China Telecom Research Institute: AI's centre of gravity shifts from models to agents, inference to hit 80% of compute by 2029
On September 12, 2026, CCTV reported on the China Telecom Research Institute's AI Infrastructure Development in the Agent Era report (2026). Its central conclusion: the centre of gravity of China's AI industry is shifting away from competition over large models and computing power towards scaled deployment and monetisation of AI agents. Rao Shaoyang, head of the institute's industry and enterprise strategy research department, provided a set of concrete numbers: China's annual token consumption is expected to reach 10^17 in 2026 and exceed 3.5 x 10^19 by 2030, a compound annual growth rate of nearly 12x; on the compute side, the agent boom will drive sustained high growth, with national computing demand expected to rise nearly tenfold per year over the next two to three years; and structurally, inference demand will far exceed training, with inference projected to take 80% of China's computing-power market by 2029. Bloomberg and The Next Web followed the same day.
Start with the three hardest numbers in the report. First, token consumption: an expected 10^17 in 2026, rising past 3.5 x 10^19 by 2030, with a compound annual growth rate of nearly 12x — not 12 percent growth, but roughly multiplying by twelve each year. Second, computing demand: nearly tenfold growth per year on average over the next two to three years, again an order-of-magnitude figure rather than a percentage. Third, a structural shift: inference computing reaching 80% of the market by 2029, pushing training into a secondary position. Read together, they point to a plain conclusion — if the investment logic of the past two years was who has the biggest model and the most cards, the logic ahead becomes who runs inference most cheaply and whose agents actually do work for customers and bring the bill back.
Why does the training-to-inference shift matter more than the headline growth? Because the two have completely different business models. Training is one-off capital expenditure: buy the cards, build the cluster, finish the run, amortise over years. Inference is operating expenditure that happens again and again — every user call burns compute. The Next Web put it well: training builds a model once, while inference is what it costs every time somebody uses one, which makes it an operating expense rather than a capital one. For enterprises, that means AI goes from being a big project to being a recurring bill. It also explains the report's emphasis on monetisation: inference cost rises linearly or worse with usage, and if the revenue side cannot absorb it, bigger scale simply means bigger losses.
The report also places this in the frame of investment scale. According to figures cited by The Next Web, Chinese technology companies are expected to spend close to 600 billion yuan (about $89 billion) on AI this year, more than a tenth of all investment in the country. Policy is moving in step: the South China Morning Post reported on September 8 that the Ministry of Industry and Information Technology, in a five-year industry plan released on Monday, set a target of 9,800 EFLOPS of intelligent computing capacity by 2030 and called for 3.8 trillion yuan (about $532 billion) in cumulative information infrastructure investment over 2026-2030. The plan explicitly calls for orderly deployment of 10,000-card and even 100,000-card intelligent computing clusters alongside inference facilities tailored to different applications, and for greater adaptation to home-grown chips. The whole arrangement extends the East Data, West Computing project launched in 2022, which shifts power-hungry computing workloads from densely populated eastern regions to cheaper land and energy in the west, and now centres on eight national computing hubs and ten data-centre clusters.
Set this report beside two other stories from the same week and a global coincidence appears. In the United States, OpenAI paused new sign-ups for its $200-per-month Pro tier because of Astra demand, with management acknowledging that tier puts the most strain on its systems; the same week Altman said he was open to signing a slowdown pact with other labs. In China, the report lands on scaled agent deployment and monetisation, whose precondition is whether inference compute can keep up. Both sides are really facing the same question: the cost structure and business model of the agent era are being redefined. They differ only in response — one side debating whether to slow down deliberately, the other planning to quadruple computing capacity over five years and roll out inference facilities. For anyone building products, both paths will show up in future model pricing and supply stability, and that macro line deserves a place in vendor selection.
🤔 Frequently Asked Questions
Q1: How big is 10^17 tokens, really?
10^17 means a one followed by seventeen zeros. That is the China Telecom Research Institute's forecast for China's total token consumption in 2026; the same report expects it to exceed 3.5 x 10^19 by 2030, with a compound annual growth rate of nearly 12x. To get a feel for the magnitude, note that token consumption is essentially three variables multiplied together: how many people use it, how long they use it, and how long each task is. Agents push the third variable to its extreme — a single task can involve dozens or hundreds of model calls, so consumption multiplies accordingly.
Q2: Why will inference overtake training?
Because the two happen at completely different frequencies. Training a model is one-off; the cluster finishes a run and moves to the next. Inference happens once per request, and the more users and the longer the tasks, the more calls there are — growing linearly or faster. The report expects inference to reach 80% of China's computing-power market by 2029, with training secondary. The commercial meaning of that shift: AI spending moves from one-off investment to ongoing operating cost, and attention moves from how many cards to buy to how much each call costs and how to push that down.
Q3: Where do the report's numbers come from?
The report is titled AI Infrastructure Development in the Agent Era (2026), published by the China Telecom Research Institute and carried by CCTV on September 12, 2026, with the specific figures provided by Rao Shaoyang, head of the institute's industry and enterprise strategy department. Chinese outlets including Cailianshe and Sina Finance republished it that day, and Bloomberg and The Next Web covered it in English. One distinction matters: this is an industry research report and forecast from a research arm of China Telecom, a state-owned carrier. It reflects institutional views and forward projections rather than realised statistics — every figure for 2026, 2029 and 2030 is a forecast.
Q4: What does this mean for developers and independent builders?
Three. First, inference cost becomes a core competitive factor rather than a rounding error — the 80% figure means industry resources will tilt toward inference efficiency, and engineering levers like unit pricing, caching and batching keep gaining value. Second, agent billing models are still immature: charging per task while making dozens of model calls behind it can eat the margin, so count the real number of calls per task at design time. Third, home-grown compute adaptation is now written into policy planning, so if you build for the domestic market, replaceability of models and inference infrastructure matters more and more — do not lock your architecture to a single vendor.
🛠️ Recommended Tools
- Unit Converter - Numbers like 10^17, 3.5 x 10^19 and 9,800 EFLOPS are easy to blur together; convert to the same magnitude before comparing
- Timestamp Converter - The report's 2026, 2029 and 2030 are forecast milestones; converting them into your own product windows beats daydreaming
- Regex Tester - For token counting and log parsing, test your regex first before wiring it into a pipeline — it saves a lot of wasted inference
The most valuable thing about this report is not the big numbers but the fact that it puts monetisation and inference computing side by side in its core logic. For two years, AI discussion in China has focused almost entirely on model parameters and card counts; now a research institute is emphasising that inference will reach 80% and that the industry's centre of gravity is shifting to deployment and monetisation, which means the yardstick has changed — whether it runs matters less than whether the maths works. The line I would keep is the compound annual growth rate of nearly 12x: it means token consumption next year will be an order of magnitude larger than this year. For product teams that multiplier is both an opportunity and a pressure, because cost grows with it — unless the architecture was built for inference efficiency in advance.
Summary
On September 12, 2026, CCTV reported that the China Telecom Research Institute published its AI Infrastructure Development in the Agent Era report (2026), concluding that China's AI industry is shifting from competition over large models and computing power towards scaled deployment and monetisation of AI agents. Rao Shaoyang, head of the institute's industry and enterprise strategy department, gave these figures: annual token consumption is expected to reach 10^17 in 2026 and exceed 3.5 x 10^19 by 2030, a compound annual growth rate of nearly 12x; national computing demand is expected to grow nearly tenfold per year over the next two to three years; and inference is projected to take 80% of the computing-power market by 2029, far exceeding training. Bloomberg reported the same day that agents will drive close to tenfold annual growth in China's computing demand over two to three years, with inference at 80% by 2029; per figures cited by The Next Web, Chinese tech firms are expected to spend close to 600 billion yuan (about $89 billion) on AI this year, more than a tenth of all investment in the country. On policy, the South China Morning Post reported on September 8 that MIIT's five-year plan targets 9,800 EFLOPS of intelligent computing capacity by 2030 and calls for 3.8 trillion yuan (about $532 billion) in cumulative information infrastructure investment over 2026-2030, with orderly deployment of 10,000-card and 100,000-card clusters plus inference facilities. Primary sources: CCTV (via Cailianshe and Sina Finance), Bloomberg, The Next Web, South China Morning Post.
Sources: 财联社（转央视新闻） · 新浪财经 · Bloomberg · The Next Web · South China Morning Post