A Rather Expensive Way to Summarise a PDF

There is a certain comfort in large numbers, chiefly because they cease to mean anything after the twelfth zero. When one is told that Anthropic, a company whose primary contribution to human civilization is a remarkably polite piece of software named Claude, is preparing for an initial public offering that could value it at two trillion dollars, the mind does not panic. Instead, it gently shuts down, much like an overworked laptop trying to process a particularly dense spreadsheet. Two trillion dollars is, by any sensible metric, an absurd amount of money. It is enough to purchase several small European nations, or, perhaps more realistically, to pay for the electricity required to keep the world's chatbots running for another fortnight.

To put this figure in perspective, two trillion dollars is roughly equivalent to the valuation of SpaceX, a company that spends its time launching actual, physical rockets into the actual, physical heavens. SpaceX deals in fire, steel, and the terrifying physics of escaping gravity. Anthropic, on the other hand, deals in the delicate art of text prediction. Its primary achievement is a digital assistant that, when asked to write a sonnet about a toaster, does so with an almost painful level of sincerity, before politely asking if there is anything else it can do to make your afternoon slightly more tolerable. It is a fascinating contrast: on one side of the ledger, we have the conquest of the cosmos; on the other, we have a very expensive way to summarise a PDF.

The investment bankers orchestrating this financial spectacle are, of course, entirely unbothered by the lack of physical output. In the modern economy, physical assets are increasingly viewed as a bit of a nuisance. They require maintenance, they are subject to the laws of friction, and, worst of all, they occupy actual space. A digital algorithm, by contrast, exists in a state of pure, unblemished potential. It does not need to be painted, it does not rust, and it never demands a pension. It merely requires a continuous supply of electricity, which is currently being harvested by melting salt at five hundred degrees Celsius in the American desert—a detail that feels entirely reasonable to anyone who has ever tried to explain the modern tech sector to their grandmother.

One cannot help but admire the sheer optimism of the valuation. A two-trillion-dollar price tag suggests that the market believes Anthropic has not merely built a very clever autocomplete tool, but has instead discovered the secret to manufacturing thought itself. There is a quiet, almost religious faith among investors that if you feed enough data into a sufficiently large cluster of graphics cards, eventually something resembling wisdom will emerge from the other side. It is the modern equivalent of the infinite monkey theorem, though with significantly more venture capital and slightly fewer bananas.

I recently spent an afternoon watching a colleague attempt to use one of these multi-trillion-dollar algorithms to draft a polite email to our landlord regarding a leaking pipe. The algorithm, displaying the cognitive depth of a seasoned diplomat, produced a three-hundred-word masterpiece that managed to express deep empathy for the landlord's busy schedule while gently suggesting that water should not, ideally, be cascading through the light fixtures. The landlord, being a biological entity of a more traditional design, replied with a single, unpunctuated sentence: "will look on monday." It was a beautiful demonstration of the modern technological divide: a two-trillion-dollar brain being utterly defeated by a man named Arthur who owns a ladder.

This brings us to the central paradox of the artificial intelligence boom. We are spending sums of money that would make the pharaohs blush to automate tasks that humans generally perform while slightly distracted. We have built a cathedral of silicon and cooling fans, all so that we can generate slightly more persuasive marketing copy for artisanal soap. The bankers assure us that this is merely the beginning, and that soon these algorithms will be managing our pensions, diagnosing our illnesses, and perhaps even deciding who gets to be the prime minister. One hopes they will do so with the same polite deference they currently bring to the toaster sonnets, though there is always the slight worry that a minor formatting error in a database could result in the accidental annexation of Belgium.

There is a quiet library in Somerset where the loudest sound is the occasional turning of a page. I often think of it when reading about the latest developments in Palo Alto. In that library, a century of human thought sits quietly on wooden shelves, requiring no electricity, no cooling systems, and certainly no investment bankers. It is a remarkably efficient arrangement. If you wish to know what Seneca thought about anger, you simply open a book. Seneca does not charge you by the token, he does not hallucinate facts about the Roman senate, and he never pauses to tell you that his servers are currently experiencing high demand.

Yet, the march of progress is relentless, and we must all prepare for the day when our sentences are fully unionised and our spreadsheets are capable of feeling a mild sense of existential dread. When Anthropic finally makes its debut on the stock exchange, there will be the usual flurry of ringing bells, expensive champagne, and breathless television commentary. The founders will become billionaires, the bankers will buy slightly larger yachts, and the rest of us will return to our desks, where we will ask Claude to summarise a document we have no intention of reading, so that we can send an email we had no desire to write, to a colleague who will use another chatbot to draft the reply. It is a magnificent, self-sustaining ecosystem of digital noise, and at two trillion dollars, it is practically a bargain.