How much computing power does the Agent need? Barclays: More than 70% of total computing demand, with expenditures reaching 300 billion!

In the fierce competition among giants, the AI industry is at a rapid development turning point.

Recently, Barclays Capital released a research report on the future development of AI. The Ross Sandler analysis team predicts that by 2026, the daily active users (DAUs) of consumer AI will exceed 1 billion, while the adoption rate of enterprise AI agents may account for 5% of the 7 billion software tasks globally.

At that time, the demand for inference computing may account for more than 70% of the total computing demand for general artificial intelligence (GenAI), with inference computing demand exceeding training computing demand and reaching 4.5 times that of the latter.

However, existing chip resources may struggle to meet this demand. The report predicts that the industry needs to increase chip capital expenditures by four times the current forecast, totaling nearly $300 billion.

Barclays: AI Chip Demand to Surge 14 Times by 2027

The adoption speed of AI products is accelerating. Technology giants represented by OpenAI and Meta have surpassed 200 million AI users. At the same time, new voice products such as Google's NotebookLM and consumer and enterprise agents are beginning to enter the market, and the application of AI in real scenarios is gradually maturing.

In the coming years, the demand for computing resources will far exceed supply. Barclays expects that by 2025, the number of GPUs and custom chips needed for training and inference AI models will be 250% higher than current widespread forecasts, and by 2027, this number could even reach 14 times.

“We predict that NVIDIA GPUs currently hold about 80% of the inference market. However, as large tech companies continue to introduce customized ASIC chips, this proportion is expected to drop to around 50% by 2028.”

As AI application scenarios expand, computing demand is also growing rapidly. The report predicts that by 2026, the demand for inference computing may account for more than 70% of the total computing demand for general artificial intelligence (GenAI), with inference computing demand exceeding training computing demand and reaching 4.5 times that of the latter.

However, existing chip resources may struggle to meet this demand. The report predicts that the industry needs to increase chip capital expenditures by four times the current forecast, totaling nearly $300 billion.

Nevertheless, the unit cost of AI inference is rapidly decreasing at a rate of over 90% every 18 months, a cost-effectiveness improvement known as the "Jevons Paradox," which means that as computing costs decline, AI consumption will significantly increase.

Three Stages of AI Applications: From Chatbots to Enterprise AI Employees

Barclays expects that in the coming years, the development of AI technology may present a wave-like trend. The report divides the development of AI into three stages:

• 2023-2024: Exploration period represented by chatbots and Copilots.
• 2025-2026: The era of agents begins, with significant growth in the adoption of consumer and enterprise AI agents
• 2027-2028: The popularization of AI employees and consumer robots will profoundly change productivity models.

Although it is widely believed that AI frontier laboratories are suffering significant losses, a Barclays report emphasizes that these institutions may be profitable on each large language model (LLM). For example, OpenAI's GPT-4 may have contributed nearly $2 billion in profits over the past two years, demonstrating strong economic potential.

Barclays believes that AI technology is advancing at an unprecedented pace, with simultaneous increases in demand for AI in both consumer and enterprise markets.

By 2026, the daily active users of consumer AI will reach 1 billion, while the adoption rate of enterprise AI agents may account for 5% of the 7 billion software tasks globally. At that time, the demand for inference computing will exceed the demand for training computing, reaching 4.5 times that of the latter.

The AI industry is entering a turning point, expected to lead the global economy into a new growth cycle