Why the installation costs more than the computer

Two purchases that look like one — and the labor that does not get cheaper as models improve.

By VK Singh Vashisht — Datura Labs

A desktop AI box now costs about $4,700. A subscription on top of it runs a capable local model over your files, on your hardware, with nothing leaving the machine. Anyone evaluating a TalkingDB installation should be asking why ours begins at $25,000.

The answer is that those two things are not the same purchase, and we would rather say so in public than one call at a time.

What the box buys

The box buys a reader. It assumes the corpus already exists — that your documents are files, in folders, on that machine, in formats a model can parse. Given that assumption, it is very good, and it is getting better on a clock we do not control and would not bet against.

If your records look like that, buy the box. We will tell you so on the call, and it will be a short call.

What the engagement buys

Most archives do not look like that. The ones we are built for look like this:

Twenty-five years across four machines, two of them dead.
Scanned documents where roughly one page in five came off the glass rotated, producing text that reads as noise to every search you have ever run against it.
Three cloud accounts, one of which belongs to a company that no longer exists.
Mail in formats that predate the tools you would use to read it.
Material that must never be indexed at all, mixed into the same directories as material you need answered tomorrow.

Getting that into a state where a reader helps is the engagement. Ingest, extraction, OCR correction, deduplication, classification, and a wall between what may be retrieved and what may not. None of it is inference. All of it is labor, and it is the part that does not get cheaper as models improve.

Three commitments the hardware does not make

Grounding and abstention. A system that answers from your records must also decline to. “Not in your records” is a correct answer, and a system that will not produce it is not usable for anything consequential. We treat the abstention rate as a measured property of an installation, not a claim about it.

Egress you can audit. Not a permission dialog. A deterministic check with an exit code, a default that denies, and an append-only log your counsel can inspect after the fact. Policy enforced by withholding capability, not by instructing a model.

A name on the agreement. The installation runs under a counsel-reviewed agreement with a named principal accountable for it. For a family office carrying live litigation, that is not a soft benefit.

What we do not have yet

No published sample deliverable. The only corpus this has run against at full scale is our own, and no anonymized artifact has cleared review. Until one does, TalkingDB is listed as inquiry rather than as a storefront — the same gate that governs every queued marketplace on this site.

We would rather be a listing you can evaluate honestly than a storefront you cannot.

The honest version of the comparison

The local-first argument has been won. It was won by people with more capital and better silicon than we have, and we think that is good. What their work does not address is the archive itself: getting it in, keeping the wrong parts out, and standing behind the answers.

That is the whole of what we sell. If you do not need it, you should not buy it.

Inquiries go to the principal. There is no mailing list and no follow-up sequence.

Datura Labs builds agents and infrastructure directly on frontier models. This is the work.

The installation described here: TalkingDB.