How much are customers actually paying for artificial intelligence? Not how much is being invested in chips and data centers, but how much real revenue it has generated. Until now, no one has been able to answer that question precisely. A team led by Azeem Azhar of Exponential View has published a report that sheds light on the matter. Over the past twelve months, people and companies have spent more than $110 billion on generative AI.
Most of the debate about AI has revolved around supply. Chips, energy, water, and gigantic data center buildings. We know plenty about supply. Demand was a fog of press releases and round numbers that no one could piece together. The problem has two causes. Most of the money flows to private companies such as Cursor or ElevenLabs, which are under no obligation to disclose anything. And large publicly traded companies hide their AI revenue within larger segments, making it impossible to extract from their financial statements.
The second problem is even more insidious. The same dollar is claimed by several companies at once: the cloud provider, the model developer, and the company with an application built on top of it. If you spend a dollar with Anthropic on Claude and Anthropic sends fifty cents to Amazon to run it, that can easily be counted as a dollar and a half. In reality, however, the customer paid only one dollar.
Complex calculations
The authors therefore adopted a simple rule. They count the dollar spent by the end customer and do not count it again as it moves through the supply chain. If an application collects one hundred dollars, sends sixty of that to the model developer, which then spends another thirty on hosting, the report records one hundred dollars. Not one hundred and ninety.
The model is based on data from more than a thousand companies. Every revenue figure has a source: financial statements, audited accounts, earnings call transcripts, or credible reporting. Where a private company’s revenue appears in the accounts of a public company, the authors traced it through the cloud. OpenAI can be captured this way through Azure, and Anthropic through Bedrock. Leaked information and companies’ own statements were scored according to their credibility.
This produced a line-by-line financial model rather than an estimate. That is the difference between “the market could be worth X” and “this company actually collected this much.”
Revenue has finally caught up with the hardware bill
In the first quarter of 2026, global AI revenue outside China reached roughly $25 billion. The authors estimated depreciation on data centers and chip infrastructure over the same period at $21 billion. For the first time—actually, for the second quarter in a row—the money coming in exceeded the cost of wear and tear on buildings and hardware. In other words, the machine is beginning to pay for itself.
But beware: margins are very weak. Depreciation consumes most of the revenue, and quarterly depreciation alone comes nowhere close to covering all the investments accumulated historically, let alone leaving room for a healthy profit. Azhar and his colleagues summarize it this way: large, but still small. And still early. Even for the companies spending the most on AI, it is a rounding error in their accounts. Uber spends around $1,500 per engineer, which barely moves the needle.
Cheaper tokens? Revenue will keep growing
This is a good place to dispel a widespread misconception. Many people expect total industry revenue to shrink as AI prices fall. The report shows that the arrow points in exactly the opposite direction.
Every ten percent drop in price attracts twelve to eighteen percent more token consumption. Total spending therefore rises even as the price per unit falls. This is not a new phenomenon. William Stanley Jevons described the same thing with coal as early as 1865. Make a resource cheaper, and people consume more of it, not less. Anyone expecting cheaper AI to reduce revenue is therefore mistaken.
The report makes one more proposal. It retires the token as a measure of value. It works fine as a billing unit, but using it to measure intelligence makes no sense. Counting tokens to measure a model’s intelligence is like counting pages to measure the quality of a library. Instead, the authors propose a quality-weighted token. This takes into account how many of the tokens produced actually reach the user as visible output and how capable the model behind them is. Raw volume versus actual usefulness. That is a major difference.
Shortcomings of the report
The $110 billion figure is more of a lower bound. The authors deliberately excluded several major areas. They do not count the internal benefits of AI at Meta or Google, such as when recommendation systems increase advertising revenue. They do not count the savings that internal tools generate for major technology companies. They exclude professional services and systems integration. And they do not include China. If global revenue outside China amounts to around $25 billion per quarter, Chinese revenue sits somewhere on top of that, and no one has measured it here.
In other words, the actual AI economy is probably even larger than the headline figure suggests.
Faster than the internet and mobile
When its growth rate is compared with previous waves of technology, AI is growing roughly three times faster than the internet or mobile phones did at the same stage. At the same time, the entire industry is still small. Compared with US GDP, AI revenue remains just a rounding error. For now, companies are mostly focused on efficiency and cost savings, although that mix is slowly changing. Many companies have moved beyond the pilot-project stage, but they are only beginning to learn how to scale AI fully and embed it into their operations.
In conversations with corporate leaders across Europe and the United States, from manufacturing and insurance to pharmaceuticals, Azhar keeps hearing the same thing. They intend to invest more in AI. And half of the executives surveyed believe their own jobs depend on whether they can master AI.
Source: intelligence.exponentialview.co
