Huawei has published a forecast that global annual token consumption will grow 100,000-fold by 2035, with agents generating more than 90% of that traffic. The figure comes from the company's Intelligent World 2035 report, released days before Huawei Connect 2026 opened in Shanghai. For businesses, the number matters less as a prediction than as a signal of where infrastructure spending is heading: a token is a unit of model output, and each one consumes power, memory bandwidth and network capacity.
What the two reports contain
Huawei released two documents. The first, Intelligent World 2035, carries the subtitle Turning Vision into Action and names ten directions rather than the ten megatrends of the 2025 edition. The second, the Global Digitalization and Intelligence Index 2026, was built with the Institute of Economics at Tsinghua University and assesses 90 countries, sorting them into three stages labelled Builder, Adopter and Frontrunner. David Wang, in the foreword, wrote that agentic AI is a key variable in the transformation and that the direction is clear, but bringing the vision to life demands concrete action. Huawei has run this series for three years: the 2024 edition mapped trends, the 2025 edition named ten megatrends.
The reports separate agentic AI from the assistants most people use. A chatbot responds when prompted; an agent perceives, reasons, plans, calls tools and keeps learning, which the report describes as a far heavier workload. That distinction drives the token forecast, because each step in an agent's loop generates output. The ten directions read as a product roadmap rather than a vision statement. Huawei wants computing clusters to scale 100-fold and the cost of an agent task to fall 1,000-fold, and argues the industry must move to SuperPoD-style systems to get there. Another direction covers storage and memory systems with causal accuracy and traceable provenance, the argument behind the context memory storage cluster Huawei launched in Shanghai. A further direction is the Tau Scaling Law, the chip design method unveiled in May, which proposes replacing geometric scaling with time scaling. There is also an Agent OS direction, summarised as coordination, execution, memory and connectivity multiplied together and raised to the power of evolution.
The economics sit in the second report. Huawei and Tsinghua forecast more than $27trn in cumulative AI economic value over the next five years, and expect global investment in digital and intelligent infrastructure to grow at a compound annual rate of 19.14%, passing $4trn by 2030. The index reclassifies the core factors of production for a digital economy as data, ICT talent, and digital and intelligent technologies, then argues that integrating networking, computing and storage makes industrial upgrading possible. Chinese policy is moving in the same direction: 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. The framing is not disinterested. A 100,000-fold increase in token consumption requires enormous new infrastructure, and Huawei manufactures that infrastructure. The same caution applies to agent-economy projections from American vendors, whose methodologies have been found thin.
What this means for business
For companies adopting AI, the practical consequence is a change in what they buy. If agent workloads dominate, spending shifts from training capacity to inference, from one-off model runs to continuous execution, and from software licences to systems that combine networking, computing and storage. A small company will notice this mainly through the price and availability of agent platforms: Huawei's enterprise agent platform reaches markets outside China on 30 December, and its latest AI cluster service follows on 30 November. A large organisation with its own data centres faces a different question, because cluster scaling and power limits become procurement criteria rather than abstractions. The report's own stated purpose for the index is to let each country identify its own areas of focus, since there is no universal path to an intelligent economy.
What the forecast does not mean is that token demand will grow as stated. It is a vendor projection, and Huawei sells the equipment the projection requires. The Tau Scaling Law direction reflects sanctions that limit access to the newest lithography; the power and thermal direction reflects grid limits in AI data centres; the cluster scaling direction compensates for per-chip performance Huawei cannot yet match. When evaluating any agent platform, the questions worth asking are which inference workloads it supports, how memory and provenance are handled, and what happens to cost per task as volume rises. Europe's position illustrates the stakes: ECB president Christine Lagarde said this month that Europe must build its own AI capacity or risk being cut off, and Huawei is selling into that shortfall.
The marker to watch is the commercial calendar rather than the forecast itself. Huawei's AI cluster service enters markets outside China on 30 November and the enterprise agent platform on 30 December. If both ship on schedule and attract buyers in regions short of compute, the agent-centric infrastructure thesis moves from report to procurement, and the token figures become a planning assumption rather than a vendor claim. If adoption stalls, the ten directions remain a roadmap for hardware Huawei still needs to place.
