Huawei says AI will burn 100,000 times more tokens by 2035

“Visions are painted in words, but measured in deeds,” David Wang wrote in the foreword to Huawei’s latest forecast. The company published two reports days before Huawei Connect 2026 opened, and they explain most of what it launched there.

The first is Intelligent World 2035, subtitled Turning Vision into Action. The second is the Global Digitalization and Intelligence Index 2026, built with the Institute of Economics at Tsinghua University.

Huawei has run this series for three years. The 2024 edition mapped trends and the 2025 edition named ten megatrends. This one names ten directions and says what would have to be built.

“Agentic AI is a key variable in this transformation,” Wang wrote. “The direction is clear, but bringing our vision to life demands concrete action.”

The number that drives everything else

One forecast sits underneath the rest. Huawei expects global annual token consumption to grow 100,000-fold by 2035. It expects agents to generate more than 90% of that traffic.

The report separates agentic AI from the assistants most people use. A chatbot responds when prompted. An agent perceives, reasons, plans, calls tools and keeps learning, which is a far heavier workload.

That is a claim about electricity as much as about software. A token is a unit of model output, and every one of them costs power, memory bandwidth and network capacity.

The report describes a shift from an application-centric digital world to an agent-centric one. In that world, software perceives its environment, reasons, calls tools and talks to people continuously. Each of those steps generates tokens.

Chinese policy is already moving the same way. A state report in September said the country’s AI industry is shifting from models to agents, with the compute burden moving from training to inference.

Ten directions, and what they give away

The ten directions are the useful part, because they read as a product roadmap rather than a vision statement.

The first direction is the broadest. Huawei argues that artificial general intelligence requires going physical, and that this needs advances in language intelligence, embodied intelligence and scientific intelligence at once. Another direction treats vehicles as mobile embodied agents.

Huawei wants computing clusters to scale 100-fold and the cost of an agent task to fall 1,000-fold. It argues the industry has to move to SuperPoD-style systems to get there, which is precisely what it spent Connect 2026 selling.

Another direction covers storage and memory systems with causal accuracy and traceable provenance. That is the argument for the context memory storage cluster Huawei launched in Shanghai.

A third is the Tau Scaling Law, the chip design method Huawei unveiled in May. It proposes replacing geometric scaling with time scaling, which matters a great deal to a company that cannot buy the newest lithography.

Two more are worth noting. One calls for breaking power and thermal limits in AI data centres. The last one covers security and privacy protection for autonomous agents.

There is also an Agent OS direction, summarised by a formula. Huawei writes it as coordination, execution, memory and connectivity multiplied together, then raised to the power of evolution.

The $27trn index

The second report is the economics. Huawei and Tsinghua forecast more than $27trn in cumulative AI economic value over the next five years.

They also expect global investment in digital and intelligent infrastructure to grow at a compound annual rate of 19.14%, passing $4trn by 2030. The index assesses 90 countries and sorts them into three stages, labelled Builder, Adopter and Frontrunner.

The framing is the interesting bit. The report reclassifies the core factors of production for a digital economy as data, ICT talent, and digital and intelligent technologies.

It then argues that integrating networking, computing and storage is what makes industrial upgrading possible. Huawei sells all three, which is worth holding in mind while reading.

How to read a vendor forecast

None of this is disinterested. A 100,000-fold increase in token consumption requires an enormous amount of new infrastructure, and Huawei manufactures that infrastructure.

The same caution applies to every agent-economy projection currently in circulation, including the ones from American vendors. We have already gone through those numbers and found the methodologies thin.

What a vendor forecast is genuinely good for is reading the vendor. These reports tell you what Huawei believes, what it is building towards and where it thinks its constraints are.

On that reading the tells are clear. The Tau Scaling Law direction is about sanctions. The power and thermal direction is about grid limits. The cluster scaling direction is about compensating for per-chip performance it cannot yet match.

Where Europe sits in it

The index puts countries in three buckets and says there is no universal path to an intelligent economy. Its stated purpose is to let each country identify its own areas of focus and map a pathway suited to its conditions.

Europe is short of the compute, and its own institutions keep saying so. ECB president Christine Lagarde said this month that Europe must build its own AI capacity or risk being cut off.

Huawei is selling into that shortfall. Its latest AI cluster service reaches markets outside China on 30 November, and its enterprise agent platform follows on 30 December.

Huawei’s report closes with five commitments: take a long-term approach, stay people-centric, uphold tech for good, remain open, and think in systems. They are the same commitments every large vendor makes.

The difference is the hardware attached to them. Huawei is one of very few companies publishing a decade-long forecast and simultaneously shipping most of the equipment that forecast requires.