On that Monday morning you were headed to your doctor checkup you have booked a week ago. The physician pulled up your file - scans, labs, medication history, the records accumulated of years at the same hospital and half of it didn’t load even though the images exist. They were taken on the hospital’s own scanners the week before. But the hospital decided to rent its data storage from a vendor rather than build an IT department of its own (the quality of the imaging requires more space). So, the platform that stores and serves those records back to clinicians belongs to a private vendor under contract with the hospital network and that contract is currently in dispute. There is a disagreement over renewal pricing that has, as of that morning, resulted in the contractor controlling access pending payment. The scan hadn’t gone anywhere, it is on the cloud but nobody can see it.
The hospital hasn’t lost a single patient record just it has lost the ability to read them.
Usually, this would be considered as a technology problem but this time it is not the case. The reason has nothing to do with firewalls or login screens. The reason lies in a very old story wearing new clothes - somebody built the pipeline, somebody else came to depend on it, and the person who paid for it to be built discovered, at a convenient moment, that dependence is a form of power which he can use, yet without restrictions.
Thus, this is not really a piece about data or about AI, however often those two words end up sharing a sentence lately. It is about the older and less exciting question of governance and system of values: who gets to decide what happens to a thing everyone depends on, and who handed them that authority in the first place.
Dependence is a form of power.
Your data but not your terms
The obvious next question “what is the policy?“ has an unsatisfying answer: not enough or not fast enough. A policy is a promise and promises get broken. What you would actually want is not a promise but a rule built into the infrastructure itself, so the data only moves on the terms its owner set, mechanically, whether anyone downstream feels like respecting the spirit of the thing that day.
Europe is, at least, trying to build exactly that kind of rule: the EU Data Act’s provisions on cloud switching, in force since September 2025, require a vendor to let a customer move its data to another provider or back onto its own servers - within a matter of months, and ban the fees for doing so outright from 2027 onward. The point is not that a vendor can never sit on a harsh invoice but that a vendor is no longer allowed to make leaving slow or expensive, which was most of what turned a billing dispute into leverage in the first place. Although it sounds like a good rule, it only works, where there is an institution behind it old enough and well-funded enough to move fast, and that is not most places.
The researcher Kate Crawford argues in her book Atlas of AI, that data is still widely treated as something simply lying around for the taking, collected without much restriction, used without much context, by industries with every incentive to keep it that way rather than see it questioned or regulated. Pete Howson, who studies cryptocurrency and international development, has a less diplomatic term for one version of it: “crypto-colonialism.” Projects tried out first on people too poor, or too legally unprotected, to object.
Project Worldcoin illustrates this pattern well. Tools for Humanity, the company behind it, co-founded by OpenAI’s Sam Altman, sent metallic orb-shaped scanners around two dozen countries and, by early 2022, had scanned the irises of nearly half a million people - most of them in the developing world, for about $50 in cryptocurrency a head, in exchange for what the company called Proof of Personhood. MIT Technology Review’s investigation that spring, built on more than thirty-five interviews across six countries, turned up recruitment events staged to look like government aid handouts and a consent form that admitted, in writing, that Worldcoin didn’t yet meet European privacy standards. The investigation also documented local operators in Indonesia paying village officials for access to crowds; the operators involved are independent contractors rather than company employees, and Worldcoin’s own response called the incidents isolated and not sanctioned company practice.
Kenya’s experience in this project became the most visible one. Tools for Humanity started piloting there in May 2021, two years after the country’s first data protection law which it turned out, had no regulator behind it fast enough to catch a foreign biometrics project moving at that speed. The pilot ran two full years before anyone in government stepped in. By August 2023, when the queues outside a Nairobi convention center made international headlines - about six weeks into a rollout already underway across dozens of Kenyan cities where the company had collected iris data from hundreds of thousands of Kenyans and shipped it to servers in Germany without the legally required consent or impact assessment, the government suspended the project within days. It still took until 2025 for a court to order the data deleted, and until early 2026 for regulators to confirm the erasure. It took five years between the first scan and the confirmed deletion showing that the law got there eventually, yet it wasn’t built to move at the speed of the thing it was supposed to stop.
The point is that a rule wired into the pipeline only solves half the problem, because there is always a stretch of time where the law hasn’t caught up yet. What happens in that stretch depends on something else entirely: whether the company itself has anyone in the room who answers to a mission, or just whoever happens to hold the most shares.
One coin, two sides
Every company, when you get down to it, is answering two separate questions that most people assume come attached together: who gets paid and who gets to decide. In an ordinary company they are tacked together - whoever owns the shares owns both the profits and the steering wheel, which is exactly why a majority shareholder having a bad year (everyone does sometimes!), a change of heart, or simply a tempting enough offer can quietly turn a company’s mission into whatever the new owner wants it to be next.
Steward-ownership is type of governance which solves the trust problem differently than most corporate governance does. Instead of hoping managers will stay true to the mission or law to catch it in time if they don’t, it takes the temptation off the table entirely - permanently splitting who controls the company from who profits off it. Steward-ownership is a kind of governance where the voting control is designed to stay with people accountable to the company’s purpose, usually through a trust or foundation built specifically so those votes can’t be traded away like ordinary shares. The money, meanwhile, keeps flowing to investors, founders, and employees exactly as it would anywhere else. Nobody works for free and nobody’s capital and time goes unrewarded; what is different is that nobody gets to simply cash out the mission along with their shares. For a founder, it means building something that can outlive your own patience for running it. For the investor, it means something less romantic and, in its way, more persuasive such as durability. If you are an employee, a customer, or a hospital sharing patients’ scans with a vendor, this is what protects you: nobody can buy the company overnight and hand it to a new owner who never heard of the mission or doesn’t care about it.
Nobody gets to simply cash out the mission along with their shares
The company that has run this type of governance the longest is Robert Bosch GmbH, and it has now been running it for over sixty years. When Robert Bosch died in 1942 without an obvious successor, his will eventually split the business in two: a charitable foundation, the Robert Bosch Stiftung, holding about 94 percent of the shares and every euro of dividend that comes with them, and a separate industrial trust holding roughly 93 percent of the voting rights and almost none of the capital. One side gets rich and gives the money to health and education causes while the other runs the company and neither can do the other’s job. Six decades on, Bosch still exists, still invests billions into research every year, and has never once had to keep away an activist shareholder demanding it strip itself for parts.
A much smaller, quieter proof of concept is Ecosia, the Berlin-based search engine that donates its ad revenue to tree-planting. In 2018, its founder Christian Kroll, gave up his own right to ever sell the company. He found a nonprofit called the Purpose Foundation which holds a single golden share - about one percent of the voting rights, no economic stake whatsoever - whose only power is to veto a future sale of Ecosia, or any change to its mission. Turns out you don’t need a century or a family fortune or an AI-safety board to pull this off. It still quietly works exactly as designed. You need just one clause and a founder willing to close off his own exit.
Anthropic is the newer, considerably less tested version of the same bet, aimed at a newer and much faster-moving problem. In 2023 it created the Long-Term Benefit Trust: five trustees, picked for expertise in AI safety, national security, and public policy, holding no shares, no dividends, no upside if the company’s value climbs. What they hold is a special class of stock whose only power is electoral - the right to appoint a growing share of Anthropic’s board, eventually a majority. Yet, Anthropic still has ordinary investors chasing ordinary returns; the Trust’s job is to make sure that when the gap between “profitable” and “responsible” is deeper, the people holding the board seats answer to the mission rather than the share price. Anthropic itself calls the arrangement an experiment rather than a template.
Where this gets difficult
What is difficult in the time of agentic economy is the truth that building a decentralized dataspace is slower and more painful than pointing everyone at one big platform, because it requires many independent companies to agree on shared standards before most of them see much benefit. And the hardest truth is: the governance structure only protects a mission if the people inside it are worth trusting with the veto - and worth trusting isn’t something an agreement can certify on its own.
Anthropic’s own Long-Term Benefit Trust, already mentioned, has drawn exactly this challenge from an AI-governance researcher Zach Stein-Perlman who runs AI Lab Watch, whose widely-circulated analysis argues that the Trust may be considerably weaker than it appears: stockholders, he reports, retain the power to enforce or override it through supermajority votes whose exact thresholds Anthropic has declined to publish, and he suggests that reluctance itself is telling. Someone still has to guard the guardians, and no legal document yet invented does that job unassisted.
Back to Monday morning
Picture the same Monday again, same contract dispute, same difficult morning but this time the infrastructure was built differently from the start. The hospital keeps its records with whoever hosts them. What is different is that it was never locked into just one vendor with no way out, and no single company was ever handed the only door its data had to pass through.
Two things made it work. One is right: the EU Data Act’s cloud-switching provisions mean the hospital could always walk to another vendor within months, without paying a penalty to leave so an unpaid invoice was never going to become a hostage situation in the first place. The other is structure: a shared standard, governed by an entity nobody could quietly buy and repoint, that lets the hospital set the terms under which a record moves onward to (a specialist across the border, to the patient’s own phone) and take those terms back.
In the agentic economy, the word ethics is becoming heard so often that people stop asking what is actually being done about it. The fix here has nothing to do with curing greed because people will keep fighting over invoices for as long as there are invoices. The change on offer is smaller: eg. when the physician opens the file, it is going to be there. Not that anyone will stop disagreeing just that disagreement no longer comes with a switch that turns off somebody else's lights.
The author researched, fact-checked, and edited this piece, using AI tools for structural drafting and grammar checks along the way. All facts, figures, and quotes were independently researched and verified by the author.
The Infrastructure of the Commons is a biweekly publication on data sovereignty, agentic systems, AI governance, and the critical infrastructure decisions that will define the next decade.
