Could the United States Seize Anthropic at a Discount?

Anthropic's IPO marketing kicks off this week. As is typical for investor disclosures in the age of AI, its publicly filed S-1 is full of warnings to investors, from ballooning computing costs to the possible extinction of humanity. Despite such a broad sweep, the S-1 may be missing a particular risk: that the United States government might try to seize control of the company, and that it might be able to do so at a steep discount to Anthropic's $1.4 trillion target IPO valuation.

The federal government (and its state and local counterparts) has broad power to seize private property: anything from land to copyrights and patents to corporate stock. The Fifth Amendment's Takings Clause requires "just compensation" for the seizure of private property, but doesn't otherwise limit *what* the government can seize.

How might this work? Imagine a company with a single class of common stock: 100 million shares outstanding at $2/share. A very simple model of a seizure is that the government would seize all of the company's common stock and pay shareholders $2/share. The government would then control the company, at a cost of $200 million.

So far, so simple. But things might get complicated, even in our simple example. The company's shareholders might try to go to court to dispute the price — they might argue that an acquisition would command a premium, for example. If the president attempted to unilaterally seize the company, the shareholders might argue that the president needed to get approval from Congress before doing so; that the emergency wasn't extreme enough to warrant executive action. But if a court decides the government appropriately authorized the seizure, it would probably take effect, with a compensation of more or less $200 million.

But public debt is spiraling, and maybe the government wants to pinch pennies. Instead of spending $200 million to buy all the stock, it could spend $102 million to seize 51% of the stock. As the majority shareholder, the government would effectively control the company: it could hire and fire board members and executives. There still might be lawsuits: here, the minority shareholders might argue that this was an "effective taking", especially if it destroys the value of their shares. That's not a sure thing: it's an easier case if the government shuts the company down, but less so if the government continues to run the business more or less as is.

Here's where it gets weird. In this toy example, the company had a single class of stock: control a majority, and you control the company. A shareholder's voting power in the company was the same as their economic interest in it. But many tech companies aren't organized this way. Instead, they have multiple classes of stock that give founders increased voting power. For example: Mark Zuckerberg owns about 13% of Meta stock, but controls 61% of its voting power. At Alphabet, while each class of stock has equivalent economic rights, its Class B stock for founders and other company leaders grants 10 votes per share, while its Class A stock grants 1 vote. Alphabet's Class C stock carries no voting rights, and trades at a slight discount. The newly for-profit OpenAI is controlled by a nonprofit that holds 26% of its stock but has the power to appoint all of the company’s board members. For tech founders who need to raise capital, this dual-class structure has been an appealing solution to attract outside investment without giving up control over the company's direction.

Anthropic takes this structure to an extreme. A long-term benefit trust will control a "Class T" share, which permits it to appoint a majority of the company's board of directors. And a "Founder LLC" will own a "Class F" share, which has 50.1% of voting power. Anthropic's S-1 states that neither class of share will have any claim to the economic value of the company—all or nearly all of the company's economic value will lie with other shareholders. But: controlling both the Class T and the Class F share will effectively control the company. As Anthropic says in its amended S-1: "Under this governance structure, we expect that the LTBT and the Founder LLC will collectively have the ability to elect all members of our board of directors, and the Founder LLC will have the ability to determine the outcome of all matters submitted to a vote of our stockholders generally for the foreseeable future."

So what would happen if the government tried to seize the company? Not the whole of Anthropic, but just the shares that control how the company is run. The reason is almost immaterial: say the company was 90% of the way to an especially dangerous or difficult-to-control technology; or produced a bioweapon in its wet lab; or a new hot war ripped away whatever restraint the government may have had. Or maybe the company gets on the government's bad side again, or the government decides it wants to control an AI lab.

What would the government be required to pay, in order to seize two shares with zero economic value? Here, the "just compensation" would likely not be zero—controlling Anthropic is worth something, after all. But it might be well south of $1.4 trillion; downright affordable, even. Economic shareholders may have a surprisingly weak case for compensation: Anthropic’s Public Benefit Corporation status allows directors and officers to take actions that may advance the company’s purpose even when it does not maximize benefits to shareholders. Anthropic’s S-1 explicitly warns that the company’s Trust may exercise its powers in a way that is adverse to other shareholders’ economic interests. Still, complications are likely: shareholders would sue; engineers could flee; the market would respond. Seizing two controlling shares would be a fairly novel legal approach that a court might view with skepticism, to say nothing of where it might land on the subject of just compensation. But in the right circumstances—or the right emergency—the government might take the risk.

This vulnerability is not unique to Anthropic. Other AI companies with imbalanced governance structures may be seizable for less than their full economic value. This is the cost of concentrating control in a small number of people, while working on a capable, potentially dangerous technology.

This may be unlikely to happen anytime soon. But if it does happen, the AI industry may have only itself to blame. After all, AI companies have spent enormous energy warning the public about the danger of their work, and how they cannot be trusted to wield this power alone. What if one day the government finally believed them?