AI enters the efficiency era as Wall Street looks beyond large models

Artificial intelligence has moved beyond its initial expansion phase, marked by tools like ChatGPT, and is now entering a new era of business adoption. Companies are no longer solely seeking the most powerful AI models but are prioritizing those that offer greater efficiency and a clear return on investment, according to analysts.

“

The first phase of AI was about demonstrating that the models worked; the next consists of demonstrating that they generate enough value to justify the expense.

The consumption of 'tokens', the unit used to measure AI model usage and associated costs, has become a critical efficiency metric. As companies scale up their AI initiatives, the cost of tokens transitions from a technical consideration to an operational expense that financial directors closely scrutinize. A recent UBS survey revealed that 60% of companies have implemented spending restrictions on AI to maximize their return per dollar invested.

“

When companies move to production, token costs stop being a technical issue and become an operating expense that financial directors start to monitor.

This shift signifies a move from an era focused on creating models to one centered on maximizing their performance. Karl Freund, founder of Cambrian-AI Research, notes that as AI models are used millions of times and AI agents emerge, the market is pivoting towards efficiency. This evolution is expected to alter the value distribution within the AI sector, with value potentially shifting towards the layers that enable businesses to leverage AI effectively.

“

When AI focused on training and developing new models, time-to-market and performance were most important. Now that those models are used millions of times and AI agents arrive, the market is turning to efficiency.

Goldman Sachs reports that companies heavily utilizing AI consume three times more tokens than average companies, driving demand for chips, memory, and data centers. The next phase is anticipated to be characterized by the expansion of inference, the use of trained models to perform tasks with new data, and sustained investor interest in chip manufacturers and cloud providers. While Nvidia, AMD, Intel, and Micron have seen significant stock gains driven by AI infrastructure demand, the future may require more comprehensive solutions beyond just chip sales.

“

If AI becomes cheaper and more accessible, value will shift to the layers that help companies use it effectively.