China expects to burn 100 quadrillion tokens this year
Two demand numbers landed today that say where AI compute is actually going, and a law takes effect tomorrow that decides who gets to keep the chip know-how behind it.
China's annual token consumption is expected to hit 100 quadrillion this year and exceed 35 quintillion by 2030, according to a report presented by China Telecom's research institute at the 2026 China Computing Conference. Rao Shaoyang, who leads industry and enterprise strategy at the institute, paired the token curve with a claim about what it is made of: Chinese compute demand growing close to 10x a year for the next two to three years, and inference taking roughly 80% of the country's compute market by 2029. The supporting figures are as pointed as the headline number — China's biggest tech companies spending close to RMB 600 billion on AI capital expenditure in 2026, more than 10% of the national total, and a market-research forecast cited in the report of about 80 million active agents globally this year rising to 2.2 billion by 2030. The logic underneath is the part worth stealing: agents consume tokens continuously while they work, so the token count stops tracking how many people asked a question and starts tracking how much work got delegated. Independent audit of the methodology is not public, which is the caveat to carry — but an inference-heavy, agent-driven demand mix is the same conclusion every chip vendor's own filings keep reaching.
South Korea's rewritten espionage law takes effect September 13, the statute's first broadening in more than seven decades, and it points directly at the semiconductor and AI process know-how foreign rivals keep hiring away. Article 98-2 stops defining espionage as an act committed for North Korea alone and extends it to any foreign country or "equivalent organization," with a three-year mandatory minimum. What lit the fuse is the CXMT case: a former Samsung engineer who transcribed more than 600 sequential steps of DRAM process technology, helping China reach its first 10nm-class DRAM, was sentenced to seven years — under trade-secret law, because the espionage statute would not reach him. A bill filed August 31 would name foreign corporations explicitly as covered beneficiaries, add designated "national core technologies" to protected subject matter, and raise the floor to ten years. Note the blast radius: the revised law does not distinguish adversaries from allies, so a transfer to a US, Japanese or European employer sits in the same legal frame — which is why Korean defence exporters say one loophole closed and another opened.
A 35B-parameter model post-trained on its own research trajectories lifted a Frontier Science score from 1.7 for the base model to 46.2 for a July build, according to Chen Sisheng, an AI professor at Shanghai Jiao Tong University. He described the loop at the "Agent Post-Training: the nature of intelligence and the next Scaling Law" forum at the Inclusion·Bund Conference on September 12: an agent explores a research environment, its trajectories become training data, and the stronger model runs the next round. Others on the same bill mapped the rest of the stack — Ant's Ling model lead Zhou Jun on building agents that hand off to a human when evidence, authority or risk crosses a line; an Nvidia researcher on a deliberately small PyTorch-native reinforcement-learning framework, chosen partly so coding agents can read all of it; Meituan LongCat's Gu Qi on harness design for long-horizon work. The shared caveat is the interesting bit: several speakers said the bottleneck is no longer producing results, it is verifying them.
What to watch: whether China's inference-share forecast shows up in carrier and cloud capex guidance next quarter.
Is total token consumption a useful measure of AI demand, or does it just reward verbose agents? Tell us in the comments.
