New York .- Artificial intelligence (AI) has entered a new phase of corporate adoption after nearly four years of the expansion of tools such as ChatGPT, Claude, or Gemini, in which companies are no longer looking for the most powerful models, but rather those that offer greater efficiency and return on investment, according to analysts consulted by EFE.
“The first phase of AI was about proving that the models worked, the next is about proving that they generate enough value to justify the spend,” Shay Boloor, chief market strategist at Futurum, explains to EFE.
For the expert, the consumption of ‘tokens’, the unit used to measure the usage of models and their cost, has become a key metric of efficiency.
Each question, answer, or piece of code consumes ‘tokens’ and, the more complex the task, the higher the cost: “When companies move to production, ‘token’ costs stop being a technical issue and become an operating expense that CFOs start to monitor,” he adds.
Corporate limits on AI spending
According to a UBS survey of 130 companies, 60% have already implemented restrictions on AI spending to maximize the return per dollar invested.
More Read
The bank cites cases such as an employee who spent up to 35,000 dollars a month on ‘tokens’ or equipment that exceeded weekly limits by between 100% and 200%.
Karl Freund, founder of Cambrian-AI Research, believes this shift reflects the transition from a phase focused on creating models to one focused on maximizing their performance.
We recommend reading:
More Read
“When AI focused on training and developing new models, time-to-market and performance were the most important things. Now that those models are being used millions of times and AI agents are arriving, the market is shifting toward efficiency,” he explains to EFE.
This stage will modify the distribution of value in the sector, according to Boloor: “If AI becomes cheaper and more accessible, value will shift toward the layers that help companies use it effectively.”
Semiconductors and cloud, on the rise
In a recent report, Goldman Sachs notes that companies that use AI the most consume three times more ‘tokens’ than average companies, a trend that is driving demand for chips, memory, and data centers.
He adds that the next phase will be marked by the expansion of inference, that is, the use of already trained models to execute tasks with new data, and by greater investor interest in chip manufacturers and cloud providers given the current absence of developers like OpenAI and Anthropic from the stock market.
Nvidia, a leader in data center chips, has accumulated a rise of nearly 13% in the last six months, while AMD has soared 167%, Intel 150%, and Micron 186%, favored by the demand for AI infrastructure.
But selling chips will no longer be enough: “They will continue to be important, but they will have to offer complete solutions to cloud operators,” points out Freund.
Thus, cloud providers are among those benefiting from the expansion of hardware, notes Morgan Stanley, including Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle, as well as digital infrastructure companies like Cloudflare, which rose 45% in the last semester.
China and open models gain ground
The pressure to reduce costs favors cheaper, open-source models, such as those developed by China’s DeepSeek, Alibaba, or Tencent, according to Boloor.
For the expert, this trend will intensify competition among AI developers and make it difficult for a single provider to maintain high prices for a long time.
“The market is moving towards a layered model: the most powerful systems will be used for the most complex tasks, and cheap models will take on routine, high-volume jobs,” he explains.
Freund agrees that the industry is heading towards greater standardization: “There are already more than 2.9 million models available on Hugging Face, where researchers and companies share AI models. The models of the future will be like the compilers of the past. And nobody makes money from compilers,” he states.
In this scenario, companies capable of applying AI to specific operations, managing data, and developing agents, such as Salesforce, Microsoft, or Palantir, gain importance.
In fact, the CEO of the latter firm, Alex Karp, recently criticized the generalist model based on ‘tokens’, considering that it “increases costs for users and reduces the return on investment”.
For Boloor, the first phase of AI rewarded those who built the most powerful models, and the next will reward those who make the technology “cheaper, more useful, safer, and more integrated” into businesses.
