--- title: "China's 'AI Plus' Is Not a Slogan. Here's What It Actually Changes" date: 2026-10-19 category: China AI Decoded site: NeuroAI canonical: https://neuroai.site/a/na-decode-ai-plus-what-changes language: en --- # China's 'AI Plus' Is Not a Slogan. Here's What It Actually Changes > AI Plus is a procurement and adoption programme with dated usage targets, so the numbers that matter are tokens and terminals rather than benchmarks. The saleswoman in the Foshan mall holds the handset at arm's length and speaks a sentence into it. The screen writes a meeting note, then a reply to a supplier, then a shopping list. The customer is a workshop manager replacing a three-year-old phone. He does not ask which model is inside. He asks whether the trade-in discount covers his old device, and whether the note-taking survives his accent. That second question is where Chinese industrial policy now lives. ## Key takeaways - The State Council's AI Plus (人工智能+) opinions target **>70%** adoption of new-generation intelligent terminals and agents by 2027, and >90% by 2030 (document, cited by Guangming Daily). - The 2026 Government Work Report added "smart economy" (智能经济) as a category and put "computing–power coordination" (算电协同) into the report for the first time (SCIO briefing, 20 March 2026). - CAICT puts the AI industry above RMB 1.2 trillion in 2025 (≈ US$169bn / HK$1.32tn), up 40% year on year, with the model layer up 189% (China Economic Net). - National token use averaged 140 trillion a day by March 2026, against roughly 100 billion in early 2024 (National Data Administration). - More than 6,600 AI companies operate in China, about 15% of the world total, and applications take 55% of industry revenue (CAICT, cited by China Economic Net). ## The instrument is a purchase order The programme is not mainly funding research. It is funding installation, then counting what was installed. The State Council document names six priority fields and two dated numbers: more than 70% adoption of new-generation intelligent terminals and agents by 2027, and more than 90% by 2030 (document, cited by Guangming Daily, March 2026). Neither number measures model quality. Both measure installed base. Three instruments do most of the work: - **Terminal trade-ins.** Consumer subsidy schemes lower the effective price of a handset, a kiosk or an in-car system, so a buyer upgrades before the old device dies. - **Public-cloud model vouchers.** Vouchers reduce the cost of a first purchase of model services, which moves a small firm's decision from building to renting. - **Agent promotion and local pilots.** City and provincial programmes fund a named deployment — a hospital queue, an inspection line — and publish the result as a case. Each instrument ends in a purchase order. None requires a new model. That is the design: the constraint is demand and integration labour, not research capacity. ## What the 70 per cent is counting A target needs a definition, and "intelligent terminal" was a marketing phrase. On 8 May 2026, MIIT, SAMR and MOFCOM published GB/Z 177-2026, a national standard series that grades terminal intelligence. Once a device carries a grade, a subsidy can attach to it and a buyer can defend the purchase in an audit. CAICT's figures explain the diagnosis. The AI industry passed RMB 1.2 trillion in 2025 (≈ US$169bn / HK$1.32tn), up 40% year on year, with the foundation layer up 59% and the model and framework layer up 189%. Applications were 55% of the total (China Economic Net, 13 October 2026). Supply is growing quickly; the plan is aimed at the other side of the ledger. ## Tokens are the closest thing to a meter By March 2026, national token consumption averaged 140 trillion a day, against roughly 100 billion in early 2024 — an increase of more than a thousandfold, and up more than 40% from 100 trillion at the end of 2025 (National Data Administration). ByteDance's Doubao went from about 100 billion tokens a day in May 2024 to more than 120 trillion by March 2026 (industry data, cited by CGTN). Those figures are useful because they measure consumption, not capability. They count work asked of a model. They do not say whether the work was worth doing, which is why they are a meter and not a verdict. ## Four ways a subsidy becomes a prop - A handset bought for the discount and never opened again. - A pilot that ends when the municipal grant ends, leaving a licence and no process. - An agent given permissions but no redesign of the workflow it was meant to replace. - A scoreboard that counts contracts signed rather than hours saved. None of these is exotic; all are standard failure modes of deployment subsidies. What makes the Chinese version testable is that the targets are dated and the counts are published, so failure eventually shows up as a missing number rather than an argument. ## Watch the meter, not the leaderboard Three indicators will settle whether the programme is working: - Token volumes from the National Data Administration, and whether growth comes from agents rather than chat. - Terminal grading registrations, which show whether the 2027 target is met by devices or by definitions. - Procurement documents, where the agent standard becomes a requirement rather than a suggestion. The 2026 Government Work Report put "smart economy" (智能经济) forward as a category and added "computing–power coordination" (算电协同) for the first time (SCIO briefing, 20 March 2026). Both phrases point the same way: demand and electricity are now planned together, while the model race is left to companies that have to sell something. ## Honest limitations The 70% and 90% targets reach English readers through a document quoted by Guangming Daily; the original definition of "adoption" is not public in translation, and it may count shipments rather than active use. Token figures are national averages from an agency that does not publish its method, and company token counts are self-reported. No official series shows how many installed assistants are opened monthly, or what share of subsidised devices are bought by businesses rather than households. The causal claim — that subsidies produce usage rather than purchases — is untested in any public evaluation. ## What readers can do now - **Operators:** pick one workflow, measure it for a month, then buy the terminal. A device that arrives before the process is defined becomes a subsidy receipt. - **Investors:** separate revenue that depends on public procurement from revenue that survives without it; the first is a policy cycle, the second is a business. - **Researchers:** request the procurement notices behind any adoption claim. The definitions in those documents are more informative than the headline percentage. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-decode-ai-plus-what-changes Free to quote with attribution and a link to the original.