November 9, 2026 · Infrastructure
Storage is the bill nobody reads
Dropbox just paid $350 million for object storage. If you run agents, you already have the same problem — and unlike tokens, it compounds without showing up on any dashboard.
Dropbox's $350 million purchase of Tigris Data is being reported as an AI infrastructure bet, which it is. The more useful reading is what it says about where the money goes once a company starts running agents.
Tigris does object storage: S3-compatible, multi-region, placing data near the compute that reads it. Dropbox is folding it into something called DBX Agent Cloud, sold as storage you can buy on its own. A file-sync company is now selling buckets.
Why storage, of all things
Because agents are prolific writers. One task can leave behind a transcript, a tool-call log, a handful of screenshots, intermediate files, an embedding set and a snapshot of the workspace. Multiply by a few thousand runs a month and the storage line grows faster than the model line ever did.
And unlike tokens, storage does not announce itself. There is no per-request meter on a dashboard, no friendly usage chart that turns amber. You find out at the end of the quarter, when an invoice arrives with a number nobody can explain.
Three numbers that actually matter
Stored. Total gigabytes, and the growth rate month over month. Flat storage with growing usage usually means retention is doing its job. Growing storage with flat usage means something is accumulating that nobody owns.
Egress. What you pay to read it back. This is where object storage gets expensive, and where the architecture decision hides. If your compute and your buckets sit in different regions or on different vendors, every read is a line item.
Requests. Per-operation charges look trivial until they are not. Agents are chatty: they list, stat and re-read the same objects constantly, because that is what a loop does.
The trap called memory
Every product now wants to remember. Memory means a store that grows with every conversation, plus a retrieval path that reads it back on every turn. It is a good feature and a compounding cost, and most teams ship it before deciding what gets forgotten.
The fix is dull: a retention rule per object type, written down where an engineer can find it. Transcripts ninety days. Tool logs thirty. Screenshots seven, unless flagged. Embeddings rebuilt rather than kept forever. Deleting has to be a feature with a button, not an incident.
Run the arithmetic once
Take a modest agent deployment: two thousand tasks a day. Each leaves a 40 KB transcript, five screenshots at 300 KB, and a few hundred kilobytes of logs and intermediate files. That is roughly 1.7 GB a day, or 600 GB a year before versioning and indexes, which usually triple it.
At typical object storage rates, 1.8 TB costs less than a round of coffees each month. So storage is not the problem. Egress is. If each task reads 20 MB of context back out, you are moving 40 GB a day across the wire, and at ten cents a gigabyte that is four dollars a day doing nothing but looking things up.
That is the shape of the bill nobody budgets for: a small storage line, a growing retrieval line, and a request count that nobody monitors because it never looked important.
Why this is an acquisition and not a launch
Because the margin lives in the plumbing. Storage is cheap to run at scale and expensive to run badly, and whoever owns the pipes can price below what you would pay to stitch the same thing together from three vendors.
It also hints at where the business is going. Dropbox cut nine percent of its workforce last month and is now buying infrastructure. That is a company trading headcount for a substrate, and it is a bet that the next decade of the product is less about the interface and more about what sits underneath it.
What to do this quarter
Pull the storage bill and split it by product. Most teams cannot, and that is the first finding. Then take the three largest prefixes and ask who owns each. At least one will belong to a feature that was quietly deprecated last year.
Then adopt one rule: nothing is stored without an expiry date. Not because storage is expensive today — it is not — but because retrieval, backup and migration costs all scale with the pile, and those are the ones that hurt the week you decide to change vendors.
Compute gets the attention because it is where the model runs. Storage is where the business accumulates. One is a bill you can forecast. The other is a bill you discover.
We publish the pricing model, minimum commitment and export-readiness notes for every tool in the library, and we flag it when a vendor changes one without changing the headline price.