China Telecom Bets AI Agents Will Meter Compute

China Telecom Research Institute said on September 12 that inference will take 80 percent of the country’s compute market by 2029, as AI agents go commercial. The same report expects those agents to drive close to tenfold yearly growth in computing demand over the next two to three years.

The forecast did not come from a model lab. It came from the research arm of a state-owned carrier that already bills customers in tokens, the unit that meters each run of a model.

A State Carrier Wrote the Forecast

The paper is titled Research Report on the Development of AI Infrastructure in the Agent Era (2026). Rao Shaoyang, director of the Industry and Enterprise Strategy Institute at the China Telecom Research Institute, tied the demand spike to agents that hide the model and finish a whole task, which burns tokens at every planning step, tool call and retry.

He said China’s yearly token use will hit 100 quadrillion in 2026 and more than 35 quintillion by 2030, “with a compound annual growth rate approaching 12 times.” That line is a forecast from one institute, and the public write-ups do not show the counting rules behind the 12-times claim. The direction is still the point the carrier wants believed: the scarce resource is no longer a training run you do once, it is the bill for running the thing all day.

Zhiyong Huang, vice president of China Telecommunications Corporation, made the same shift explicit a day earlier at the AI for Good Global Summit. Industry attention, he said, is moving from model training to inference services, and agents are becoming a digital workforce. He also named the strains that come with that, including data security, weak guardrails, a wider AI divide and rising energy use.

AI capabilities should be made as universally accessible, affordable and usable as water and electricity.

Zhiyong Huang, Vice President, China Telecommunications Corporation, AI for Good Global Summit 2026

Water and electricity are metered. That is the business a carrier already knows how to run.

Tokens Are the New Meter

China Telecom has been packaging that meter since May, when it began selling token plans the way it once sold minutes and data. Consumer tiers start at 9.9 yuan a month for 10 million tokens, about $1.40, and rise to 49.9 yuan for 80 million. Developer and business packs run from 39.9 yuan to 299.9 yuan a month.

CHINA TELECOM TOKEN PLANS

The National Data Administration said average daily token use in China rose from about 100 billion in early 2024 to more than 140 trillion in March 2026. That is the measured load sitting under Rao’s yearly forecast, and it is why the three state carriers now talk about token factories the way they once talked about base stations.

China Telecom said intelligent computing revenue rose 95 percent in the first half of 2026, lifting a broader intelligent-business line to 31.1 billion yuan. In June it put a public face on the pitch, launching the XINGCHEN TokenHub token service platform and an AI Token Global Service Ecosystem Alliance at its Global Partner Conference in Shanghai. Liu Ying, executive vice president, told partners the company wants token operations to be the bond that holds a global computing network together.

Chinese technology companies are expected to spend close to 600 billion yuan (about $89 billion) on AI this year, a sum the institute put at more than a tenth of all investment in the country. The National Bureau of Statistics said internet companies spent 81.8 percent more on equipment in the first seven months of 2026 than a year earlier. The cash is still going into halls and chips. The report’s claim is that more of what those halls do will be inference, billed by the token.

How China Telecom Already Sells Agents

Huang did not speak only in forecasts. He said the company has built more than 110 industrial models and more than 350 intelligent agents, serving 37,000 enterprises, and that its global cloud already connects to over 300 mainstream large models through more than 230 nodes.

AGENTS ALREADY ON THE NETWORK

• Farm pests: An agricultural model for fruit and vegetable growers cut crop losses by more than 30 percent, Huang said, while lifting farm incomes.
• Factory power: An energy agent for a railway vehicle maker forecasts supply a day or a week ahead and improved energy efficiency by about 8 percent.
• Home watch: Smart screens sold to households take dialect voice commands and alert carers if an older person living alone falls or stays still too long.
• Carrier staff: 981 digital employees across 23 job types raised network service-handling efficiency by 30 percent and cut fault-processing time by 14 percent year on year.
• Office agent: TeleAgent, built on the Xingchen model, is sold as a one-click path into those models for office work.

China Mobile and China Unicom have followed with their own token packs and, in Unicom’s case, a token supermarket. The overlooked party in a story about models is the phone company, because the product that scales is not a new checkpoint, it is a billable unit riding a network the state already owns.

Huang put numbers on that network: optical transmission covering 97 percent of Chinese cities, about 200 million fibre broadband users, and more than 1.54 million 5G base stations co-built with China Unicom. He cited 1,500 megawatt-scale data centres with rack power above 3.2 gigawatts, and said total computing power across China Telecom’s own network had reached 91 EFlops, a company figure, not the national stock.

The energy side is already a design problem. He described an underwater data centre 35 metres down in the East China Sea, tied to offshore wind, cooled by seawater, with a green-energy supply rate above 95 percent and electricity use down 22.8 percent, and a cave site in Gansu running on hydropower, wind and solar. If agents become a digital workforce, the power bill becomes a political one.

Inference Hardware Is the Bottleneck Now

Long agent sessions break a different piece of the stack than training does. The limit is often the KV cache that holds context, not the next training cluster. At MWC Shanghai on June 24, China Mobile’s Hubei unit and Huawei said a live-network test of Huawei’s AI Inference Acceleration Solution, using OceanStor A800 storage, an Ascend A3 SuperPoD and Unified Cache Manager, produced a 372 percent rise in token throughput on long sequences.

LIVE-NETWORK INFERENCE TEST IN HUBEI

On GLM-5.1, tokens per second rose 313 percent at a 64K sequence length before that 372 percent jump at 128K. A China Mobile Hubei representative said the province sits about 10 milliseconds from the country’s eight major computing-power hubs, and that the same kit can lift throughput by over 50 percent on agent chat and code generation. Michael Qiu, president of Huawei’s global data storage marketing and solution sales, said carriers launching token packages show that large-scale agent use “has clearly entered a new phase.”

The same pattern is showing up on the global wire. Jeetu Patel, Cisco’s president and chief product officer, told a Goldman Sachs conference that agentic AI still has adoption below 2 percent in enterprise and consumer markets, yet agents already consume about 60 percent of global inference capacity, with token use up 14 times since February. He put the extra load on the network at about 450 percent more bandwidth than a person doing the same task.

These agents tend to be far more consumptive on network bandwidth than humans. About 450% more network bandwidth gets consumed by an agent to conduct the same task that a human would conduct.

Jeetu Patel, President and Chief Product Officer, Cisco, Goldman Sachs Communacopia conference

A product that almost nobody has fully adopted can still dominate the compute bill. That is the second bill hiding behind the model race.

Beijing’s Compute Target Assumes the Same Shift

The Ministry of Industry and Information Technology published its 15th Five-Year Plan for the information and communications industry on September 7, with the document dated August 12. It sets a target of 9,800 eflops of intelligent computing by 2030. China’s intelligent computing stood at 2,185 eflops at the end of June, up 177 percent from a year earlier, and the country had 52 facilities with more than 10,000 accelerator cards each. Hitting 9,800 eflops means more than four times the June stock.

The plan also calls for 3.8 trillion yuan in cumulative information infrastructure investment from 2026 to 2030, and for industry revenue of 4.1 trillion yuan by 2030. It wants orderly rollout of 10,000-card and 100,000-card clusters, more work to fit domestic chips, and inference computing facilities tailored to different applications. Beijing is not only counting total flops. It is already drawing a separate line for the machines that run models rather than train them.

Huang’s 91 EFlops is what one carrier says it holds. The 2,185 eflops figure is the national intelligent-computing stock MIIT published for the end of June. They are different piles, and both can grow at once if the token forecast is even half right.

Seven Factories Face a November Deadline

Europe is trying to buy the same kind of capacity on a slower clock and a thinner cheque. On July 30 the Commission opened a call for up to seven AI Gigafactories, backed by up to €10 billion in EU and national funds and meant to unlock at least €20 billion in private money, a package it put at more than €30 billion. The plants are supposed to serve training, inference and fine-tuning, and to sit alongside an existing network of 19 AI Factories.

THE EUROPEAN GIGAFACTORY CLOCK

1. July 30, 2026: The Commission opens the tender for up to seven AI gigafactories.
2. November 12, 2026: Bidding closes.
3. Early 2027: Award decisions are expected from the EuroHPC Joint Undertaking.
4. 2027: Construction is due to start after contracts are signed.
5. Within 18 months of signature: Selected plants are expected to begin operations, which would put the first machines in 2028 if the awards land on time.

Eighteen member states have signed a joint procurement agreement. An earlier informal call drew 77 proposals covering 60 sites in 16 member states. The live tender splits into two lots: up to four projects that can receive €100 million in EU money in phase one and €400 million in phase two, and up to three larger ones that can receive €200 million then €800 million. The Commission also signed letters of intent with AMD, NVIDIA and Qualcomm so consortia can buy the chips.

That is a plan for buildings, still out to bid, with the private half hoped for rather than booked. Chinese technology companies, on the institute’s 600 billion yuan line, are spending on AI this year at a pace that dwarfs one European public envelope. The European sites may well run inference when they exist. They do not exist yet.

Open Weights Do Not Cancel the Inference Bill

Chinese labs have spent the past year putting strong models into the world as open weights, which looks like the opposite of a metered service. It is not. A free checkpoint still has to run somewhere, and the token-hungry ones make that somewhere expensive. Shipping weights can be a way to win users when you cannot yet sell a polished inference product at global scale. Selling tokens at home is how a carrier turns the same demand into rent.

The institute’s 12-times growth line may or may not hold. What is already measurable is the daily token count, the token price list, the Hubei throughput test and a national plan that funds inference sites by name. Huang’s water-and-electricity line is the strategy in one sentence: make the model invisible, then charge for the current.

European bids close on November 12, 2026. By then China Telecom will still be selling 9.9 yuan token packs, and the report on the carrier’s desk will still say inference takes the market by 2029.