Huawei’s 2035 AI Agent Forecast: 900 Billion Agents and 90% of Token Traffic

Huawei’s Intelligent World 2035: Turning Vision into Action presents a highly expansive vision of autonomous AI. Its headline projection is that, by 2035, 900 billion AI agents could serve a global population of 9 billion, while autonomous agents generate more than 90% of global AI-token traffic. 72

Those figures should be read as Huawei’s strategic scenario and infrastructure roadmap—not as an independently validated prediction of how many agents will exist. But the report usefully explains the scale of the technical shift Huawei expects: digital services would move from individual applications responding to people toward persistent software agents that plan, use tools and act across systems.

Why agents would consume far more compute

A conventional chatbot generally handles a discrete request and returns an answer. Huawei’s envisioned agents would operate in longer loops: they perceive inputs, retain context, reason about options, make decisions, call external tools and interact with people or other systems. Reuters reported that Huawei expects this continuous operation to drive a major rise in token consumption and to require substantially more computing power. 2

Huawei forecasts a 100,000-fold increase in annual AI-token consumption by 2035. Its roadmap calls for compute clusters that are 100 times larger and for a 1,000-fold reduction in the cost of agent tasks, alongside architectures designed for efficient token throughput rather than just single-chip performance. 2021

The pressure would not fall on processors alone. A large population of interacting agents would also require responsive, reliable networks; durable storage and memory; and data centers with sufficient electricity and cooling. Huawei describes an “agentic Internet” capable of connecting people and agents, and projects a 100-fold increase in data traffic over the next decade. 7

The 10 technology priorities in Huawei’s roadmap

Huawei frames its report around 10 directions for the agentic-AI era:

1. Artificial general intelligence (AGI)
2. Larger computing clusters and SuperPoD-style systems
3. New storage and memory systems
4. Intelligent connectivity
5. Chip-design methods focused on latency and energy efficiency
6. An Agent OS
7. Device intelligence
8. Intelligent driving
9. AI data-center power and cooling
10. Security and privacy for autonomous agents 1721

Together, these priorities describe a full stack: chips and clusters at the bottom; storage, connectivity and operating software in the middle; and intelligent devices, vehicles and security systems at the edge.

Why physical-world AI is central to Huawei’s AGI argument

Huawei argues that AI must “go physical” to reach AGI. In its view, models need to interact with real environments in real time—combining perception, cognition, decision-making and action—and learn from the results of those actions. 18

That is a stronger claim than simply making language models better. The report links AGI to embodied intelligence, scientific intelligence and high-fidelity world models. In practical terms, this points to systems such as robots, intelligent vehicles and devices that can sense their surroundings and adapt their behavior, rather than software that operates only through text and screens. 1821

It remains an aspiration, not proof that physical embodiment is the only route to AGI. Huawei’s position is that real-world feedback will be necessary for agents to handle complex, changing tasks more robustly.

Security becomes an infrastructure problem

When agents can access data, APIs, software tools, credentials or physical equipment, the consequences of failure extend beyond an inaccurate response. The relevant concerns include unauthorized actions, data exposure, manipulated tool calls, compromised agents and the difficulty of auditing or stopping an automated process.

Huawei therefore places security and privacy among its 10 core priorities. Reuters reported that the company sees new systems for keeping independent software secure and controllable as necessary alongside the projected growth in agent activity. 217

Huawei has also introduced a Xinghe AI Network Security Agentic SOC, intended to help enterprises build autonomous security-operations systems. 22 Its proposed AI-powered firewall framework emphasizes intelligent detection, real-time defense, reliable and high-performance protection, and autonomous operations. 19

These systems may help organizations detect and respond to threats more quickly, but they do not eliminate the need for basic agent controls. High-impact deployments still need clear identity boundaries, least-privilege access to tools and data, logs that support investigation, and defined points for human escalation. Security tools are most useful when they are part of a broader control architecture rather than treated as a substitute for one.

China and the US: different emphasis, shared control concerns

Huawei’s roadmap favors scaling deployment while building infrastructure, technical standards and controls around the resulting risks. China has publicly identified concerns that advanced AI could replicate, seek power or evade human control, and Reuters reports that Beijing is developing regulation aimed at preventing such outcomes. 3

The difference from many US frontier-AI debates is primarily one of emphasis. In China, the policy posture described by Reuters leans more heavily on state oversight while promoting rapid diffusion of AI through the economy. In the United States, debates have placed greater attention on whether frontier systems should be slowed or gated until developers can show stronger safeguards against misuse and loss of control; company-level practices play a comparatively prominent role. 3

Neither framing treats control as irrelevant. The practical disagreement is over whether controls should chiefly be built in parallel with accelerated deployment or demonstrated before the most capable systems are widely deployed.

The strategic message behind the forecast

Huawei’s report is also an argument for integrated AI infrastructure. At the scale it imagines, AI capability depends on clusters, networking, storage, operating software, devices, power systems and cybersecurity—not solely on an accelerator chip.

That approach aligns with China’s broader push to develop and deploy AI across its economy. Brookings notes that the United States retains an edge at the technological frontier, particularly in compute scale and model performance, while China is advancing through efficiency, open-source diffusion and integration into the real economy. 1

The report’s lasting value is therefore less the precision of its 2035 numbers than the challenge embedded in them: if autonomous agents become persistent, tool-using participants in digital and physical systems, the bottleneck will be the entire infrastructure and governance stack required to make them useful, affordable, secure and controllable.