Data, not algorithms, is the lasting layer of AI

An investor in Mercor asked about the data suppliers behind AI, naming Mercor, Handshake, Surge, and Scale as companies above a billion dollars in revenue, with others doing hundreds of millions. Anastasios (Andy) Kyrtsos, the chief executive of Arena, a company that tests how AI models perform on real work, said he has a thesis for why their growth will last.

His line was short. “Data plus GPUs equals model,” he said. GPUs, the chips used to train AI, get most of the attention, but he argued that data is the other half of the equation.

He explained it with cars and gasoline. In economics, he said, some goods are complements, meaning demand for one drives demand for the other. More cars sold means more gas sold. Data, he argued, is what he calls a scaling complement to AI models. As models get bigger, get more numerous, and spread to more businesses training their own versions, the need for data grows with them.

That is why he believes data is "less so of a commodity than even GPUs." He said data only becomes irrelevant once humans do, which would mean artificial general intelligence has been reached.

Frontier labs, the handful of companies building the largest models, now spend about 10 to 20 percent of what they spend on computing power on data, he said. If computing keeps growing and models keep scaling, Kyrtsos said, the data market will reach what he called “at least a hundred billion dollars by 2030 if not a trillion.”

The questioner then posed a hypothetical, that Anthropic and OpenAI could be worth $3 trillion to $5 trillion, and asked whether that could make a data provider worth $100 billion. Kyrtsos answered, “I think these companies will easily be worth hundreds of billions of dollars.”

The same discussion raised the obvious objection, that these providers depend heavily on a few customers such as OpenAI, Anthropic, and Meta. Kyrtsos dismissed the concern, saying investors had become too fussy about revenue concentration, and pointed to chipmaker TSMC and to government suppliers such as Anduril as hugely concentrated businesses that do very well.

Sourcing that data is the hardest and dirtiest part of training, he said, work nobody wants to do, while the underlying design, a standard method called the transformer, is now widely understood.