Huawei Forecasts 900 Billion AI Agents by 2035 as Compute Demand Accelerates

Huawei is making one of the most aggressive forecasts yet for the rise of autonomous artificial-intelligence agents, predicting that the technology could reshape how computing capacity is consumed over the next decade.

Reuters reported on September 16 that Huawei expects as many as 900 billion active AI agents globally by 2035 and believes agents could account for more than 90% of AI token traffic. The figures come from Huawei’s Intelligent World 2035 report and should be understood as the company’s long-range forecast, not an established outcome.

Even if the eventual number is much lower, the direction matters. AI is moving from systems that mainly answer questions toward software that can take actions, use tools, coordinate tasks and operate with less continuous human input.

What an AI agent actually is

An AI agent is software that can receive a goal, reason about steps, use digital tools and take actions toward completing a task. A conventional chatbot generally waits for a user prompt and returns an answer. An agent can potentially continue working through a multi-step process.

For a business, that could mean monitoring inventory, preparing reports, reconciling records, scheduling work or responding to routine customer requests. More advanced agents could coordinate with other agents, creating large volumes of machine-to-machine communication.

That helps explain why Huawei’s forecast focuses on token traffic. If software agents continuously read data, reason, communicate and act, they can consume far more model inference than occasional human prompts.

Why 900 billion agents does not mean 900 billion robots

The forecast is easy to misread. Most agents would be software processes rather than physical machines. A single company could operate thousands or millions of specialized agents across servers, devices and applications.

An agent might exist only to watch a narrow data stream or complete a repetitive workflow. Multiple agents could also serve one person, vehicle, factory or software platform.

That makes agent counts fundamentally different from user counts. The economic question is how much useful work those agents perform and how much computing they consume while doing it.

Compute demand could become enormous

AI inference already requires substantial data-center capacity. Autonomous agents can increase that requirement because they may generate many model calls while planning and completing a task.

If billions of agents operate continuously, demand for processors, memory, networking and electricity could grow rapidly. This would reinforce the infrastructure cycle already driving large investments in data centers.

Earnyx has covered how AI hardware demand is reshaping factories and supply chains. Agentic AI could extend that cycle by shifting compute demand from periodic human interaction toward persistent software activity.

Efficiency will matter as much as raw computing power

A future with huge numbers of agents is only economically practical if the cost of inference keeps falling. Running a powerful model for every small task would be unnecessarily expensive.

Businesses are therefore likely to use a mix of models. Smaller models can handle routine classification and extraction, while larger models can be reserved for difficult reasoning. Caching, local processing and specialized chips can further reduce cost.

The winning architecture may not be one giant model doing everything. It may be a layered system that routes each task to the cheapest model capable of completing it reliably.

Security becomes a much bigger problem

Agents are more powerful than chatbots because they can take actions. That also makes mistakes more consequential.

An agent with access to email, financial systems or cloud infrastructure could cause damage if it follows malicious instructions, misinterprets a request or receives excessive permissions. Attackers may also try to manipulate agents through data they read from websites, documents or messages.

Huawei’s report highlights security and privacy as important technologies for an agent-heavy future. Businesses considering agents today should already apply familiar security principles: minimum necessary access, approval requirements for high-risk actions, detailed logs and clear separation between trusted and untrusted data.

Human oversight will not disappear

Automation does not eliminate accountability. Companies still need people responsible for defining goals, reviewing outcomes and handling exceptions.

The level of oversight should depend on the consequences of an error. An agent drafting an internal meeting summary can operate with more freedom than one initiating payments or changing production equipment.

This suggests that enterprise adoption will progress unevenly. Low-risk administrative workflows can automate quickly, while regulated and safety-critical processes may require extensive controls.

What the forecast means for software companies

If agents become a dominant interface, software could shift from applications designed primarily for humans toward services designed to be called by other software.

That increases the value of clean APIs, structured data and reliable identity systems. A company may need to know not only which employee is authorized but which agent is acting on that employee’s behalf.

Pricing models could also change. Software sold per human seat may be awkward when one employee supervises dozens of automated agents. Vendors may increasingly charge for usage, transactions or outcomes.

Why Huawei is emphasizing the agent future

Huawei has substantial interests in computing infrastructure, networking and enterprise technology. A future in which AI traffic expands dramatically would create demand across those areas.

That commercial context is important when interpreting a forecast as ambitious as 900 billion agents. It is a scenario from an industry participant with incentives to highlight rapid AI expansion.

Still, the underlying trend is visible across the technology sector. Major software and cloud companies are already building agent frameworks and tools that allow models to interact with external systems.

What businesses should do now

Companies do not need to prepare for hundreds of billions of agents. They do need to understand which workflows are suitable for limited, measurable automation.

A sensible starting point is a process with clear inputs, clear success criteria and reversible actions. Businesses can measure time saved, error rates and operating cost before expanding deployment.

They should also track the total cost of inference. An agent that saves ten minutes of employee time but consumes expensive model calls may not create meaningful value.

The bigger picture

Huawei’s 2035 forecast is intentionally long-range and highly uncertain. No one can know today whether the world will operate 900 billion AI agents nine years from now.

What is more certain is that AI systems are becoming increasingly capable of acting rather than merely answering. That transition changes the infrastructure, security and economics of AI.

If agents become persistent digital workers, the next computing boom will not be measured only by how many people use AI. It will also be measured by how many automated processes are running continuously on their behalf — and whether the value those processes create can justify the enormous computing resources behind them.