An analytical essay by Charles Sun, published at Lawfare (under its Cybersecurity & Tech and Foreign Relations & International Law sections) on May 6, 2026. The essay argues that tighter US export controls on advanced chips to China reinforce, rather than disrupt, the Chinese incentive system that drives AI capability acquisition, by deepening the resource dependence that motivates firms to align with state priorities. It introduces a named framework, Proactive Elite Alignment Theory (PEAT), and anchors its analysis on China's generative-AI registration system (da moxing bei'an, 大模型备案) and on the Manus AI acquisition saga as the central attempted-exit case.
Summary of argument
Sun's central claim is that export controls do not merely fail to disrupt China's AI incentive system but strengthen it: by restricting the supply of advanced foreign chips, the controls deepen Chinese firms' dependence on state-subsidized domestic compute, which is the resource dependence that motivates firms to align with state priorities. On this account the policy tool designed to constrain capability acquisition reinforces the system that motivates it.
Sun frames the mismatch as structural rather than operational: US enforcement tools — the Entity List under the Export Administration Regulations, IEEPA sanctions, and Export Control Reform Act prosecutions — operate on the supply side, while the Chinese incentive system that drives behavior operates on the demand side.
The essay organizes the Chinese incentive architecture into three institutional layers, which Sun labels the gate, the field, and the backstop.
The gate: registration as a survival condition
Under the 2023 Interim Measures for the Management of Generative AI Services (issued by the CAC together with six ministries), any firm offering generative AI services to the Chinese public must complete formal registration (备案). Sun characterizes this not as voluntary but as a regulatory prerequisite for market access. By the end of 2025, 748 generative AI services had completed bei'an with CAC, and 435 applications and features drawing on registered models had completed a separate filing process (登记). Registration requirements include a security self-assessment report exceeding 100 pages, training-data provenance documentation, content-safety benchmarks covering 31 enumerated safety risks across five categories (per CSET analysis), and emergency-response protocols for noncompliant output.
The revised Cybersecurity Law (effective January 1, 2026) for the first time incorporated AI governance provisions into China's statutory legal framework. Unregistered models face app-store removal, fines, or operational suspension.
Sun's analytical claim is that registration functions as a survival condition rather than a tier of privilege: the gate filters for alignment not by rewarding the compliant but by eliminating everyone else, so that the firms inside the gate are the only firms that exist. What Sun argues distinguishes the generative-AI registration system from prior Chinese content regulation of film, television, and publishing is that its compliance requirements reach into production inputs — training data, training methodology, model weights, and content-safety benchmarking — not just finished outputs.
The field: differential subsidies channeling compute choice
Inside the gate, Sun describes compute vouchers with chip-origin-differentiated reimbursement rates that create dependency on domestic compute:
- Shenzhen (March 2025): the first batch of "training power vouchers" (训力券), roughly 200M yuan to about 40 firms, with individual awards up to 10M yuan, within a 4.5B yuan combined digital-economy and tech package.
- Beijing Yizhuang Economic Development Zone: 100M yuan per year in compute vouchers, with an individual cap of 20M yuan, reimbursing 40% for domestic AI chips and 30% for nondomestic.
- Hangzhou: 250M yuan per year, reimbursing 30% for domestic compute and 20% for others.
- National policy (January 2026): eight central ministries led by MIIT issued the "AI Plus Manufacturing" joint guideline, the first national-level policy to reference "compute vouchers" as a policy instrument.
Sun's structural argument is that the differential is not large in any single transaction, but that over a year of model training a 10-percentage-point subsidy premium on domestic compute translates into millions of yuan, so the price signal accomplishes what no directive needs to. On Sun's reading the subsidy structure creates dependency rather than only cost reduction, because model architectures, engineering workflows, and cost structures optimize around the subsidized platform. The resulting switching costs are structural rather than contractual: no clause prevents exit, but reengineering an entire production pipeline does.
The backstop: the Manus AI saga
Sun presents Manus AI as the central attempted-exit case. Manus was built on Anthropic Claude (a US-developed foundation model) and served a global subscription user base rather than depending on domestic procurement. In mid-2025 it relocated its headquarters from Beijing to Singapore, and earlier in 2025 it secured US venture-capital investment (Benchmark's $75M among others, per Bloomberg) alongside existing Chinese backers. In December 2025 a Meta acquisition was announced for $2 billion. On April 27, 2026, China's foreign-investment security-review mechanism — an interagency body led by the NDRC and the Ministry of Commerce — ordered the parties to withdraw the deal, citing only "laws and regulations," and exit bans were imposed on the co-founders.
Sun's claim is that Manus was the rare Chinese AI firm that had systematically minimized state-ecosystem dependencies, both by design (a US foundation model and a global subscription base) and by later strategic choice (a Singapore headquarters and US venture capital), and that the clearest available case of an attempted exit ended in a state veto. For Sun the case demonstrates that the incentive architecture has a coercive backstop even for firms that reduce their economic pull, and that the backstop is operative rather than latent.
The PEAT framework
Sun's framework, Proactive Elite Alignment Theory (PEAT), developed in prior research on Chinese state-business dynamics, identifies four structural variables that the essay argues predict when firms in China's technology sector preemptively align with state priorities:
- Central strategic signaling — the State Council's designation of AI as a national priority, the eight-ministry joint AI guidelines, and "computing-electricity coordination" in the 2026 Government Work Report.
- Cascading local escalation — local governments competing to over-implement (the Shenzhen, Hangzhou, and Beijing subsidy race), which Sun describes as structurally resembling China's pandemic-era "dynamic zero" campaigns.
- Resource dependence — operating through differential subsidy rates that lock firms into domestic compute.
- Organizational capacity — the speed with which Chinese AI firms restructure engineering pipelines around subsidized infrastructure.
The export-control feedback loop
Sun's central causal claim is a four-step feedback loop. First, export controls restrict the supply of advanced foreign chips to Chinese firms, the intended effect. Second, restricted foreign chip supply increases dependence on domestic compute, whose pricing is shaped by state subsidies and whose availability is channeled through government-backed programs, deepening the resource-dependence variable. Third, deeper resource dependence strengthens the incentive for firms to align with state priorities: by making domestic computing cheaper through differential reimbursement, the economically rational choice and the politically preferred choice point in the same direction, so firms align because the price structure makes alignment the path of least resistance. Fourth, aligned firms use every available pathway to acquire capabilities that unlock further state resources; Sun argues that distillation of American frontier models is one such pathway, framed not primarily as espionage or defiance but as a rational response to an incentive structure that rewards capability acquisition from every available source. The loop closes when distillation triggers calls for tighter export controls, which further restrict foreign chip supply, deepening resource dependence and strengthening the alignment incentive.
Sun acknowledges a boundary condition: the loop holds only as long as domestic computing is sufficient to produce competitive models. If the performance gap between domestic chips and restricted foreign alternatives became unbridgeable, the incentive structure would break down, since resource dependence without resource adequacy produces failure rather than alignment. Sun cites Lennart Heim's RAND analysis that China can still develop competitive models despite compute constraints, because frontier training consumes only a fraction of total compute capacity, and argues that the boundary condition has not been reached.
Engagement with the US policy debate
The essay positions itself against two recent US-side arguments. Sun engages with Joe Khawam (Just Security, February 2026), who proposes phased sanctions escalation targeting Chinese AI distillation and treats adversarial distillation as behavior that sanctions can deter by raising costs; Sun's response is that this framing misses the incentive system on the Chinese side. Sun engages with Ryan Fedasiuk (War on the Rocks, April 2026), who focuses on regulatory tools to limit diffusion of Chinese open-source models in response to national-security threats from China's open-source AI expansion; Sun offers the same critique. Sun's position is that both are correct about the tools available to the United States, but that neither engages the incentive system on the Chinese side, which Sun treats as supply-side approaches that are incomplete rather than wrong.
Relationships
- supports: PEAT — Proactive Elite Alignment Theory (created — canonical anchor), China Generative AI Registration (da moxing bei'an / 大模型备案) (created — da moxing bei'an anchor)
- deploys-in: Export Controls (AI) (Sun's feedback-loop section)
- deployed-by: Cyberspace Administration of China (CAC) (CAC, per the 2023 Interim Measures and the 2026 Cybersecurity Law revision)
- related: Distillation, Adversarial Distillation, Distillation, Sovereign AI (Product Concept), US-China AI Competition: Different Races, Different Metrics, AI Diffusion, China's Not the Problem. We Are. (NYT Interesting Times, Douthat-Chan, May 14 2026) (Chan's "multiple races" framework — Sun's PEAT explains why China can compete on the diffusion and applications races even without parity on the frontier-model race)
- contradicts: Khawam's "sanctions-as-deterrence" framing and Fedasiuk's regulatory-tool framing, both treated as supply-side incomplete (not wrong, but insufficient)
Sources
Raw Sources/The Incentive Architecture Export Controls Cannot Reach.md— full text- Published at Lawfare: lawfaremedia.org