AI Strategy · Data & Archives

Keep the Bronze: Cheap Comprehension Just Repriced Every Archive You Own

I deleted a twenty-year mailbox because it was unreadable dead weight. That decision was correct when I made it — and it is wrong now, and I didn’t do anything to make it wrong. The economics moved under the decision. They just moved under yours too.

By Scott Farrell · LeverageAI

Somewhere in the last year I deleted a Lotus Notes NSF file. Twenty years of email inside it, from my time at IBM in 1995 onward — clients, negotiations, decisions, the whole texture of a working life. I hadn’t opened it in ages. The Notes clients were Java, the machine was ancient, the software had fallen apart around the format, and even when it did run, the search was useless: full-text, sure, but you had to already know what you were looking for. It was dead weight. So I got rid of it.

I wish I hadn’t.

Here is the uncomfortable part: that deletion was correct at the time. Given what an NSF file was worth the day I deleted it — an unreadable container, a useless search, no path to meaning — clearing it out was the sensible, disciplined, tidy-minded thing to do. I wasn’t careless. I was right. And I’m still holding the loss, because the premises my decision rested on changed underneath it. The file didn’t get more valuable because I did something clever. It got more valuable because comprehension got cheap.

That is the whole argument, and it applies to you: the delete-or-keep decisions you made under the old economics have been silently invalidated. Every old drive, every dead-format mailbox, every shoebox of DV tapes in the cupboard just got repriced — and nobody sent you the notice.

Why the old decision was right (and why that matters)

For decades, the reasonable person’s syllogism about an unreadable archive went like this: I can’t open it → I can’t search it → it has no usable value → delete it. Every step of that was true. The industry even had a name for where the data went: the digital dark age — “a lack of historical information in the digital age, which is a direct result of outdated file formats, software, or hardware.”¹ Note what that definition blames. Not the bytes. The container.

Stewart Brand made the point cleanly back in 1999, and it stood unchallenged for a quarter-century:

“Digital storage is easy; digital preservation is not. Preservation means keeping the stored information cataloged, accessible, and usable on current media, which requires constant effort and expense … there is no business case for archives.”²

“No business case for archives.” That was the load-bearing assumption under every purge decision anyone ever made. Your files might be perfectly intact, Brand wrote, and still “as unrecoverable as if they never existed” — WordStar in CP/M on a Kaypro, a 5¼″ floppy nobody can read.² The information survives; the way in dies. And if there’s no way in and no economic case to build one, then keeping the file is just paying rent on a coffin.

I want to be fair to that logic because it was good logic. This isn’t a story about people being foolish hoarders who should have known better. It’s a story about careful people making correct decisions whose premises then expired. That’s a more unsettling situation than carelessness, because you can’t fix it by being more diligent. You can only fix it by noticing that the price changed.

The repricing: two costs collapsed, and value moved

Two things happened, close together, and between them they dismantled every step of that old syllogism.

First, comprehension got cheap — absurdly cheap. The cost of running a language model over a pile of text has fallen off a cliff. Stanford’s AI Index puts numbers on it: the cost of querying a model at GPT-3.5 quality “dropped from $20.00 per million tokens in November 2022 to just $0.07 per million tokens by October 2024 — a more than 280-fold reduction in approximately 18 months.”³ Andreessen Horowitz calls it LLMflation: for a model of equivalent performance, “the cost is decreasing by 10x every year … The cost of LLM inference has dropped by a factor of 1,000 in 3 years.”⁴ Reading and understanding a mailbox — the expensive part — went from a research budget to a rounding error.

Second, access got cheap too. The dead-container problem — no client, no driver, no vendor — used to be terminal. It isn’t anymore, because an AI agent can now synthesise the adapter on demand: go and find what the internet knows about a format, write a driver against it, and open a container whose entire ecosystem died fifteen years ago. Format obsolescence just stopped being a death sentence. (I watched this happen to my own NSF files this month — there is no Python connector for Notes, so the agent wrote its own in C++ and opened them. That capability is a whole story of its own, which I told separately in The Backup Is the API; I’m not going to re-derive it here.) The point for this argument is narrow and large: every “dead” archive on earth got repriced again, not just cheap to comprehend but cheap to reach.

Value was never in the data. It was in the activated data — and activation just became an industrial process you can run whenever you want.

This is the piece people miss. An archive’s worth was always conditional on an activation step — opening it, understanding it, making it searchable in the “ask it anything” sense. That step used to be impossible or uneconomic, so the latent value read as zero, so deletion felt free. Now activation is cheap and repeatable. In the language I’ve used before, this is intelligence turning from an operating expense into a capital asset: you spend comprehension once and keep the compiled understanding forever. Which means the raw archive is no longer junk — it’s feedstock. As I put it to myself in the middle of all this: even if you don’t think you want the information, once it’s wikified it’s actually useful.

And the repricing is retroactive. It doesn’t wait for new archives; it re-scores decisions you already made. Within a day of working through this argument, I’d gone from regretting a deleted mailbox to running a salvage operation reconstructing a corpus back toward 1995. Same me, same files, opposite decision — because the only thing that changed was the price of understanding.

Keep the bronze

Here’s where the data engineers already have the right vocabulary, and it gives this rule its name. The medallion architecture organises data into three tiers: “bronze (raw), silver (validated), and gold (enriched).”⁵ The bronze layer is the landing zone — “raw, unvalidated data” stored almost verbatim, deliberately with minimal cleanup, precisely so nothing is thrown away before you know what you’ll need.⁵ Its whole job is to “preserve the original state of the data for auditing, compliance, or reprocessing.”⁶ Silver and gold — the clean, structured, queryable layers — are built on top, later, cheaply, and as many times as you like. I’ve written before about the bronze tier for organisational data; the point here is that the rule scales all the way down to a person and a cupboard of old drives.

So the rule is exactly that: keep the bronze. Keep the raw layer — all of it, immutable, cold, off to the side. The map (the searchable, structured, activated version) can always be built later. But it can only ever be built over territory that still exists.

Storage is cheap, comprehension is now cheap — so deletion is the only irreversible operation in the whole stack.

That’s the entire asymmetry. Keeping is reversible: if you never activate an archive, you’ve lost a trivial storage cost. Deleting is irreversible: you’ve foreclosed every future activation of that data, forever, to save that same trivial cost. When one side of a bet is “a few dollars” and the other side is “gone for good,” you don’t need the option to pay off often for keeping to be correct.

The storage-cost sanity check

“Keep everything” sounds like hoarder-talk, so make it disciplined with one comparison: the cost of keeping the bronze vs. the option value of a future activation. Do the first half honestly and the second half will usually dwarf it.

The cost of keeping has been falling for forty years on a curve so regular it has a name — Kryder’s Law, the storage equivalent of Moore’s Law.⁷ Backblaze, who buy hard drives by the tens of thousands, put hard numbers on it: the cost of a gigabyte fell “from over $1 million in 1981 to less than $0.05 in 2011,”⁸ and their own per-gigabyte drive cost dropped 87.4% between 2009 and 2022, to around $0.014 per gigabyte.⁹ At cloud-archive rates you can now sit raw data at roughly $6 per terabyte per month, with the first slice free.¹⁰ A lifetime of email, documents and photos is measured in tens of gigabytes; keeping it costs about what you’d tip a barista.

Set that against what you’re protecting. Industry estimates hold that roughly 80% of all data is unstructured and around 90% of it is never analysed;¹¹ most of what organisations store is “cold” — 60–70% of enterprise data hasn’t been touched in over 90 days.¹² The old read on those numbers was “you’re wasting money storing junk.” The new read is that the map simply hasn’t been built yet — the territory is enormous and the activation era just started. The one instinct to drop is the fatalistic one Gartner’s analysts describe: storing everything “just in case, assuming storage is cheap and analysis can come later. But later rarely comes.”¹³ That was fair warning in 2015. In 2026, “later” arrived. Activation is the thing you can now actually do.

Mount it read-only: the archaeologist’s rule

“Keep the bronze” needs one operational discipline or it’s just sentiment. Treat every old archive like a dig site: the site is read-only. Mount those disks with the write path off. Extract into scratch space. Do all your census, dedupe and sifting on copies. Let deletion remain the one operation nothing automated is ever allowed to perform — the single irreversible step, gated behind deliberate human intent. Ten parallel threads of enthusiastic AI chewing through your backups is a wonderful thing, right up until one of them decides to “clean up.” Enthusiasm deserves a fence.

Proof: a working life, assembling itself in scratch space

I’m not arguing this from the whiteboard. I’m mid-salvage, and it’s the most fun I’ve had on a project in a while.

It started because I went looking for that NSF file. I didn’t find it where I expected — but on a backup server I found a stack of old Acronis disk images, and I pointed an agent at them with nothing more than a North Star: here’s what I’m trying to do, have at it. It built the tooling to look inside the images, found the NSF files, and then found something I’d forgotten entirely: a projects archive — all the documentation, everything I built in that company for twenty years. That’s probably worth more than the email.

Then it went to work, and the shape of the work is exactly the medallion pattern wearing its original clothes — a gold miner with a pan. The Acronis images are the ore. A Postgres database, where it’s deduping and cataloguing every file across machines and years, is the crush-and-wash. The wiki it’ll eventually compile is the smelt. Crucially, you don’t activate from raw — you collect, census, dedupe, let cheap deterministic passes concentrate the ore, and only then spend comprehension on what’s left. Where there was no driver for the dead format, it wrote one. Where the old Notes replicas were partial and differently dated, it used the identifiers the original software left behind to stitch them into a single record more complete than any copy that ever existed intact.

Two archives are coming back, and they’re complementary. The email is the as-operated layer — the negotiations, the decisions, the relationship texture. The projects file is the as-designed layer — what the company actually intended and built. Together they’re the full blueprint of my own former business. The legacy-organisation reading I’ve specced for clients is getting its first complete run on me. I’ve become my own first enterprise engagement.

The pendant recorders capture forward; this digs backward — the bronze layer of an entire working life, assembling itself in scratch space while I sleep.

None of this was possible eighteen months ago at any sane price. All of it is cheap now. And — this is the part that should stop you — if I’d deleted those Acronis images the way I deleted the NSF file, there would be nothing to salvage. No amount of cheap comprehension rebuilds territory you erased. (There’s a further chapter to this — turning a life’s archive into a living memory aid — but that’s a story for another day. Today’s rule is simpler: don’t delete the raw material for it.)

Do this now: the archive inventory

Before you next tidy a drive, clear a mailbox, or recycle an old machine, run an inventory instead of a purge. You don’t have to activate anything today — you just have to stop foreclosing the option. Point an AI agent (read-only) at your storage and ask it to find, not delete:

What comes back is a map of your own option value — the appreciating assets you were about to throw away. Keep the bronze. Build the map when you’re ready. But do it in that order, because the map can always be built later — and only ever over territory that still exists.

If you’re sitting on decades of organisational archives — old mailboxes, project drives, dead-format backups — and you’re not sure whether they’re a liability or an asset, that’s exactly the repricing this piece is about. I’m Scott Farrell; I help mid-market teams turn that kind of bronze into governed, activated capability. Before you delete anything, run the inventory. What would you find on your own drives?