The Chain Links Don't Lie: Deconstructing the Governance Theater Around Artificial Intelligence
A leaked position paper circulated among Washington policy circles last month contained a phrase that should alarm anyone tracking the intersection of geopolitics and emerging technology. "The moment has arrived," it read, "for a Geneva Moment in artificial intelligence." The reference to the 1955 Geneva Summit—the last time global powers attempted to negotiate away existential risk—is telling. But unlike the nuclear problem, where physics provided neat boundaries and verifiable metrics, AI governance operates in a domain where the threat model remains contested, the actors are uncountable, and the verification mechanisms exist only as theoretical constructs.
I have spent the better part of a decade tracing wallet clusters, auditing smart contract bytecode, and mapping liquidity flows across DeFi protocols. The analytical discipline required for on-chain forensics—relentless verification, falsifiable hypotheses, evidence before conclusion—offers a useful framework for evaluating policy claims. When I examine the current discourse around US-China AI governance negotiations, I find the same pathologies that precede market collapses: confident assertions unsupported by verifiable data, analogies that break down under technical scrutiny, and a systematic avoidance of the metrics that actually matter.
This analysis examines the structural failures in the current AI governance framework, identifies the concealed competitive dynamics that mainstream reporting obscures, and quantifies what the actual risk parameters look like when stripped of diplomatic euphemism.
The Verification Problem Nobody Wants to Discuss
Nuclear arms control succeeded—not perfectly, but structurally—because physicists had already solved the verification problem. Enrichment facilities emit detectable radiation signatures. Centrifuge cascades consume predictable amounts of electricity. Warhead geometries produce distinguishable satellite imagery. The International Atomic Energy Agency could send inspectors into known facilities and obtain meaningful assurances about unknown ones through correlated observables.
AI development has no equivalent verification pathway. The training process for a frontier model requires computational infrastructure, certainly—thousands of GPUs humming in climate-controlled data centers—but the output is pure mathematics encoded in floating-point tensors. A model's weights can be encrypted, transferred across borders in milliseconds, or distributed across a network of anonymous compute clusters with no physical footprint detectable by satellite. The training data itself may be proprietary, synthetic, or scraped from sources that leave no forensic trail.
Consider the practical implications for any "AI inspection regime" modeled on IAEA protocols. American inspectors arrive at a Chinese AI laboratory and request access to training runs, datasets, and model weights. The laboratory complies—providing carefully curated evidence of a system operating well below capability thresholds. Meanwhile, the actual frontier model trains on infrastructure registered under a dozen shell entities, distributed across universities, state研究所, and commercial partners, with weights extracted incrementally through standard API calls. This is not a hypothetical scenario; it is the operational reality of how dual-use technology has always moved through geopolitical boundaries.
The behavioral verification problem compounds the technical one. Nuclear weapons require nation-states with identifiable command structures, identifiable delivery systems, and identifiable facilities. AI capabilities can be developed by a graduate student with access to cloud compute credits, an open-source framework, and six months of uninterrupted training time. The actor set is not nine nuclear states; it is every research institution, every well-funded hobbyist, every state-sponsored program, and every non-state actor willing to invest the capital.
Current discourse treats AI governance as a problem of coordinating nine rational actors at a conference table. The actual problem is coordinating an effectively infinite set of actors, many of whom have strong incentives to defect from any agreed framework the moment it constrains their competitive position.
The Compute Trap: Why Hardware, Not Ethics, Will Determine the Outcome
Follow the gas, not the hype. The same principle applies to geopolitics. When policy analysts discuss AI governance in terms of "responsible development frameworks" and "shared safety standards," they are describing a negotiation theater that will produce press releases but not binding constraints. The actual negotiation—the one that will determine whether US-China AI competition resolves into accommodation or decoupling—concerns compute.
Semiconductor supply chains represent the only AI-relevant resource that is both strategically critical and physically traceable. NVIDIA's H100 and H800 chips are manufactured at TSMC facilities in Taiwan, designed using EDA software primarily produced by American companies, and shipped through logistics networks that can be monitored, restricted, and surveilled. Unlike model weights or training methodologies, a chip is a physical object that can be counted, tracked through customs declarations, and intercepted if necessary.
This is why the semiconductor export control regime instituted by the Bureau of Industry and Security represents the only meaningful lever in US-China AI negotiations. Every other governance framework—every joint statement on safety, every bilateral working group on AI ethics—functions as either a diplomatic pressure release valve or a public relations exercise designed to manage domestic constituencies in both countries.
The evidence for this claim is not speculative. In 2022, the Biden administration imposed sweeping export controls on advanced semiconductors and chip-making equipment to China. In 2023, those controls were expanded to close loopholes around third-party chip transfers. The cumulative effect has been to force Chinese AI developers into creative compliance strategies: purchasing H800 chips designed for the export market, investing in domestic alternatives like Huawei's Ascend architecture, and developing training methodologies that achieve comparable results with reduced computational requirements.
The DeepSeek moment—where a Chinese laboratory demonstrated frontier-level capabilities using an order of magnitude less compute than Western competitors—represented a qualitative shift in this dynamic. If compute limitations no longer reliably constrain capability development, the export control lever loses effectiveness. American negotiators arrive at any AI governance dialogue with a shrinking set of credible threats and an increasing awareness that the technical assumptions underlying their position have been invalidated.
This creates a peculiar strategic situation. The United States has an incentive to formalize AI governance frameworks precisely because its unilateral technical advantage is eroding. A verifiable treaty structure, even one with weak enforcement mechanisms, would "freeze" the competitive dynamic at a moment favorable to American interests. China, conversely, benefits from the current ambiguity—it can continue developing capabilities while using governance rhetoric to delegitimize American export controls and position itself as a responsible stakeholder in the emerging global AI order.
Neither party is negotiating to prevent catastrophe. Both are negotiating to preserve strategic advantage while managing the appearance of international responsibility.
Governance as Competition: The War of Institutional Frameworks
Wallets connect the dots. Institutional frameworks connect the narratives. The battle over how AI governance is structured—who participates, which principles are prioritized, what enforcement mechanisms are legitimized—represents a second-order competition that will shape the strategic environment for decades regardless of whether any binding agreements are reached.
The European Union recognized this dynamic early. The AI Act, finalized in 2024, establishes a risk-based regulatory framework that treats certain AI applications as inherently high-risk and subjects them to conformity assessments, transparency requirements, and human oversight mandates. The framework is ambitious, contested, and imperfect—but it positions European standards as the default reference point for any multinational corporation seeking to operate across jurisdictions.
American and Chinese frameworks operate on fundamentally different premises. The US approach emphasizes voluntary commitments, sector-specific guidance, and reliance on market incentives for safety development. China's framework centers on content governance, algorithmic recommendation controls, and state administrative registration requirements for large-scale models. Neither framework is interoperable with the other; a model trained under American safety protocols would likely fail Chinese content requirements, while a Chinese-registered model would struggle to demonstrate compliance with American data handling standards.
The emergence of "bilateral AI governance frameworks" between the US and China would not create a unified global standard. It would ratify a bipolar regulatory order, with American and Chinese frameworks serving as the two poles around which other nations must orient. Smaller states—particularly those in the developing world that lack domestic AI industries to protect—would face pressure to align with one pole or the other, effectively choosing between access to American technology markets and access to Chinese investment and infrastructure.
This dynamic explains the otherwise puzzling attention that AI governance receives in diplomatic circles. The goal is not to prevent AI risks—a goal both parties pay lip service to while undermining each other's credibility on implementation. The goal is to establish the definitional framework through which "responsible AI" gets interpreted globally. Control the definition, and you control the market access conditions for every other nation.
The Compliance Tax: How Governance Fragmentation Destroys Innovation
For market participants—AI developers, infrastructure providers, enterprise adopters—the governance fragmentation creates a compliance cost structure that favors incumbents over challengers.
Consider the operational requirements facing a mid-sized AI company attempting to serve customers in both American and Chinese markets under bilateral governance frameworks. The company must maintain two distinct model versions: one trained on datasets and architectures compliant with American safety standards, another meeting Chinese content governance requirements. Content moderation pipelines must filter outputs according to two distinct regulatory logics. Data handling practices must satisfy American data sovereignty norms in US operations while meeting Chinese cybersecurity requirements in PRC operations. Legal review must account for two regulatory regimes with potentially conflicting enforcement priorities.
This is not a hypothetical burden. It is a description of the operational reality facing any multinational technology company navigating existing US-China regulatory divergence in adjacent domains—cloud services, data localization, cross-border data transfer. The incremental cost of operating under dual compliance regimes is not linear; it is multiplicative, because requirements in one jurisdiction may create conflicts that require resolution engineering in the other.
The beneficiaries of this fragmentation are not AI safety or global stability—they are the large incumbents with dedicated compliance teams, regulatory affairs departments, and the capital to absorb multi-jurisdictional legal risk. Google, Microsoft, Amazon, Alibaba, Tencent, and ByteDance can absorb the compliance tax. The startup seeking to build the next foundation model cannot.
This represents a structural transfer of competitive advantage toward incumbents that is rarely acknowledged in governance discussions. Every framework that creates compliance complexity without creating safety benefits disproportionately benefits those with existing market power.
The Military Exception: Where Actual Progress Becomes Possible
If the broader AI governance framework is theater, one narrow domain presents genuine opportunities for structured negotiation: military applications of artificial intelligence.
Autonomous weapons systems—drones, loitering munitions, targeting algorithms—operate in an institutional environment more similar to traditional arms control than to general AI governance. The actors are identifiable (nation-states with defense ministries), the systems are observable (weapons procurement follows predictable supply chains), and the risks are acute (accidental escalation through misclassification or command-and-control failure).
Both the United States and China have expressed concern—through separate channels, not joint statements—about the risks of AI-enabled autonomous weapons operating without meaningful human oversight. The concerns are not identical in framing or priority, but the operational overlap is sufficient to support confidence-building measures: hotlines between military AI systems, notification requirements for autonomous weapons tests, and limits on AI integration in nuclear command-and-control systems.
This is the domain where the nuclear analogy actually holds. Military AI applications are state-controlled, procurement-dependent, and operate within command structures that can be engaged through diplomatic channels. Verification is imperfect but tractable—drone procurement can be tracked through supply chain monitoring, autonomous weapons tests can be observed through national technical means, and command system architectures can be assessed through signals intelligence.
Progress in military AI governance would not solve the broader problem of frontier model safety or existential risk mitigation. But it would establish precedents for cooperative verification, create institutional channels for managing crises, and demonstrate that structured engagement is possible even between adversaries with conflicting interests. The alternative—allowing AI integration into military systems without any coordination—creates precisely the instability that arms control frameworks were designed to prevent.
The Signal Architecture: What Actually Matters for Monitoring
For market participants and policy analysts seeking to separate signal from noise in AI governance reporting, the following indicators represent meaningful data points:
First, and most important: semiconductor export control policy adjustments. Any reversal, relaxation, or reinterpretation of current export restrictions represents a fundamental shift in the negotiating posture of both parties. The compute regime is the only lever that matters; everything else is negotiation theater.
Second: establishment of official intergovernmental dialogue mechanisms. The existence of a structured channel—joint working groups, regular ministerial consultations, technical-level exchanges—matters more than the content of any single meeting. Institutional continuity creates accountability structures that constrain defection.
Third: military-specific confidence-building measures. Hotlines, notification requirements, and limits on autonomous weapons integration represent the one domain where verifiable progress is possible. Statements about "general AI safety cooperation" without military-specific content are statements about public relations, not policy.
Fourth: compute threshold reporting requirements. American executive orders currently require frontier AI developers to report training runs exceeding 10^26 FLOPs. Any bilateral framework that incorporates mutual reporting of compute-intensive training activities would represent a genuine verification mechanism—imperfect, but meaningful.
The absence of any of these indicators from a US-China AI governance announcement suggests the announcement is primarily domestic political theater with minimal international policy content.
The Incumbent's Advantage: Who Benefits from Fragmentation
Code is the only witness that cannot be corrupted by narrative pressure. The data on AI market concentration tells a story that governance discussions systematically obscure.
The five largest AI laboratories by capability—OpenAI, Anthropic, Google DeepMind, Meta AI, and the Chinese consortium led by ByteDance and Alibaba—collectively control the infrastructure, talent, and capital necessary to absorb compliance costs that would be prohibitive for smaller competitors. Each has dedicated regulatory affairs teams, international legal counsel, and the ability to operate parallel compliance structures across jurisdictions.
This is not a coincidence of market dynamics; it is a structural consequence of governance fragmentation. Every compliance requirement that increases fixed costs relative to variable costs favors larger players with existing market power. The startup that might challenge incumbent positions cannot afford the legal infrastructure necessary to navigate bilateral compliance regimes. The result is a market structure that becomes progressively more concentrated under the guise of safety regulation.
For investors and market participants, this suggests a counterintuitive risk profile. AI governance frameworks, particularly those that create compliance complexity across jurisdictions, should be understood as barriers to entry that favor existing incumbents. The beneficiaries of US-China AI governance negotiations are not global stability, not AI safety, and not the international community—they are the five or six companies with the capital to operate simultaneously in American and Chinese regulatory environments.
The Open Source Variable: Where the Analogy Completely Breaks Down
Nuclear weapons cannot be open-sourced. A warhead design posted on GitHub would be useless without access to weapons-grade materials, specialized manufacturing equipment, and the technical expertise to assemble a functional device. The barrier to nuclear proliferation remains physical, not informational.
AI capabilities do not have this structural protection. The Llama model family, developed by Meta, demonstrated that frontier-adjacent capabilities can be released as open-source software, distributed globally, and modified by anyone with sufficient technical skill. Subsequent iterations by the open-source community have produced models that match or exceed the performance of proprietary systems on specific benchmarks.
This creates a fundamental tension in any governance framework that excludes open-source development. State actors can negotiate compute restrictions and compliance requirements for proprietary models developed by identifiable corporations. They cannot prevent a graduate student in Nairobi from downloading an open-source model, fine-tuning it on domain-specific data, and deploying it for purposes that would violate any bilateral agreement.
The governance implications are significant. Any US-China framework that focuses exclusively on corporate development of frontier models will be increasingly irrelevant as capabilities democratize through open-source distribution. The nuclear analogy, already strained at the verification level, completely collapses when applied to open-source development, which by design has no center to target for compliance.
What Comes Next: The Institutionalization of Ambiguity
The most likely outcome of any US-China AI governance framework is not cooperation and not conflict, but the institutionalization of strategic ambiguity. Both parties gain domestic political value from appearing to engage in responsible AI governance. Neither party gains sufficient benefit from binding commitments to accept the verification and enforcement costs they would require.
The announcement will use the language of cooperation. The implementation will reflect competition. The monitoring will be impossible. The compliance costs will be borne by smaller market participants while incumbents absorb the overhead. The military domain will see marginal progress in confidence-building measures while civilian AI development continues without meaningful coordination.
This is not a pessimistic conclusion; it is a structural prediction based on the incentive alignment of the relevant actors. Both the United States and China have strong incentives to maintain the appearance of governance engagement without accepting the constraints that genuine governance would require. The announcement is not the policy; the announcement is the noise around the policy.
For market participants, the signal is clear: treat AI governance announcements as political events with limited policy content, and focus instead on the metrics that actually determine competitive dynamics—compute access, talent distribution, research velocity, and the regulatory environments that shape market structure. The negotiations will continue. The frameworks will multiply. The verification will remain impossible. The incumbents will continue to benefit. The fundamental competition will proceed on its own logic, indifferent to the diplomatic theater surrounding it.
The question is not whether AI governance frameworks will be announced. They will be. The question is whether any of them will create accountability structures sufficient to constrain the competitive dynamics that actually matter. The evidence suggests the answer is no—and that market participants should position accordingly, treating governance rhetoric as background noise and focusing on the hardware, talent, and infrastructure that actually determine AI capability development.