Who Gets Paid When You Ask AI a Question?
Let’s follow the money behind every AI answer.
You've reached your free article limit
You've read 0 of 1 free Pro articles.
A company beats earnings estimates, sales look good, and management sounds pleased. Then the stock falls, leaving you wondering what everyone else saw in the report.
So you put the report into a stock research app we’ll call “Earnings Desk” and ask what’s going on. The app points you to a weaker forecast a few paragraphs down. You still check the numbers yourself, but now you know where to look.
That answer took seconds. But getting it onto your screen took the efforts of several businesses, and they all want to get paid.
Start With the App
Say you pay Earnings Desk $20 a month. The company lets you build a watchlist, keeps relevant filings organized, and makes it easy to ask questions about them. But it doesn’t have to build the AI model that answers you.
It could send your question, along with the earnings release and the company’s quarterly filing, to a model provider such as OpenAI or Anthropic. The provider runs the model, sends back a response, and Earnings Desk shows it to you. You might never know that Earnings Desk is plugged into Claude or OpenAI’s GPT models.
The model was trained earlier, when its developer built its ability to work with language and information. Using those abilities to answer your question is called inference, and it costs money each time. Model providers typically charge for small pieces of text called tokens, counting both what the app sends in and what the model generates in response.
Take Nvidia’s latest 10-Q (NVDA), filed in August. The text of that filing comes to roughly 32,000 tokens. At GPT-6 Sol’s standard prices, processing the entire document costs about 6.5 cents. If producing the answer uses another 1,000 tokens, you’re at roughly 7.5 cents altogether.
If Earnings Desk sent that whole filing with each of a dozen questions a day, the bill would reach about $27 over 30 days, before any discounts for reusing the same material. That’s already more than the $20 subscription, before Earnings Desk pays its staff or covers any of its other bills.
One customer like that won’t break the business. But if enough customers use the app that way, adding subscribers could make the problem bigger. Earnings Desk could charge heavy users more, use a cheaper model for simpler questions, or find other ways to reduce the cost per answer. Whatever it does, the answers still must be good enough to keep customers coming back.
You can see why model prices matter. OpenAI says its new GPT-6 Sol and GPT-6 Luna improved on their predecessors in its evaluations, while the prices charged to apps are roughly half those of the comparable GPT-5.6 models. Sol’s price for a million input tokens fell from $4 to $2. That gives an app like Earnings Desk more room to serve customers profitably, provided the model handles their questions well.
What if the App Chooses Another Model?
Earnings Desk could also choose one of the downloadable models from Meta’s (META) Llama family and newer Muse Glimmer models, Alibaba’s (BABA) Qwen family, or DeepSeek. These are called open-weight models. Developers can get the trained model and run it on computers they own or rent. Each release comes with its own ground rules for using it.
Why would a company give a model away after spending so much to build it?
Meta says wider use brings outside improvements, feedback, and tools it can put to work itself. For Alibaba, sharing Qwen can help attract cloud customers who like the model but would rather pay Alibaba to run it than manage the hardware themselves.
Getting the model doesn’t make the answers free. Someone still has to run it, keep the computers running, and handle all those customer questions.
Earnings Desk could take on that work itself or hire Fireworks or Baseten, two privately held companies known as inference providers. They run models for other businesses and send the answers back to apps like Earnings Desk. Their latest funding rounds valued Fireworks at $17.5 billion in July and Baseten at $13 billion in June.
In that arrangement, the model’s creator may receive nothing from an individual answer, while the inference provider and the company supplying its computers both get paid.
Someone Supplies the Computers
Those answers need computing power. Amazon (AMZN), Microsoft (MSFT), Alphabet’s Google (GOOGL), and Oracle (ORCL) operate enormous cloud businesses that rent out computing resources. You’ll often hear them called hyperscalers. Their data centers serve all kinds of customers and workloads, including AI.
Then there are CoreWeave (CRWV) and Nebius (NBIS). They’re examples of neoclouds, a name for cloud companies that focus heavily on the computing power AI developers need. They rent out access to machines equipped with powerful chips for training models and producing answers.
A company can play more than one role. Google, for example, develops Gemini models and sells cloud services. Earnings Desk could buy several pieces from Google or choose a model from one supplier and computing power from another.
Following the money raises a few useful questions about these businesses. Can the app keep customers while covering the cost of their questions? Can a model provider earn enough to pay for building and running its models? And if the app changes models, which suppliers keep the business?
If Earnings Desk finds a less expensive model that gives us equally useful answers, we may never notice the change. We’d keep paying for the app, and the company supplying its computing power could keep getting paid, too.
We’ve followed the answer as far as those computers. The next part of the story starts inside them, with the chips that do the work and the companies that put them there.
At the time of publication, Byrne had no positions in any securities mentioned.
