S3 costly + un-tunable at exabytes Magic Pocket: custom block store immutable ≤4MB SHA-256 blocks → aggregated into 1GB buckets Fresh data: n-way replication simple, fast to write, more storage cost → converts to erasure coding as it ages Cold data: erasure coding same durability, far less raw storage HDDs + (2018) petabyte-scale SMR drives Migrated 90%+ of user data in ~2.5 yrs; designed for 99.9999999999%+ annual durability ~$74.6M operating-cost savings over two years (per Dropbox’s SEC S-1, not the blog) Owning the storage stack unlocked tuning the cloud could never expose.
Match the durability strategy to the data's age. Replicate hot data for speed, erasure-code cold data for cost — a tiering Dropbox could only tune because they owned the stack.
90%+
user data migrated in ~2.5 yrs
12 nines
designed annual durability
~$74.6M
2-yr savings (per S-1)

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

Want feedback on your design?

The weekly teardown

One real-world architecture, every week.

How real companies actually built it: the design, the trade-offs, and what to say about it in an interview. Free, and one click to unsubscribe.

Related articles