The First Trillion

The first organization to report one trillion dollars of revenue in a single year will be OpenAI. It will get there by selling intelligence to one billion customers at an average order value of $100 a month, and the line item on the receipt will say usage credits. That is $1.2 trillion a year, which clears the mark with room to spare. I expect the first trillion-dollar year to be reported for 2030, and I will be specific below about what has to happen for the date to hold and what would make me wrong about the name.

• The arithmetic1,000,000,000 × $100 × 12$1.2 trillion / yrA billion customers is fewer people than already open ChatGPT in a given week.
• The threshold$1,000,000,000,000 ÷ $1,200833 millionThe number of $100-a-month customers it actually takes to cross a trillion. The billion is margin.

I said in the first piece that a billion people will pay $100 a month for ambient intelligence by 2030, and I left the payee open between two companies. This piece closes it. The bill goes to OpenAI first, and the reason is not that OpenAI has the better model. It is that OpenAI has the billion people, and it has already built the meter.

Two clarifications before the argument, because they are where forecasts like this usually cheat. First, I mean reported revenue for a full year, not a run-rate. Run-rates flatter: OpenAI ended 2025 with more than $20 billion of annualized revenue by its CFO’s account, and its audited revenue for the whole of 2025 was $13.07 billion. A trillion-dollar run-rate will arrive a year or so before a trillion-dollar year. Second, I mean gross revenue billed to customers, which is the number Amazon and Walmart report, and the number the comparison has to be made against.

Go to Settings, then Usage, in ChatGPT. There is a balance there, and a button to add to it. OpenAI’s own help page describes the thing plainly: credits are a pay-as-you-go add-on that kicks in after your plan’s included usage is spent. They are prepaid. They expire twelve months after purchase. They are non-refundable except where the law requires. And if automatic reload is available on your account, you set a minimum balance and a target balance, and when the meter drops below the minimum, OpenAI charges your card for the difference.

Read that paragraph again as a description of a business rather than a feature. Prepaid, expiring, non-refundable, auto-replenishing. That is not a subscription. It is a top-up card for a utility, and it is the same instrument that carriers used to take billions of people from a phone they owned to a bill they pay.

• PrepaidMoney in before intelligence out. Cash arrives ahead of cost, which is the opposite of how the company has funded itself so far.
• Expires in 12 monthsUnused balance is revenue anyway. Breakage is a line item every gift-card issuer already knows how to book.
• Auto-reloadA minimum, a target, a card on file. The customer decides once, and the purchases happen without a decision ever again.
• Spans featuresOne balance feeds Codex, ChatGPT Work, and the spreadsheet add-in today, and whatever is added to the list tomorrow.

Today the meter is small. The help page limits it to Codex, ChatGPT Work, and ChatGPT for Excel, and says credit purchases are available to a limited group of Free and Go users. That is exactly how a company tests a meter before it puts one on every house. The features it is attached to are the ones where the intelligence works on its own for a while, which is the product the earlier pieces in this series were about. The chat window stays free or stays $20. The meter is for the part that runs while you are not looking.

I chose the phrase average order value deliberately. Nobody is going to set a price of $100 a month, and I am not predicting one. The $20 plan will still exist. So will a $2,000-a-month balance held by someone whose agents run all day. A hundred dollars is the mean of a very wide distribution, and it only takes a modest tail to get there. Here is one shape that does, and I want to be clear that it is mine, drawn to show the arithmetic, not a forecast of the tiers.

Notice what the meter does to the word customer. Under a subscription, a person is either paying or not, and OpenAI disclosed 50 million paying subscribers against 900 million weekly users in February 2026, a conversion rate of about one in eighteen. Under a meter, a Free user who buys $5 of credits once is a customer, at an order value of $5. Conversion stops being a wall and becomes a slope. The billion is not a billion subscribers. It is a billion people who touched the meter in a year, most of them lightly, some of them very hard.

A subscription asks a billion people the same yes-or-no question. A meter asks each of them a different one: how much.

The honest version of this section starts by conceding the scoreboard. By revenue run-rate, Anthropic is ahead right now. It disclosed a $47 billion run-rate when it closed its Series H in May 2026, and a figure of $65 billion was reported for late July. Bloomberg reported in August that OpenAI’s run-rate had passed $40 billion, roughly double where it ended 2025, and cautioned that the two companies may not be measuring the same way. Anthropic has been winning the enterprise buyer, and I said in the first piece that either company could collect the bill.

But the race to a trillion is not a race for revenue per customer. It is a race for the number of customers, and there the two are not close. OpenAI disclosed more than 900 million weekly users in February; reporting in late July had it nearing a billion. Sensor Tower’s estimate for Claude’s app, reported around the same time, was about 56 million monthly users, growing far faster in percentage terms and still an order of magnitude smaller. The meter is a distribution product. It converts attention into small charges, and you need the attention first. A billion people already have the app on their phone. The meter is a settings-page away from all of them.

There is a second reason, and it is about temperament rather than scale. OpenAI is the company that shipped a consumer product to the world before the world asked for it, and then priced it with a free tier so wide that 900 million people walked through it. Anthropic’s instinct has been to sell to the people who already know what they want. Both are legitimate. Only one of them is how you get a billion people to put a card on file.

The company most likely to beat OpenAI to the first trillion is not an AI lab at all. Amazon reported $716.9 billion of net sales for 2025, up 12 percent. If it simply keeps doing that, it reports a trillion-dollar year for 2028. Walmart is a hair behind at $713.2 billion, but growing at 4.7 percent it does not arrive until the 2030s. So the real question this piece has to answer is whether OpenAI can get from $40 billion to a trillion before Amazon covers the last $283 billion at a jog.

The answer depends entirely on whether OpenAI can hold its current pace, so let me state the pace. Twenty billion at the end of 2025, forty billion by July 2026: a doubling in about seven months, which annualizes to a little more than tripling. Bloomberg also reported that Greg Brockman told staff the run-rate grew more than 20 percent month over month in July. If OpenAI sustains roughly tripling, the run-rate passes a trillion in 2029 and the first full trillion-dollar year is 2030. If it settles to merely doubling, the run-rate crosses in 2031, and Amazon wins by three years or more.

So the claim reduces to this. Amazon needs to slow from 12 percent to something like 10, which would push its trillion to 2029, and OpenAI needs to keep roughly the growth it has shown this year for three more years, which puts its run-rate across the line in 2029 and its reported year in 2030. I think both happen, and I think the second is more likely than the first. Retail at Amazon’s size has already mean-reverted once. A meter attached to a billion people who are only just starting to let the intelligence work unattended has not begun to.

There is one more thing the Amazon comparison clarifies. Amazon’s trillion, when it comes, will be a trillion dollars of other people’s goods passing through a very good pipe. OpenAI’s will be a trillion dollars for a thing that did not exist as a category five years ago, sold one meter at a time, with the cash arriving before the cost. The number will be the same. What it says about the world will not.

As with the rest of the series, I want the claim to be checkable rather than just confident. Here is what is already true, what still has to happen, and where I am most likely wrong.

Already true

Roughly a billion people use ChatGPT every week, by OpenAI’s February disclosure and July reporting. The meter exists, with prepaid balances, twelve-month expiry, and automatic reload, on OpenAI’s own help pages. The run-rate doubled in seven months. And no company on earth has reported a trillion-dollar year, so the title is open.

What has to happen

The meter has to go from three features and a limited test group to the default way every account beyond the chat window is billed. The features attached to it have to become things people run unattended, so that spending is a decision made once, not a hundred times. Roughly tripling has to hold for three more years, which is a rate no company has held at this scale; and the compute to serve a billion metered customers has to be built, which is the constraint every lab is now fighting on the grid rather than in the lab.

Where I am probably wrong

Amazon. If it simply keeps its 2025 growth rate, it reports a trillion-dollar year for 2028, and OpenAI would need to do better than tripling to beat it. If OpenAI’s growth settles to merely doubling, this piece is wrong about the name and right about the number. The second risk is the same one I named in part one: the bill may be collected by whoever owns the device or the carrier plan, with OpenAI paid wholesale, in which case a billion people pay $100 a month for intelligence and it shows up as someone else’s trillion. The third is that I am reading a July growth spurt as a three-year trend. Company-reported run-rates are chosen numbers, disclosed when they are flattering, and the audited 2025 figure was $13 billion, not twenty.

If in 2031 the largest revenue line on earth belongs to a retailer, you will be able to hold this piece up and say so. If it belongs to a company that bills a meter, I want the record to show the meter already had a name in September 2026, and it was not subscription.