As autumn was drawing to a close, an artificial intelligence company that builds large models handed a very thick document to the United States securities regulator. People outside the circle rarely read such material, but this one is worth a few extra words, because what sits inside it is not a new model. It is a set of books.
A Document Lays Out the Books
That document is a prospectus: the long bill of particulars a company lays out for everyone to see before it offers shares to the public. Whether the sale goes through is another matter; being willing to write down revenue, costs, customers and commitments item by item is itself a threshold.
In the past, talk about companies like this ran on launch events, interviews and hearsay: a lot of listening, very little looking. This time, for the first time, there is a public set of accounts that can be read line by line, where the money comes from and where it goes set down in black and white rather than left to whoever tells the best story.
What is interesting is that the document does not say how many shares will be sold or at what price; even the price range is left blank, to be filled in later. What it can answer right now is how the company earns and spends today, not what price tag it will be given on the day it lists.
Keeping those two things apart makes the whole thing easier to read without drifting. The excitement tends to gather around the column that has not been written yet, while the part that actually teaches you something is the pages already filled in. They ask for no imagination, only numbers and notes that line up row by row.
Half-Year Revenue, in Black and White
The line that catches the eye first is revenue: in the first half of this year, the company recorded about US$16.3 billion. For a company only a few years old, that is unusual, and several outlets checked it against the filing with broadly consistent definitions. It is not a figure someone tossed out casually.
A more intuitive comparison is to set it beside the company's own revenue for the whole of last year. What half a year produced is already several times what a full year produced before; what lies between them is not a little growth but a business that has changed scale altogether.
As for the growth rate, the definition has to be settled first. The filing discloses the first-half figure compared with the same period a year earlier, while the percentages quoted in the press differ, some calculated over the half year and some over a single quarter. Different methods, and the numbers naturally do not match.
Rather than memorize a percentage that is easy to mix up, it is better to keep the steadier line: compared with a year ago, this is no longer a business of the same size. How fast it grew is one thing; growing fast enough that people have to specify which stretch of days they mean is itself telling.
That revenue line also prompts a plain question: who paid it. The answer sits mostly with companies, using the product to write software and handle the miscellaneous work of an office, rather than with ordinary people signing up on a whim because something looked new.
Who Actually Supplied the Valuation
As for what the company will be worth after listing, that is the column most easily misread, and it calls for particular care. The prospectus itself gives no price range; the number of shares and the price are both still blank, so there is no figure the company has set for itself.
The range circulating outside comes from estimates among investors and from reports quoting people familiar with the matter, landing roughly between US$1.5 trillion and US$2 trillion. These are market discussions, not a number given by the company, and they are still shifting with time.
A week or so can change the story: the figure the same people mentioned a little earlier and the one they mention today need not be the same one. Freezing it into a settled conclusion turns a live discussion into a price tag that appears already signed.
Mistaking discussion for conclusion also skips the several steps in between. The filing is not yet effective, the size and price of the offering come later, and the valuation only settles at the moment of actual pricing. Until then, every number is somebody else doing the arithmetic on its behalf.
So for an ordinary reader, the useful thing to take from this column is not the range but where it came from: who is estimating, on what basis, and whether the company has acknowledged it. Getting those three straight is more use than memorizing one very large number.
The Other Side: The Compute Bill
If the revenue line looks good, the other side should not be skipped either: holding up that revenue is a compute bill just as striking. Training models and answering requests both depend on banks of machines running day and night, and none of the electricity, the equipment or the floor space comes free.
The filing also lists long-term compute commitments stretching many years out. The money is not paid in one go; it is spread across years to come, like signing a lease that runs a very long time, not needed yet but already recorded on the books as an obligation.
The loss line needs taking apart too. A large part of it comes from changes in the fair value of convertible notes and similar instruments as the company's own valuation rose. That is an accounting measure, not cash actually spent over the same period, and the two should not be added together as though they were the same.
Only when these lines are read together does the picture come out whole: revenue climbing quickly on one side, compute promises pressing years into the future on the other, and between them a loss that is mostly a matter of accounting. Pick out any single line and you will read the books crooked.
In the same filing, there is one quarter in which the company recorded a profit at the operating level for the first time. It is an unshowy detail, but it pushes the idea of simply burning money back a step: in that stretch at least, the services sold covered the operating costs of the day.
From a Story to a Spreadsheet
The customer column deserves a glance as well. Paying corporate customers widened considerably over the year, moving from an earlier small circle into a much broader set of industries, and existing customers spent more than they had the year before, suggesting that once people got used to it they were reluctant to switch.
Writing software, handling documents, running a chain of steps for someone: that is where most of this money goes. What companies keep paying for is probably not how stunning a demo looked, but whether the thing can sit inside their own work day after day.
Set revenue, compute, losses and customers on a single page, and for the first time the company is no longer just a story about the future. A story can reach a long way; a set of accounts will only recognize what has already happened, and here the two are being weighed together for the first time.
For an ordinary reader, that is a useful nudge: when you look at an AI company, beyond the lively launch there is a set of accounts that can be checked line by line. The launch makes you look up; the accounts give you somewhere to land. Without the second, judgment is left hanging.
Stories reach far, accounts stay close. Where a company has actually got to can now be checked on this page before you listen to the adjectives from the stage; and the blanks that have not yet been written into the filing might as well be left blank until they are filled in.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.