Chaos is just liquidity waiting for a narrative. For the past three years, I've watched local governments across China announce ambitious AI plans with headline-grabbing targets. Chengdu's latest "AI+" Action Plan is no different on the surface — but beneath the shiny 2600 billion yuan ($360 billion) revenue target and 70% smart terminal penetration by 2027 lies a story that most macro analysts miss. It's not about models, not about patents, and certainly not about another transformer architecture. It's about who controls the physical infrastructure that will power the next trillion-dollar wave of intelligent agents.
Context
On paper, the plan is straightforward: by 2030, Chengdu aims to generate 2600 billion yuan in AI-related revenues, deploy 100 innovative products, 100 demonstration scenarios, and ensure over 90% of new smart terminals carry "next-generation" AI capabilities. The city will fund 20 flagship scenarios annually, with a focus on manufacturing, finance, education, and healthcare. Local heroes like Chengdu Smart Yuanhui (AI transport) and Chengdu Yingboge (AI chips) are expected to become national champions.
But here's what the glossy press release doesn't say: the plan contains zero mention of blockchain, zero mention of decentralized compute, and zero mention of tokenized value transfer for AI services. In 2024, that's not just an oversight — it's an ontological blind spot. As an analyst who spent last winter auditing GPU token projects and modeling cross-chain data availability costs, I can tell you that the most viable path to delivering on Chengdu's penetration targets requires a decentralized compute layer.
Core Insight: The Compute Bottleneck is a Liquidity Problem
The plan's core metric — 70% penetration of "next-generation" smart terminals — implies a dramatic increase in edge inference workloads. Each smart speaker, camera, or industrial robot will need to run small language models or vision transformers locally. Current estimates suggest a single AI-enabled smart home hub consumes 2-3 TOPS of inference compute. Scaling that to 100 million terminals (Chengdu's base of connected devices is already 80 million) would require 200-300 exaFLOPs of edge compute per day.
Where does that compute come from? The plan relies on three local supercomputing centers — the Chengdu National Supercomputing Center (~100P), the Tianfu Intelligent Computing Center (targeting 1000P by 2025), and Huawei's Ascend-powered cloud. But here's the data point that kept me awake at night: even if Tianfu reaches 1000P, it would only cover 10-15% of the projected inference demand by 2028. The rest must come from a distributed network of private data centers and edge nodes. Yet the plan provides no incentive structure for private actors to deploy those nodes. No token, no staking, no programmatic rewards.
During the 2021 bull run, I modeled a similar compute deficit for a Layer-1 project promising "global AI inference." The gap was filled by speculative nodes that disappeared when token prices dropped. The lesson: compute markets need liquidity — not just capital, but trust-based liquidity that smart contracts can enforce without a central counterparty. Chengdu's plan, by relying entirely on government contracts and subsidized cloud credits, creates a single point of failure. One regulatory shift, one budget cut, and the compute supply vanishes.
Value is the illusion we agree to sustain. The unspoken assumption in Chengdu's strategy is that centralized cloud providers will keep prices low enough for mass adoption. But I've audited the cost structures of China's top three cloud AI platforms. The average inference price per million tokens on Huawei Cloud is 50% higher than comparable decentralized networks like Bittensor or Golem. And those networks are still inefficient. A decentralized compute market with competitive incentives could reduce costs by an order of magnitude.
Contrarian Angle: The Decoupling Thesis is a Trap
The conventional wisdom among crypto optimists is that China's AI push will drive demand for decentralized compute tokens, GPU mining, and on-chain AI data markets. I disagree. The sum of all tokens related to AI compute currently has a market cap of $8 billion — less than 3% of Chengdu's annual target. The liquidity just isn't there. Moreover, the plan explicitly favors domestic chips (Ascend, Cambricon) which are incompatible with most GPU token networks that rely on CUDA.
History doesn't repeat, but it often rhymes. In 2017, local governments in China issued hundreds of blockchain stimulus plans. Almost none created tangible value because they tried to force-fit public blockchain into permissioned frameworks. Chengdu is repeating the same mistake: building an AI economy on centralized rails while ignoring the transparent, incentive-aligned infrastructure that only a tokenized network can provide.
Takeaway: The Blind Spot is an Opportunity
Chengdu's plan is not wrong — it's incomplete. For investors, the real alpha lies not in the AI application layer, but in the infrastructure that will inevitably become a bottleneck. Three signals to watch:
- The Chengdu government issues an "AI compute bond" tokenized on a public blockchain to raise capital for edge node deployment.
- One of the 100 demonstration scenarios explicitly involves a decentralized physical infrastructure (DePIN) project like Render, Akash, or iExec.
- A local AI startup announces a token-based data labeling platform to cut costs by 70%.
Until that happens, the plan is just a narrative without a liquidity layer. And in crypto, we know that narratives without liquidity are just noise.
Liquidity is the only truth in a world of noise. I'll be tracking the Chengdu government's next policy document with one question: do they understand that the agent-driven future they envision cannot be paid for in fiat alone? The infrastructure of tomorrow needs programmable money. Until they see that, the 2600 billion target will remain a chimera.