If AI use really grows 100,000-fold, who's paying for all that computing?
Keep up with IT Club
Add IT Club as a preferred source in Google Search.
Huawei’s Intelligent World 2035 report forecasts 100,000-fold growth in global annual token consumption by 2035, with traffic from agents accounting for over 90%. Businesses should treat both figures as Huawei’s forecast, then plan for the real costs behind prompts, automated workflows, infrastructure and energy.
The number, carefully stated
Huawei forecasts that global annual token consumption will grow 100,000-fold by 2035, with traffic from agents accounting for over 90%. That is a company forecast, not a prediction that has been independently proved.
- Metric: global annual token consumption.
- Date: by 2035.
- Agent share: over 90% of traffic, according to Huawei.
AI looks almost free on screen
Type a question, receive an answer and it feels effortless. Behind that interaction are processors, data centres, networking, cooling, electricity, storage and software infrastructure. A subscription may hide those costs from the person using the tool, but it does not make the work free.
A person might make 20 requests. An agent can make hundreds or thousands while researching, testing, checking, retrying, consulting other agents and using tools. That changes the volume of work even when each individual step feels small.
What Huawei is—and is not—saying
Huawei’s 16 September 2026 Intelligent World 2035: Turning Vision into Action announcement describes a move from an application-centric to an agent-centric digital world. It says every perception, reasoning, tool call and interaction can involve token generation and high-frequency token flow. Huawei then gives the 100,000-fold annual token-consumption forecast and says agent traffic will exceed 90%.
Keep this attributed: Huawei forecasts these figures. They are not guaranteed outcomes or independently validated measurements. Do not present them as current usage, and do not merge them with Huawei’s separate forecasts for computing capacity, storage, data traffic or agent numbers.
The questions businesses will need to ask
Today the question is often, “Can AI do this?” As usage becomes operational, the better question is, “Is this task worth the computing cost?” That cost may include an AI subscription, API consumption, duplicated agents, unnecessary retries, an oversized model, overly frequent runs, data transfer, storage and the infrastructure or energy supporting it.
- 1Define the business result before adding an agent or automated prompt chain.
- 2Set an owner, budget and stop condition for recurring workflows.
- 3Choose the smallest suitable model and sensible run frequency.
- 4Log calls, retries and tool use so waste and loops can be found.
- 5Review value created versus total resource cost before expanding the workflow.
This does not mean counting every token obsessively during early experiments. It means adding cost awareness when an experiment becomes a process that runs every day, touches customer data or can multiply its own activity.
Plain-English takeaway
AI is not magic. Somewhere, a computer is doing the work. Good AI management will mean controlling not only what agents can do, but also what is actually worth doing.
Sources and further reading
Huawei — Intelligent World 2035: Turning Vision into Action (primary announcement) →
Huawei — Intelligent World 2035: Turning Vision into Action (report PDF) →
The Standard — Huawei forecast coverage (Reuters-syndicated report) →
Plain-English Takeaway
AI is not weightless. Once an automated workflow becomes operational, measure the value it creates against the computing, subscription, API and infrastructure resources it consumes.
Frequently asked questions
What exactly did Huawei forecast?
Huawei’s 16 September 2026 announcement says that global annual token consumption will grow 100,000-fold by 2035, with traffic from agents accounting for over 90%. This is Huawei’s forecast, not an established fact.
Is the 100,000-fold figure computing capacity or data traffic?
Not in this forecast. The requested figure is global annual token consumption. Huawei publishes other, separate figures for computing capacity, data traffic and infrastructure; they should not be substituted for the token forecast.
What should an SME do now?
Start with the business outcome, set sensible usage and approval limits, monitor API or subscription costs, prevent duplicated or looping agents, and review whether an automated task is worth its total resource cost before scaling it.
Related Articles
AI is leaving the computer screen and entering the real world
Anthropic's Bay Area biology lab shows the next AI question is not only what a model can say, but how its work can be tested against the physical world.
Read articleCan AI Really Act as Your PA?
AI can increasingly help with email, calendars, meetings and routine administration. The sensible starting point is deciding what it should be allowed to see and do.
Read articlePilotless Aircraft Are Starting to Carry Cargo. What Happens to the Jobs?
A cargo aircraft without a pilot physically onboard shows why automation changes tasks and accountability before it necessarily removes an entire operation.
Read article