Here is a company you have heard of. It was founded with a charter from the state, which gave it exclusive rights to a technology the state could not run itself. It raised money from the public before anyone understood what it did. It built infrastructure in other countries faster than those countries could regulate it. It negotiated with foreign rulers as if it were a government, because in practice it was one. Its board was a small group of men whose names the public could not list. And when a crisis came, the state found that the company was too big to let fail, too strange to nationalise cleanly, and too useful to leave alone.
I am describing the Vereenigde Oostindische Compagnie of 1602. I could equally be describing a company whose press release landed in my inbox this morning.
I come to this with a credential I rarely get to use. My master’s degree, from Ghent University, is in the socio-economic history of overseas expansion in the early modern period, which is the academic way of saying I spent two years reading about the VOC’s bookkeeping. The rule is that you have to say where the money comes from. Most people who write about AI labs would benefit from that rule.
So here is the argument, in six mechanisms and one caveat. The AI lab is not a new kind of institution. It is a very old kind, with the paperwork hidden.
Contents
- The charter: what the state grants, the state can take back
- The IPO: the public buys the thing before the thing exists
- The foreign policy: a company with embassies
- The sovereign-adjacent infrastructure: forts, ports and inference
- The Coen problem: what the company does when nobody is watching
- The nationalisation: how it ends, twice
- Where the analogy breaks
- Who revokes a charter
The charter: what the state grants, the state can take back
On 20 March 1602 the States General of the Dutch Republic merged a handful of competing trading ventures into one company and gave it a 21-year monopoly on all Dutch trade east of the Cape of Good Hope. The charter, the octrooi, also let the company wage war, sign treaties, build forts, mint coins and execute people. The state did this because it wanted the spice trade and had no navy to spare. The company did it because a monopoly is the best business model ever devised.
The AI charter is not a single document, and that is the only real difference. It is a bundle. There is the export control regime that decides which chips may leave the United States and which countries may buy them. There is the defence contract: in July 2025 the Pentagon’s Chief Digital and AI Office gave Anthropic, Google, OpenAI and xAI contracts worth up to $200 million each to prototype national-security capabilities. There is the compute allocation, which in practice means which company gets the next tranche of Nvidia hardware and which gets told to wait.
None of this is a monopoly on paper. All of it is a monopoly in practice, because a frontier model without chips is a very expensive PDF.
The VOC understood the arrangement perfectly: the charter was a lease, not a deed. Every renewal had to be negotiated, and the States General extracted money and favours each time. The AI labs are learning the same lesson in real time. In June 2026 the US Commerce Department ordered Anthropic to suspend all foreign-national access to its two most capable models, including for the company’s own employees. Anthropic called it a misunderstanding and switched the models off anyway, because that is what you do when the charter holder calls. The directive was lifted on 30 June, eighteen days later. The company survived. The lesson was delivered.
The Heren XVII would have recognised every beat of that story. They would also have asked why Anthropic did not send a gift.
“Octrooi van de VOC, 1602”, the first page of the charter, Nationaal Archief.
The IPO: the public buys the thing before the thing exists
In 1602 anyone with 3,000 guilders could buy into the VOC. Many did. The subscription raised roughly 6.4 million guilders from about 1,100 investors, and the shares could be sold on to other people, which is how a square in Amsterdam became the first stock exchange. The investors were buying a promise. The company had not yet sent a single ship under its new name. They were, in the modern phrase, pre-revenue.
The AI labs are not publicly listed, which is the one piece of 1602 they have declined to copy. They have done something stranger. They have built a public market that is private. In February 2026 Anthropic raised $30 billion at a $380 billion valuation. By June, after a $65 billion Series H, the valuation stood at $965 billion. OpenAI, over the same stretch, restructured into a public benefit corporation with its nonprofit foundation holding 26 percent and Microsoft holding 27. The money comes from sovereign wealth funds, pension funds through their venture allocations, the chip supplier itself, and a Japanese conglomerate whose name was last this prominent in 2000.
The VOC sold shares to the public and kept the books private. The AI labs sell shares to the private and publish the pitch deck. The outcome is the same: a large number of people who will never see the accounts have a financial interest in the company never being allowed to fail.
The Amsterdam exchange also invented, within a decade, the short sale, the futures contract and the first attempt by the state to ban both. I mention this because the people currently inventing the AI equivalent believe they are the first to do so.
The foreign policy: a company with embassies
By 1669 the VOC had 150 merchant ships, 40 warships, 50,000 employees and a private army of 10,000 soldiers. It had a capital city, Batavia, founded by storming Jayakarta on 30 May 1619. It ran its own foreign policy with the sultans of Java, the shoguns of Japan and the English next door, and it conducted that policy without asking The Hague, because The Hague was four months away by sea and had nothing useful to say.
An AI lab’s foreign policy is conducted in megawatts. In May 2025 OpenAI, Oracle, Nvidia, Cisco and SoftBank, with the Emirati state firm G42, announced Stargate UAE: a one-gigawatt campus in Abu Dhabi, the first phase of a planned five, with the first 200 megawatts due in 2026. The deal was brokered by the US president personally. The Commerce Department announced a working group to make sure the project met American security standards, which is the twenty-first-century form of sending a resident to the court of a foreign prince.
Read the structure, not the names. A company negotiates with a foreign state for land, energy and protection. A home government blesses the arrangement in exchange for the company carrying its flag. The company gets a factory abroad; the state gets a foothold it did not have to build. That is a VOC factorij with better cooling.
The VOC also learned that the home state could change its mind about which foreigners were acceptable. Nvidia learned it in August 2025, when it agreed to hand the US government 15 percent of its Chinese chip revenue in return for export licences. The 1602 charter had a word for a share of the proceeds paid to the sovereign for the right to trade. It was not “licence”. It was closer to “tribute”.
View of Batavia” (Andries Beeckman, c. 1662, Rijksmuseum), the market in front of the castle
The sovereign-adjacent infrastructure: forts, ports and inference
The VOC did not only trade. It built the things trade needed and then owned them: the fort at Galle, the harbour at Cape Town, the island of Deshima in Nagasaki Bay. Every one of them was a piece of state infrastructure on paper belonging to a company. When the company went, the infrastructure stayed, and whoever inherited it inherited a port they had not designed.
An AI lab’s forts are data centres and the power contracts that feed them. Those contracts increasingly involve the state as landlord, grid operator and customer at once. The Pentagon prototypes run on commercial models. Government departments buy inference the way they once bought paper. The labs, in turn, have started describing their work in the vocabulary of national capacity. When a private company starts calling its product “strategic”, it is asking to be treated like a port.
The historical pattern is that the port gets nationalised and the company gets a thank-you letter. In 1870 the Hudson’s Bay Company, chartered by Charles II exactly two centuries earlier, sold Rupert’s Land, 3.9 million square kilometres, to Canada for £300,000. It kept its trading posts and five percent of the land. The company lived on as a department store. The Bay closed its last stores in 2025, which is a sentence I did not expect to write in an essay about artificial intelligence, but history keeps its own schedule.
The Coen problem: what the company does when nobody is watching
In 1621 Jan Pieterszoon Coen, governor-general of the VOC, sailed to the Banda Islands with 19 ships, 1,655 Dutch soldiers and 286 Japanese mercenaries. The islands grew the world’s only nutmeg. The Bandanese had been selling some of it to the English, which the company’s charter said they could not. Coen had written the year before that it was “necessary to once again subjugate Banda, and populate it with other people.” He did. Of roughly 15,000 Bandanese, about a thousand survived. The company repopulated the islands with enslaved people under Dutch planters, and the nutmeg monopoly held for a century and a half.
I include this not as a parallel, because there is none, and I want to be clear about that. No AI company has done anything of the kind, and nothing in this essay should be read as suggesting it. I include it because of what Coen’s reasoning looked like from Amsterdam. It looked like a quarterly report. The Heren XVII received dispatches, approved budgets and paid dividends. The thing itself happened eleven thousand kilometres away, in the gap between the charter’s words and the company’s incentives.
A charter tells a company what it may do. It says nothing about what it will do when the alternative is losing the monopoly. The AI labs’ equivalent of the Banda Islands is not a place. It is every decision taken under competitive pressure that nobody in the home state will ever read a dispatch about: which safeguard gets relaxed, which customer gets served, which jailbreak gets described as “narrow”. The Commerce Department’s June directive suggests that at least one government has started reading the dispatches.
The nationalisation: how it ends, twice
The VOC did not die in a battle. It died of its own accounts. The Fourth Anglo-Dutch War cost it up to 43 million guilders in damages, loans reduced its net assets to zero, and in 1796 the board was simply sacked. On 31 December 1799 the charter was allowed to expire. The Batavian Republic took the possessions and the debt. The initials, people joked, now stood for Vergaan Onder Corruptie: perished under corruption. The joke was not entirely fair. The company had also perished under a cost structure built for a world that had stopped existing.
The English East India Company, chartered by Elizabeth I on 31 December 1600, lasted longer and ended louder. By the early nineteenth century its three presidency armies totalled about 260,000 soldiers, twice the size of the British Army. The 1857 rebellion exposed what everyone had suspected, which was that a trading company was running a subcontinent and should probably stop. The Government of India Act received royal assent on 2 August 1858; the Company’s territories, treaties and armies passed to the Crown; a Secretary of State for India took over the directors’ desk. The Company itself lingered as a shell until 1 June 1874, when it was dissolved after a final dividend. Even in death, it paid out.
Two endings, one pattern. The state does not revoke a charter when the company does wrong. It revokes it when the company becomes a liability the state cannot afford to be seen holding. The Dutch waited until the balance sheet was gone. The British waited until the army was.
The AI labs are watching OpenAI’s restructuring for a hint of which ending they get. A nonprofit foundation with 26 percent and two state attorneys general who “have no objection” is a very modern way of saying that the charter now has a clause about the public interest and nobody has yet tested what it means.
Apotheosis of the Dutch East India Company (Allegory of the Amsterdam Chamber of Commerce of the VOC) (1702 – 1746) – Nicolaas Verkolje (Dutch, 1673 – 1746)
Where the analogy breaks
I owe the reader the counterargument, and it is a good one.
The VOC had an army. AI labs have a legal department, which is worse for them and better for everyone else. Coen could storm a city; the most a model provider can do is refuse a request, and the request can go to a competitor in the time it takes to open a new tab.
The VOC had a monopoly on a sea route, which is a physical thing you can blockade. There is no Cape of Good Hope in machine learning. The weights leak, the papers are public, the Chinese labs publish on arXiv like everyone else, and the chips, for all the export controls, are a supply chain rather than a strait. You cannot build a fort on a technique.
The VOC’s shareholders were a thousand Dutch merchants. The AI labs’ shareholders are, at one remove, everyone with a pension, because the venture funds and sovereign funds that hold the stock are holding it on behalf of teachers in Ontario and nurses in Oslo. That changes what “too big to fail” means. When the VOC failed, Amsterdam lost money. When an AI lab fails, the bill goes to a far larger table, and the people at that table were never told they were sitting at it.
And there is the obvious one. The VOC traded in a product everyone understood. Nobody in 1602 needed nutmeg explained. The AI labs trade in a product whose value, risks and even definition are contested by the people building it. A charter for cloves is a charter for cloves. A charter for intelligence is a charter for whatever the word turns out to mean, and the state that grants it is granting something it cannot describe.
So the analogy breaks. It breaks exactly where you would expect a four-hundred-year-old analogy to break, which is on the questions of force, geography and scale. It holds on the questions of structure, incentive and ending. Those are the ones that matter for the next ten years.
Who revokes a charter
I want to end on the question the VOC never managed to answer, and the question the AI labs are being asked whether they like it or not.
A charter is a grant from a sovereign. The sovereign can take it back. In 1799 that meant the States General, in 1858 it meant Parliament, in 2026 it means a department in Washington and, as the June episode showed, a department can do it in an afternoon with a phone call and a memo. That is one answer, and it is the answer the labs are organising their lives around: be useful to the state, be careful with the state, do not surprise the state.
But it is not the only sovereign in the room. The VOC’s charter was granted by the Republic; its forts stood on land the Republic had never seen and its subjects were people who had never heard of The Hague. The AI labs’ charters are American; their users are not. The models are trained on text written by everyone. The thing that gives the charter its value was never the state’s to grant in the first place, and that is the part of the analogy nobody wants to finish.
I do not know who revokes that charter. I suspect it is not revoked at all, but allowed to expire, quietly, on some 31 December when the costs of renewing it exceed the costs of letting it go. That is how the VOC ended. It is how most charters end. The company does not fall. The state simply stops signing.
The Heren XVII met for the last time in 1796 and went home. There is no record of what they said. I like to think one of them looked at the ledger, looked at the map, and remarked that it had all seemed like a very good idea at the time.
Then he asked whether there would be a final dividend. There was.




