"Storage Is Cheap" Was a Loan, Not a Strategy

The keep-everything reflex rested on one load-bearing assumption: that storing it would always be cheaper than deciding about it. In 2026 the assumption sent a bill.

In early 2026, reporting on the enterprise storage market described high-capacity hard drives as effectively sold out: manufacturers' production booked through the calendar year, data-center buyers locking purchase orders into 2027 and 2028, and the steepest sequential price increases the industry had seen in nearly a decade as model-training demand consumed supply. The phrase 'storage is cheap, just keep it' quietly stopped being a law of nature and went back to being a market price.

This matters because keep-everything was never really a storage decision; it was a deferred ownership decision. Every dataset retained 'just in case' is a small unpaid bill: the cost to store it, the cost to secure it, the cost to remember why it exists. Those bills were tolerable only while the per-gigabyte line kept falling and hid them. When the storage curve stops cooperating, the deferred questions all arrive on the same invoice at once.

What it reveals is that the hoard was a confession, not an asset. Nobody hoards what they actively use; you hoard what you cannot bring yourself to delete because no one will own the deletion. The volume that looked like richness was mostly indecision with a retention policy. The FinOps Foundation's 2026 survey captured the new mood precisely: efficiency is now expected to fund the next thing, which means the cost of keeping the last thing is finally a number someone has to defend.

Watch for the rename. 'Data swamp' becomes 'training corpus'; the quarterly archive nobody reads becomes 'historical signal for the model.' Same bytes, same absent owner, new justification with better posture. The honest test is unchanged and slightly cruel: name the person who would notice if a dataset vanished. If you cannot, you are not storing it. You are paying rent on a question no one will answer.