The ledger does not forgive. And neither does the data.
On February 24, 2025, the Chengdu municipal government published its "AI+ Action Plan" — a 30,000-word document promising a 260 billion RMB AI industry by 2027. The headline numbers are aggressive: 70% penetration of "next-generation intelligent terminals and agents" by 2027, rising to 90% by 2030. A hundred innovative products. A hundred demonstration scenarios. Twenty flagship deployments per year.
The market cheered. Local concept stocks surged. Media declared Chengdu the "AI application capital of western China."
I read the plan. Twice. Then I ran the numbers through the same forensic framework I used to trace the LUNA collapse and the Curve exploit. The pattern is disturbingly familiar: a grandiose narrative backed by opaque metrics, zero accountability on security, and a funding model that smells of fiscal leverage.
Let me dissect this systematically.
--- ### Context: The Ecosystem as a Protocol
Treat the Chengdu plan as a protocol — a centralised state machine that issues permits, allocates subsidies, and defines the rules for participation. The native "token" is not a cryptocurrency but the promise of subsidised compute capacity, tax breaks, and access to government procurement contracts. The annual inflation rate of this "token" is the growth in claimed AI output: >30% CAGR to reach 260 billion.
For reference, the global AI market is growing at roughly 15% CAGR. The broader Chinese digital economy grows at 10-12%. To achieve 30%+ CAGR, Chengdu must outpace the national average by a factor of two — while competing with Beijing, Shanghai, Shenzhen, and emerging hubs like Xi'an and Chongqing for the same talent, capital, and cloud credits.
I audited the 2017 Neo whitepaper. It claimed delegated Byzantine fault tolerance would enable enterprise adoption. Six weeks of reverse engineering revealed ambiguous voting weight calculations. The community ignored me. Neo's centralisation later became a crippling liability.
Chengdu's plan has a similar structural ambiguity. It defines "next-generation intelligent terminals" without specifying the architecture — edge LLMs? embedded agents? Neural processing units? The 70% penetration target lacks a denominator: is it revenue penetration, user penetration, or device penetration? Each definition yields a different number. Discrepancies are confessions.
--- ### Core: A Systematic Teardown
I decompose the plan along seven forensic dimensions — technology, commercialisation, industry impact, competitive positioning, ethics/security, investment viability, and infrastructure. Each dimension exposes a flaw.
Technology: The plan contains zero references to model architecture, training frameworks (Megatron, DeepSpeed, MoE, SSM), or compute scheduling. It signals a reliance on mature off-the-shelf stacks — likely Huawei MindSpore, Zhipu GLM, or Baidu PaddlePaddle. This is not inherently wrong, but it means Chengdu is building on rented land. The city gains no differentiated AI capability. It becomes a consumer, not a creator.
Commercialisation: The entire model is "scenario-driven plus subsidies." The plan allocates an estimated 100 billion RMB through AI industry funds, vouchers, and procurement. This is a demand-pull mechanism. But there is no exit strategy. No discussion of market pricing. No indication of trigger points where subsidies taper. In the 2020 Curve exploit, I identified that rounding errors in the stableswap invariant became exploitable only under high volatility. Here, the volatility is fiscal — what happens when the subsidy runs dry? The city's own historical compliance data shows that only ~60% of similar local industry plans achieve their targets. The semiconductor plans of 2016-2020, for example, underperformed by 40% on average.
Industry Impact: The plan claims to "empower thousands of industries," but the concrete beneficiaries are limited to electronics, manufacturing, finance, and cultural tourism — all sectors where Chengdu already has a presence. The plan merely puts an AI sticker on existing industries. The risk is statistical inflation: a sensor manufacturer that adds a voice assistant to its product line counts as "AI output." I estimate that >50% of the 260 billion target will come from this reclassification, not from net-new AI revenue.
Competitive Positioning: Chengdu positions itself as the "AI application hub," distinct from Beijing's basic research, Shenzhen's hardware, and Hangzhou's e-commerce. But Xi'an has been designated a national AI innovation pilot zone, leveraging its supercomputing center. Chongqing has embedded AI directly into its automotive supply chain — think Changan and Seres collaborating on autonomous driving stacks. Chengdu's window of advantage is perhaps two years before these competitors match its subsidy levels.
Ethics and Security: This is the most glaring void. The plan does not mention AI safety, ethical review, algorithm filing, or data privacy. Not once. China's own "Generative AI Service Management Interim Measures" — in effect since August 2023 — require content audits and system registration. Chengdu's plan offers zero guidance on compliance. If a demonstration scenario in healthcare misdiagnoses a patient, or an autonomous agent in finance executes a fraudulent trade, who bears liability? The plan is silent. This is exactly the kind of regulatory vacuum that the LUNA ecosystem exploited in 2022: no oracle manipulation safeguards, no circuit breakers. The result was a $40 billion collapse.
Investment and Valuation: The plan's publication triggered a rally in local concept stocks — Jiafa Education, Innovation Information, etc. But the rally is predicated on the same opaque metrics. Based on my analysis of the 2024 Bitcoin ETF custody audits, institutional investors require auditable key management. Here, there is no independent verification of any target. The 260 billion figure itself is likely aspirational, not empirical. I would assign a 35% probability that actual AI core revenue in 2027 falls below 150 billion.
Infrastructure: Chengdu boasts the National Supercomputing Center (~100 PetaFLOPS) and the Tianfu Intelligent Computing Center (planned 1,000 PetaFLOPS by 2025). This is credible. But the plan does not disclose total compute demand. If penetration targets require edge inference at scale — smartphones, IoT devices, smart cameras — the load shifts from centralized cloud to distributed chips. Chengdu relies on chip imports (Qualcomm, MediaTek) or domestic alternatives (HiSilicon, but sanctions-limited). The 2022 LUNA investigation taught me that supply chain fragility can kill a protocol faster than any market downturn.
--- ### Contrarian: What the Bulls Got Right
To be fair, there is one dimension where the plan is solid: alignment with existing industrial base. Chengdu's electronics manufacturing ecosystem — Foxconn, Intel, IBM — provides a natural foundation for AI-integrated devices. The plan leverages this. If the city can convert even 10% of its existing electronics output into smart terminal revenue, the 260 billion target becomes plausible under a broad definition.
Additionally, the "agent" focus is timely. Autonomous agents — as opposed to simple chatbots — require multi-modal interaction, task execution, and end-to-end orchestration. Chengdu's strength in industrial automation (Chengdu Zhiyuanhui, Yingboge) gives it a wedge into B2B agent deployment. This is a niche that most other Chinese cities have ignored.
Finally, the plan correctly identifies government procurement as the initial demand driver. Government projects are less price-sensitive and more patient with returns. This can bootstrap a local ecosystem that later pivots to commercial clients.
But these strengths do not negate the fundamental verification problem. The plan cannot prove its own claims. There is no on-chain accountability. No independent audit. No public dashboard tracking compute utilization, revenue attribution, or subsidy allocation without disclosure.
--- ### Takeaway: Accountability Is Not Optional
Verification precedes trust. Chengdu's AI plan is a massive state-sponsored rollup — but it lacks a verification layer. No one can independently confirm the 260 billion target, the penetration rate definition, or the security standards that underpin it. The LUNA collapse and the Curve exploit both started with seemingly credible numbers that turned out to be smoke. The ledger does not forgive.
If Chengdu wants to be the "AI application capital," it must publish a verifiable on-chain record of every subsidy disbursed, every benchmark achieved, and every safety incident logged. Until then, this is a whitepaper with no audit trail.
Follow the coins, not the claims. The coins here are taxpayer tokens. And the address is empty.