Artificial intelligence may be approaching a new inflection point. Huawei estimates that autonomous AI agents will account for more than 90% of global token traffic by 2035, a shift that would multiply demand for computing capacity, storage, and networking over the next decade.
The forecast is part of Intelligent World 2035: Turning Vision into Action, the Chinese technology company’s latest report. Huawei estimates that annual token consumption could increase 100,000-fold by 2035, driven primarily by the development of systems capable of carrying out tasks with increasing degrees of autonomy.
The scale of the estimate should be understood as the company’s own projection rather than a guaranteed market outcome. Still, it offers an indication of where a significant part of the technology industry is headed: from artificial intelligence that responds to specific prompts to systems capable of working for extended periods, using different tools, and carrying out complete chains of actions.
From asking AI questions to assigning it tasks
The distinction is significant. In today’s conversational models, the user submits a query and artificial intelligence generates a response. Agents add new capabilities: they can assess a situation, reason through the required steps, make certain decisions, use external tools, and continue working until an objective is completed.
Each of these operations requires processing information. Tokens are the units into which models divide the data they receive and generate in order to process it. The greater the number of agents and the more complex their tasks, the more infrastructure will be required to keep them running.
Huawei starts from an earlier forecast that there could be about 900 billion AI agents worldwide by 2035. Based on that scenario, it has now developed a technology roadmap identifying 10 priority areas, ranging from computing capacity and networks to storage, operating systems for agents, energy, and security.
The company believes it will be necessary to increase the scale of computing clusters 100-fold and sharply reduce the cost associated with tasks carried out by agents. The economics will be especially important if these systems move beyond occasional use and begin operating continuously within companies, devices, and industrial processes.
Huawei is also accelerating its AI infrastructure
The report’s release coincides with new Huawei announcements on artificial intelligence infrastructure. During Huawei Connect 2026, held in Shanghai, the company unveiled new computing systems and an architecture designed to connect large numbers of processors so they can work in coordination.
Its announcements include infrastructure capable of scaling to as many as 1 million NPUs, processors specialized in artificial intelligence operations. The strategy seeks to offset the individual performance limitations of certain chips through large interconnected systems.
The move also comes amid strong domestic demand. Huawei acknowledges that it currently lacks sufficient production capacity to meet all demand for its AI computing equipment in China, a situation that is also constraining its international expansion.
At the same time, the company is preparing new generations of Ascend processors and aims to accelerate some launches scheduled for 2027.
The challenge will not be purely technological
The widespread adoption of agents capable of acting with greater autonomy also raises questions about security, privacy, and control. Huawei includes these areas among the 10 technology priorities it considers necessary to develop the new ecosystem.
An agent limited to searching for information poses different risks from one that can access corporate systems, run programs, conduct transactions, or communicate automatically with other applications. As its ability to act increases, so does the need to establish permissions, oversight systems, and mechanisms to control its decisions.
Ultimately, Huawei’s forecast points to an artificial intelligence that is less visible as a standalone application and far more integrated into the day-to-day operations of companies, devices, and organizations.
Nine years remain to determine to what extent these extraordinarily ambitious figures materialize. But the direction identified in the report is already apparent: the next stage of artificial intelligence is not solely about models capable of providing better responses, but about systems able to perform an increasing number of tasks on their own.
