Category: Technology | Published: 2026-09-25
The future of AI could involve vastly more automated activity than today, according to a new forecast from Huawei. The technology company predicts that worldwide AI token consumption may increase 100,000-fold by 2035, with autonomous agents responsible for more than 90 per cent of the traffic.
Those figures are striking, but they are projections rather than measured outcomes. They describe the scale of the future Huawei expects and the infrastructure it believes will be needed to support it. Understanding what sits behind the numbers is more useful than treating the headline forecast as inevitable.
If AI agents become common across businesses, public services and consumer technology, they could generate far more computing work than a person asking a chatbot an occasional question. That would affect data centres, electricity demand, networks, storage, security and the way organisations control the cost of automated tasks.
What Has Huawei Predicted About the Future of AI?
Huawei published two reports on 16 September ahead of its Huawei Connect conference. Intelligent World 2035 focuses on how technology may develop, while the Global Digitalization and Intelligence Index 2026 considers the potential economic effect.
The second report was developed with Tsinghua University's Institute of Economics. It forecasts more than US$27 trillion in cumulative economic value from AI over five years and predicts that annual investment in digital and intelligent infrastructure could exceed US$4 trillion by 2030.
These are ambitious estimates, not guaranteed results. Huawei also has a clear commercial interest in the argument because it supplies computing, networking, storage and power infrastructure. Its reports offer a useful view of the future the company is preparing to serve, but the figures should still be examined as forecasts produced by a major industry supplier.
The 100,000-fold token prediction is especially dramatic. It does not mean every person will send 100,000 times more messages to a chatbot. Much of the expected growth comes from software agents performing chains of work in the background.
What Does AI Token Consumption Mean?
AI systems divide information into units called tokens. A token can represent a whole word, part of a word, punctuation or another form of data processed by a model.
Token use includes both the information entering a system and the content it generates. If an AI assistant summarises a document, it must process the original text before producing the summary. A visible answer of a few paragraphs may therefore require the model to handle far more material than the user sees.
The difference becomes larger when an agent completes a multi-stage task. It might read a request, plan its approach, search several databases, inspect documents, compare results, correct an unsuccessful step and then prepare a final response. Every stage can consume additional tokens.
Token counts are useful because they help describe the amount of model processing taking place. They are not a complete measure of value, energy or cost. Different models, hardware and software designs can use resources in different ways, even when they process a similar number of tokens.
Why Autonomous Agents Could Drive AI Growth
A conventional chatbot usually waits for a question and returns an answer. An AI agent can be given a broader goal and use tools or connected systems to work towards it.
For example, an agent might investigate why sales have fallen, collect information from several business systems, compare periods, test possible explanations and prepare a recommendation. Several agents could divide the work, pass findings to one another and review each other's conclusions.
That activity can create substantial processing demand. Huawei identifies agentic AI as a central part of its forecast and expects agents to generate more than 90 per cent of future token traffic.
There is already some evidence that agent-based systems use more tokens than simple conversations. In a June 2025 account of its research system, Anthropic reported that agents typically used about four times as many tokens as ordinary chat interactions. Systems using multiple agents consumed roughly 15 times as many.
Anthropic's figures describe its own experience and do not prove Huawei's global prediction. They do, however, show why a shift from occasional answers to continuous, multi-step activity could increase demand quickly.
What Infrastructure Would the Future of AI Need?
Huawei identifies ten areas that it believes need further development. These include computing clusters, memory, network connectivity, chip design, autonomous systems and security.
Among its ambitions are increasing the scale of computing clusters 100-fold and reducing the cost of completing an agent task 1,000-fold. The two targets are closely connected. More AI work becomes commercially possible if the price of each completed task falls sharply.
Lower unit costs do not necessarily mean lower total spending. If an automated task becomes much cheaper, organisations may run it far more often or apply it to areas that were previously uneconomic. Overall consumption can rise even while each individual task costs less.
The same pattern is familiar from other technologies. More efficient computing has repeatedly enabled new services and increased total use. The future of artificial intelligence may follow a similar path, with efficiency improvements encouraging a much larger volume of automated work.
Memory and reliable information are also essential. An agent needs to retain useful context, retrieve the correct records and show where important conclusions came from. More processing power cannot compensate for poor data or an inability to trace a result back to its source.
Can Electricity and Data Centres Keep Up?
A large increase in AI activity would place additional pressure on data centres, electricity networks and cooling systems. However, a 100,000-fold increase in token consumption would not automatically produce a 100,000-fold increase in energy use.
The relationship depends on several factors, including model efficiency, specialist chips, cooling methods, workload scheduling and how effectively infrastructure is shared. A smaller model running on efficient hardware may process a task using far less electricity than a larger system designed for a different purpose.
Even with efficiency gains, energy availability is becoming an important constraint. The International Energy Agency's 2025 Energy and AI report projected that global data-centre electricity consumption could more than double to approximately 945 terawatt-hours by 2030.
The IEA also warned that limitations in electricity grids could delay new projects. Building a data centre is only part of the challenge. Operators need sufficient power connections, generation capacity, cooling, network access and local approval.
This means the future of AI will be shaped partly by physical infrastructure. Software may develop quickly, but electricity grids, data centres and chip supply chains take time and significant investment to expand.
More AI Activity Does Not Automatically Mean More Value
Token growth is a measure of activity, not success. An agent can consume large amounts of processing while repeating searches, following irrelevant paths or producing work that still requires extensive correction.
For businesses, the useful question is not how many tokens an AI system can process. It is whether the system completes a worthwhile task accurately, safely and at an acceptable total cost.
That cost should include more than the supplier's usage charge. Organisations should count the time employees spend preparing data, reviewing results, correcting mistakes and dealing with failed tasks. An apparently inexpensive agent may provide poor value if people have to repair its work afterwards.
Trials should therefore measure completion rates, accuracy, staff involvement, running time and the quality of the final outcome. Usage data matters, but it needs to be connected to a clear business result.
How Businesses Can Prepare for the Future of AI
The most sensible starting point is to identify tasks where automation could create measurable value. A focused use case is easier to assess than a broad plan to add AI everywhere.
Set limits before allowing agents to operate continuously. Budgets, maximum running times and restrictions on paid services can prevent a faulty task from consuming resources without producing a useful answer. High-impact actions should require human approval.
Access controls also matter. An agent should only reach the systems and information needed for its job. Giving software broad access for convenience increases the damage that can follow from an error, compromised account or poorly designed instruction.
Businesses should ask suppliers how agent activity can be inspected. Teams need to see which tools were used, what information was retrieved, where costs increased and why a task stopped or failed. Without that visibility, it becomes difficult to improve performance or investigate unexpected behaviour.
Data quality remains fundamental. Agents cannot reliably automate a process built on incomplete records, conflicting definitions or information that employees do not trust. Preparing data and clarifying responsibilities may create more value than immediately buying additional AI capacity.
What Huawei's Forecast Does and Does Not Tell Us
Huawei's reports provide a picture of a highly automated future supported by much larger computing systems. They draw attention to real pressures around chips, networking, storage, energy, security and the economics of agent-based work.
They do not establish that token use will definitely rise by precisely 100,000 times or that the predicted economic value will be realised. Technology adoption depends on cost, reliability, regulation, public confidence and whether organisations can find tasks where AI consistently improves the outcome.
The forecast also comes from a company positioned to benefit from greater infrastructure spending. That does not make its analysis irrelevant, but it makes independent assessment important.
For decision-makers, the headline number should prompt practical questions. What work would agents perform? What resources would they consume? How would the result be checked? What happens when the task fails? Who can stop it? Which data and systems can it access?
A Practical View of the Future of AI
The future of AI is likely to involve more automated activity, but volume alone will not determine its value. The organisations that benefit will be those that connect AI use to worthwhile tasks, reliable information, sensible controls and clear human accountability.
Huawei's forecast shows how large the supporting infrastructure could become if autonomous agents move into everyday use. It also highlights why businesses need to measure outcomes rather than being impressed by activity.
A system that uses more tokens is not necessarily more intelligent or more useful. The important result is dependable work completed at a cost, speed and level of risk that makes sense for the organisation.
For businesses assessing where AI could deliver genuine improvements without losing control of cost or governance, our AI Consultancy page explains how we can help.