The AI+ campaign: who pushes, who pays, where 安全 fits

So what · judgment

This is the top-level party-state mandate and the spine of the AI+ agenda; every ministry and provincial document in this chain cites it. It frames AI primarily as a growth/productivity instrument ('新质生产力'), with safety subordinated to and instrumental for deployment.

So what · judgment

Signals China wants to set international AI-safety governance norms via the UN channel and its 全球人工智能治理倡议, positioning itself as a norm entrepreneur.

So what · judgment

Shows how the abstract AI+ opinion cascades into concrete industrial mandates with quantified targets and an explicit anti-'involution' clause — a lever for tracing central-local coordination and duplication controls.

So what · judgment

The GBA is the flagship cross-jurisdiction integration case where 'one country, two systems' meets AI safety governance — the joint safety lab is a concrete cross-border evaluation institution directly tied to AI+.

So what · judgment

Included under an explicit scope rule. Represents Beijing's intent to fold Taiwan-facing integration (incl. digital/AI services to Taiwanese residents/firms) into the AI+ economic frame; should be read as PRC policy toward Taiwan, not cross-strait consensus.

So what · judgment

This is the central non-binding-but-authoritative safety-governance backbone the AI+ agenda operates within; TC260's downstream evaluation standards (below) implement it. It is the reference layer the rest of this briefing sits inside.

So what · judgment

Identifies the operational national evaluation (测评) apparatus — CAICT/AIIA benchmarks plus TC260 standards — that the AI+ opinion's 'model capability evaluation system' (建立健全模型能力评估体系) relies on. These are the key institutional nodes in the safety-evaluation apparatus.

So what · judgment

The designation system is the mechanism through which provinces compete for AI status and resources: the structural driver of the local-central race and its duplication. Note designations froze in 2024 (no new tickets that year).

So what · judgment

Central SOEs are cast as demand-aggregators and compute suppliers, not primarily as frontier-model developers — the state channels their scale ('需求规模大、产业配套全、应用场景多') to pull the private AI sector. Key structural fact for mapping who deploys vs who innovates.

So what · judgment

Confirms that telecoms (Mobile/Telecom/Unicom) and grid/energy majors are the operational backbone of state compute and vertical models — the deployers the AI+ agenda leans on, distinct from private frontier labs.

So what · judgment

The single largest national financial instrument explicitly tied to AI+; its LP being the semiconductor 'big fund III' signals hardware/compute-autonomy as the strategic core. It anchors the question of who finances the agenda.

So what · judgment

Illustrates the state-directed credit channel wrapping around the fiscal 算力券 subsidies — banks operationalize AI+ by de-risking compute purchases, coupling monetary and industrial policy.

So what · judgment

The template subnational statute other cities followed; shows Shanghai institutionalized ethics/risk-tiering years before the 2025 central opinion, giving the Atlas a baseline for local safety-institution maturity.

So what · judgment

Concrete demonstration of the compute-voucher (算力券) fiscal instrument the Atlas should catalog as the dominant local subsidy mechanism, now layered with national vouchers.

So what · judgment

Beijing was first-mover on both a local 'AI+' brand and compute vouchers; its 2026 plan explicitly carves out a safety-governance action pillar, the clearest subnational example of safety institutionalized as a named workstream.

So what · judgment

Zhejiang/Hangzhou is the DeepSeek/宇树 heartland; its voucher design explicitly privileges domestic ('国产') compute at a higher subsidy rate — an important signal for the Atlas that fiscal instruments are being used to steer toward compute self-reliance.

So what · judgment

Jiangsu shows the manufacturing-heavy province operationalizing AI+ via mandatory enterprise diagnostics and standards — a compliance-driven variant. The mandatory self-assessment system is a de-facto evaluation instrument for the Atlas.

So what · judgment

Sichuan is a rare subnational case funding actual AI safety-testing R&D (model safety-assessment standards, a system-safety test platform, and AIGC-detection) — the strongest local safety-evaluation effort in this source set outside the tier-1 cities.

So what · judgment

Shows Guangdong wiring AI+ into export/outbound ('算力出海、数据跨境') and cross-border finance — the internationalization face of the domestic agenda, and evidence of local governments competing to be the GBA compute hub.

So what · judgment

Direct evidence of provinces over-building compute for subsidy/prestige reasons, prompting central corrective pressure (energy-quota controls, anti-'内卷' clauses in the AI+制造 opinion): the core local-central friction in the rollout. The '30%' figure is an industry-research estimate, not an official statistic.

So what · judgment

Shenzhen created a government-run AI testing/evaluation (评测) capacity in law — an early subnational safety-evaluation institution the Atlas should log; also pioneers the risk-tiered 'sandbox'-style approach later echoed nationally.

So what · judgment

Shenzhen extends the voucher model to a distinct 'training-compute' (训力) instrument and couples it with a regulatory sandbox — a concrete example of local experimentation that may diverge from central pacing.

So what · judgment

Confirms a division of labor: private labs supply frontier models/innovation while SOEs supply scale/compute/scenarios and the state supplies capital and safety rules. The '智能原生企业' concept is the policy's core private-sector construct.

So what · judgment

A verified negative finding: the AI+ agenda does NOT itself court foreign AI investment. Foreign firms enter AI-adjacent markets (compute/IDC/cloud R&D) only through parallel telecom-opening and FDI-R&D tracks, under national-treatment ('内外资一致') principles. Reading the AI+ opinion as courting foreign AI investment would overstate its openness.