How Dropbox left Amazon S3 — and what the '$75M saved' claim actually rests on

What actually happened?#

Dropbox moved the large majority of its file storage off Amazon S3 and onto infrastructure it built itself. In a post on its engineering blog, Scaling to exabytes and beyond, the company announced that a project it called Magic Pocket — after Dropbox's original nickname — had, after two and a half years of work, reached the point of "storing and serving over 90% of our users' data on our custom-built infrastructure."

A later Dropbox post documents the first petabyte-scale deployment of shingled magnetic recording (SMR) drives as an extension of the same system.

The story is circulating again because Cloud Codes published a video breakdown on 12 August 2026 — "Cloud Is a Trap: How Dropbox Left AWS and Saved $75M" — arguing the migration's finances are more complicated than the headline saving suggests. That video is how this story reached us; the sections below separate what Dropbox itself documented from what the video reports.

What do Dropbox's own posts say about Magic Pocket?#

The documented facts are narrower than the coverage. From the resolved announcement: the project exists, it is called Magic Pocket, it took two and a half years, and by announcement time it was storing and serving over 90% of users' data on custom-built infrastructure. In Dropbox's words: "when our scale required building our own dedicated storage infrastructure, we named the project 'Magic Pocket.'" The portion of the post available here gives no architecture and no costs.

The SMR post adds one more documented fact via its title: a petabyte-scale SMR drive deployment extending Magic Pocket. Its full text was not available in this evidence bundle.

The specific machinery quoted in the coverage — 4 MB encrypted blocks sharded across erasure-coded cells, a sharded MySQL index, a storage engine rewritten in Rust — comes from Cloud Codes' summary (1:24, 2:10), not from the portion of Dropbox's posts available here. The same applies to the "Diskotech" 4U chassis and a claimed 2.29× density gain from SMR drives (2:40): the Dropbox hardware post the video links could not be resolved, so those details rest on the reporter's telling alone.

What does the a16z essay argue?#

The Cost of Cloud, a Trillion Dollar Paradox (Andreessen Horowitz, 27 May 2021) supplies the economic framing the current coverage leans on. Its thesis: "the pressure the cloud puts on margins can start to outweigh the benefits" as a company scales, and the essay sets out to quantify how much market capitalisation cloud spending suppresses in order to "inform the decision-making framework on managing infrastructure as companies scale."

Two things are worth holding onto. The essay is a venture firm's argument and framework, not an audit of any particular company. And its claim is conditional — repatriation pays at scale and as growth slows — which is exactly the condition the Dropbox case is used to illustrate.

Where do the savings numbers come from?#

Every dollar figure in the current coverage traces to one source: Cloud Codes' reading of Dropbox's S-1 SEC filing — "page 73," per the video (0:00). The filing itself is not among this article's resolved sources, so all of the following is reported, not verified:

• a $92.5M reduction in the AWS bill set against roughly $53M in datacenter capital expenditure (4:32);
• gross margin rising from 33% to 67% (5:45);
• a headline savings figure of $74.6M, which the video itself argues ignores the $53M of hardware depreciation that had to be absorbed "before a single dollar was saved";
• roughly 500 petabytes of data moved off S3, per the video's description;
• Dropbox retaining AWS usage in Europe and Asia after the migration (7:42).

Chart drawn from the source's reported figures. Data: https://www.youtube.com/watch?v=m5NLYxscKio.

Chart drawn from the source's reported figures. Data: https://www.youtube.com/watch?v=m5NLYxscKio.

The video also proposes five conditions an enterprise should meet before leaving public cloud. That checklist is the reporter's framework, in the spirit of the a16z essay, not an audited standard.

Where does the coverage go beyond the primary sources?#

Three gaps stand out. First, Dropbox's engineering posts publish no dollar figures at all; the entire financial narrative — savings, margins, depreciation — rests on one reporter's reading of a filing this article could not verify.

Second, the numbers as reported do not reconcile on their face. A $92.5M gross reduction minus $53M in hardware cost is $39.5M, not the $74.6M headline figure, which suggests the numbers cover different periods or bases that the video does not specify.

Third, the video's own thesis undercuts its title. Its argument is precisely that the gross saving overstates the net benefit once depreciation is counted — which makes "Saved $75M" a headline the video's content partially retracts.

On the other side, the primary material confirms the story's engineering core: the migration happened, at greater-than-90% scale, and SMR hardware followed. The coverage is building on a real, documented event rather than inventing one.

What are the limits of this story?#

• The S-1 is not among the resolved sources. Every financial claim here is single-reporter. Reading the filing directly is the one step that would firm this story up.
• One of the Dropbox hardware posts linked in the coverage could not be resolved, and the SMR post's full text was unavailable here; the hardware specifics (Diskotech chassis, 2.29× density, erasure-coding layout) are unverified.
• The a16z essay is an investor's framework published in 2021. It argues a thesis about cloud economics generally; it is not evidence about Dropbox's accounts in particular.
• The coverage in this bundle is one channel. No second independent reporter corroborates the numbers.

Sources#

Primary sources#

• Scaling to exabytes and beyond
• The Cost of Cloud, a Trillion Dollar Paradox | Andreessen Horowitz — 2021-05-27
• Extending Magic Pocket Innovation with the first petabyte scale SMR drive deployment

Reporting that surfaced this#

• Cloud Codes — Cloud Is a Trap: How Dropbox Left AWS and Saved $75M (2026-08-12)