Skip to content
zensation

Research overview

Three tracks, architecture and agenda

Publications

Preprints, software, identifiers

Methodology

Operational standards and validation

Research ethics

Fundamental rights and compliance

Resources

Code, data, citation, open science

Adopt AI

Neutral roadmap & interactive simulation

Public sector & funding

Collaborations in the public sector

Protecting public spaces

Track B — rights-preserving early warning

Technology

The architecture behind every application

Playground

Run the open memory core live

Work with usAboutOpen SourceBlog
Contact
zensation
Research overviewPublicationsMethodologyResearch ethicsResources
Adopt AIPublic sector & fundingProtecting public spacesTechnologyPlayground
Work with usAboutOpen SourceBlogContact
Blog
Analysis

The Data Is Right There. Nobody Can Reach It.

Alexander Bering
Alexander Bering
August 18, 2026 · 6 min read

In March 2026, a contract ended for a public broadcaster in St. Louis. Nothing dramatic: a storage service agreement, renewed annually since 2019. This time the renewal did not happen, because from February onwards the contracting partner stopped responding. On 6 March, the day the contract expired, access was gone. Not throttled, not more expensive. Gone.

What has happened since is the interesting part. The broadcaster sued the service provider in April. It changed nothing — by then the company had effectively ceased to exist. The data itself does not sit there at all, but in a data centre run by a large custodian in Colorado. And the custodian will not hand it over, because its customer was not the broadcaster but the vanished provider. That provider owned the infrastructure the data sits on.

So in July the broadcaster sued a second time, now against the custodian. A court has since barred anyone from deleting, altering or overwriting the data, and confirmed that it belongs to the broadcaster.

Note what did not go wrong here. No drive failed. No data centre burned down. Nobody was breached. There was no data loss — as far as anyone knows, the files are intact. Ownership is undisputed and court-confirmed. And still, for five months, nobody has been able to reach 70 years of regional history.

The chain, not the medium

The case is being told everywhere as a cloud warning: don't trust the cloud, keep backups. That misses the point, and it misses where the thing actually broke.

What broke is the chain of custody. The broadcaster had a contract with A. A had a contract with B. The data sat with B. When A disappeared, the broadcaster was left holding a claim against someone who no longer exists — and no contractual relationship at all with the party physically holding the data. B is not even behaving improperly here. B is doing what a custodian must do: it does not release someone else's data to a third party merely because that party says it owns it.

That is why a backup does not help here. A backup protects against loss. Nothing is lost. It does not protect against the rupture of the relationship through which you reach the holdings.

And it is why portability does not help either. Exportability is a property you exercise while you still have access. It presupposes exactly what is missing. The broadcaster was free to export at any time — right up to the day it no longer could, and that day arrived without warning.

Why this cuts deeper for the AI layer, not shallower

We wrote here a few weeks ago about a different mechanism: what remains when the model disappears. That was about losing a capability — a model is switched off, a vendor changes the rules, and what your AI could do, it can no longer do.

This case is the other mechanism, and it is the more uncomfortable one, because it requires no intent at all. Nothing disappears here. The chain simply grows longer, and your holdings sit at the end of it.

In a typical AI setup that chain is considerably longer than for a video archive. Between you and the state your systems have accumulated over years there is often a model provider, a memory or agent service, and a vector or database host — not rarely across three jurisdictions, with contracts that exist only pairwise. Each link is individually replaceable. The chain as a whole is not, and nobody holds a contract with it.

There is one further difference that makes the comparison genuinely uncomfortable. In St. Louis the subject is video files. If the broadcaster wins release, it has its archive back — a video format stays readable without the provider. The state of an AI knowledge layer is not like that. Embeddings without the model that produced them are columns of numbers. An index without its schema definition is a file. A memory graph without the semantics of the system that wrote it is a pile of edges. In that case you can win the lawsuit, receive every byte — and still be left holding something you cannot read.

The holdings are not gone, then. They have merely stopped being knowledge and become data waste with a title deed.

Computing locally is not remembering locally

There is a good countervailing development right now: models are becoming locally runnable. Several capable open models have appeared in recent weeks, explicitly built for always-on local agents and working on ordinary hardware. That is real progress, and it genuinely solves part of the problem.

Just the smaller part. Read the announcements and you find quantisation, function calls, context windows. About the state between sessions they say nothing. Bringing the model onto your own machine while leaving the memory with the vendor answers the sovereignty question by half — and by the less important half. The model is replaceable. The state is not.

The question that follows

The useful question after this case is not whether your vendor is trustworthy. The one in St. Louis presumably was, for seven years. The question is:

If a link in your chain fails tomorrow — not maliciously, but insolvent, acquired, or simply unreachable — at what point do you still have access to your own state, and in what form?

Anyone who can answer that has built it right anyway. Anyone who cannot is carrying a risk that appears in no backup plan, because it is not a loss risk.

In practice this means two things. First: as few links as possible, and a direct claim against whoever actually holds the material — not against an intermediary. Second, and this is the part easily overlooked: the state must exist in a form that remains readable without the system that wrote it. Otherwise you will one day win a case and be left holding bytes.

That is why we build this layer to run on your own infrastructure with its format out in the open. Not because your own servers are inherently safer — the data in Colorado is perfectly safe. But because a chain you hold yourself has no links that can disappear.


The case described is publicly documented and was covered in mid-August by Current, Gizmodo, The Desk and Tom's Hardware. Every fact stated here is supported by at least two of those sources; the proceedings were pending at the time of writing. We take the case as a mechanism, not as a verdict on those involved — its instructive value lies precisely in the fact that nobody need have acted wrongly for it to happen.

Share on XShare on LinkedIn

Related Articles

The Narrow Band

Why a working system sits on a narrow edge — and why, with AI, small errors don't add up, they compound. Precision isn't optional.

You Can't Host Trust

"Hosted in the EU" is reassuring, "open weights" sounds good. Neither answers the real question: do you actually know what your AI does with your data? You can only check if the layer that touches your data is open.

What Remains When the Model Disappears?

In June, a leading model was switched off three days after launch. Models come and go. What remains is the knowledge layer above them. The real question is who owns it.

zensationIndependent AI research lab · Kiel
ResearchPublicationsMethodologyResearch ethicsResourcesAdopt AIPublic sectorProtecting public spacesTechnologySystem explorerPlaygroundWork with usOpen SourceAboutBlogChangelog
GitHubLinkedInarXivZenodoORCIDScholarSemantic ScholarHuggingFacenpmDiscussions

© 2026 Alexander Bering / ZenSation Enterprise Solutions

RSSPrivacy PolicyLegal NoticeAccessibility