Change in the “brain” of AIs is creating a market worth more than US$100 billion

A tone of relief and optimism marked the latest earnings release conference calls at processor developers Intel (INTC) and AMD (). After watching the rise of generative artificial intelligence technologies generate little impact on their results, companies finally began to see the impact of infrastructure investments on their results this year.

The types of processors developed by Intel and AMD are increasingly important within data centers responsible for supporting the training and use of artificial intelligence. To a large extent, the adoption of AI agents is the main element behind this movement.

Since the end of 2022, with the launch of ChatGPT, companies responsible for developing language models, such as OpenAI itself or its competitors Alphabet () and Meta () have spent billions to build the computational capacity necessary to support the mass use of their applications. During this period, the industry’s main spending was focused on a specific type of processing cards: GPUs.

Study abroad

Upgrade your career!

Graphics Processing Units (or graphic processing units, in Portuguese) are chips specialized in graphic manipulations. Created around two decades ago to operate in parallel to the main computer processing unit, these cards are more efficient in performing the operations necessary to create three-dimensional objects seen on screens, called matrices.

This is why GPUs are important in video cards for video games, where there is a high need for real-time graphics processing, but they also work to improve the quality of video calls or broadcasts. It was in this market that companies like Nvidia () have consolidated themselves since their foundation.

It turns out that the principle behind creating a three-dimensional shape is very similar to the type of calculation involved in neural networks — the basis of generative artificial intelligence. “The neural network is essentially a multidimensional matrix problem. That’s why GPUs are useful for AI”, explains the CTO of Scala Data Centers, Agostinho Villela.

GPUs are efficient in the two main stages of building generative artificial intelligence: model training, where parameters are adjusted so that the AI ​​understands patterns based on data; Inference is the model’s ability to provide answers based on the request made by the user.

It was this ability that led developers of AI language models to fill their data centers with Nvidia chips and made the company the first in history to reach a market value of US$5 trillion. Since the explosion of ChatGPT, the company’s shares have appreciated by more than 1,000%.

This is also what made many analysts point to Nvidia as one of the big winners of the revolution brought about by artificial intelligence in the semiconductor industry. For much of this period, the ability of companies like Intel and AMD to benefit from the AI ​​boom was called into question.

CPUs gain relevance for semiconductors

Intel and AMD have consolidated themselves in the processor sector by developing the so-called Central Processing Units (CPUs, or central processing unit in Portuguese), chips with more generic functions in the functioning of a computer. “The GPU is very specialized in matrix operations. The CPU does everything: from starting the operating system to orchestrating, opening files and manipulating them”, points out Villela.

In data centers, at least one CPU is always required to operate GPUs. For some time, however, this proportion was much more unbalanced.

During AMD’s Q1 earnings call, CEO Lisa Su said, “We certainly see a trend toward something closer to a one-to-one configuration. Or even exactly one-to-one. In the past, that ratio was one-to-four or one-to-eight.”

There are two main reasons for the change of scenery. First, the so-called reinforcement learning (reinforcement learning, in Portuguese), process by which artificial intelligence is tested with problems for which the answer is already known. With each correct answer, she receives a kind of positive reinforcement, indicating that she is on the right path.

“O reinforcement learning It is often done through mathematical problems. The neural network needs to non-deterministically model a problem for which the answer is secretly known deterministically. This training is generally controlled by a CPU”, says Villela.

The great expectation for increased demand for CPUs, however, is related to the proliferation of generative artificial intelligence agents — tools capable of carrying out tasks autonomously for their users. They can access data from different sources and, based on it, perform an action, such as sending emails, creating reports or analyzing data.

While GPUs are more efficient for much of AI training and inference, they cannot perform all of the tasks of a CPU. “Agentic artificial intelligence involves a lot of integration: talking to a system, opening files, making a series of integrations with the external world. The CPU is the one who does this”, points out Villela. “Agentic AI is probably the biggest AI energy hog on the planet today.”

In a recent projection, CitiGroup estimated that the CPU market is expected to grow from $29.3 billion in 2025 to $131.5 billion in 2030, an annual growth rate of 35%. The bank raised sales estimates for Intel and AMD, two of the main developers in the sector.

Bank of America estimates that AMD will hold about a 50% share of a $120 billion market for CPUs in fiscal 2030 — with the remainder split between Intel and competitors like ARM.

In a recent move, Nvidia increased its stake in Intel after purchasing 214.8 million for $9.48 billion. Last month, in the quarter ending in June, above average market estimates of US$13 billion according to data compiled by Bloomberg.

This year, AMD shares have risen around 85%, while Intel has appreciated around 170%.