The Geopolitics and Economics of Pacing Frontier AI

1. A precarious inflection point

The global artificial intelligence ecosystem has reached a precarious inflection point. The rapid scaling of computational power, expansive datasets, and algorithmic sophistication has consistently outpaced the development of corresponding safety, security, and governance frameworks.

The emergence of highly agentic, multimodal frontier models has prompted leading AI researchers and industry executives to advocate for a deliberate deceleration—often termed "pacing"—of AI development. Proponents such as Anthropic CEO Dario Amodei argue that modulating the rate of capability improvements by just one to two years could provide the necessary runway to align models, secure critical digital infrastructure, and establish robust international verification regimes. The consensus among safety advocates is that slowing the rate from "extremely fast" to "somewhat fast" surrenders relatively little strategic advantage while dramatically mitigating catastrophic risk.

However, artificial intelligence is no longer merely a commercial enterprise; it has become the fundamental currency of modern geopolitical power, economic productivity, and military capability. The proposal to pace AI development introduces a complex matrix of strategic pros and cons across the United States, the European Union, and China. A unilateral slowdown by any single sovereign actor risks ceding the technological frontier to systemic rivals, whereas a multilateral slowdown faces near-insurmountable hurdles in verifiable enforcement, supply chain sovereignty, and international trust.

2. The mechanics of pacing (not a full stop)

The contemporary debate regarding an AI slowdown is largely driven by the frontier laboratories themselves, rather than by statutory mandate. Pacing the frontier does not advocate for an outright halt to all AI research and development. Rather, it proposes modulating the rate at which model capabilities cross critical, potentially dangerous thresholds, ensuring that safety, security, and alignment research can advance synchronously with raw capability.

Leading AI developers—including Anthropic, OpenAI, and Google DeepMind—have established Responsible Scaling Policies (RSPs) or equivalent frameworks such as the Preparedness Framework and the Frontier Safety Framework. These mechanisms tie the deployment and continued training of advanced models to specific capability thresholds or critical capability levels (CCLs).

• ·Chemical, biological, radiological, and nuclear (CBRN) risks: whether a model meaningfully uplifts a malicious actor beyond what search engines provide.
• ·Cybersecurity: whether a model can autonomously discover and exploit zero-day vulnerabilities faster and more cheaply than expert human teams.
• ·Autonomous replication and adaptation (ARA): whether an agent can acquire resources, secure compute, evade shutdown, and spawn copies in unmonitored environments.
• ·Automated AI R&D: whether a model can fully automate entry-level machine learning research, compressing years of aggregate progress into a fraction of the time.

3. RSPs, thresholds, and transparency

When a model approaches these thresholds, developers publicly commit to heightened safeguards. For instance, Anthropic's AI Safety Level 3 (ASL-3) requires advanced security standards against state-sponsored espionage and robust deployment constraints before a model crossing these capability thresholds can be utilized.

Civil society groups have noted a concerning shift in how thresholds are defined. While earlier iterations of Anthropic's RSP utilized strict quantitative benchmarks (for example, a 50% aggregate success rate on autonomy evaluations), Version 3.4 transitioned to qualitative descriptions such as the ability to automate an entry-level remote researcher. Critics argue this shift reduces transparency and accountability, potentially allowing laboratories to shift goalposts as commercial pressures mount.

• ·Anthropic — Responsible Scaling Policy (v3.4): AI R&D automation, CBRN uplift, cyber offense; mitigation via AI Safety Levels and qualitative thresholds requiring affirmative safety cases.
• ·OpenAI — Preparedness Framework: high/critical risk across CBRN, cyber, and AI self-improvement; pre-deployment capability elicitation and deployment pauses at critical levels.
• ·Google DeepMind — Frontier Safety Framework: critical capability levels with a two-tier structure designed to catch risks early in pre-training.
• ·METR — capability elicitation: time horizon doubling roughly every seven months, ARA success rates; third-party red-teaming via Hawk and MALT datasets.

4. Independent evaluation and METR

The credibility of RSPs relies heavily on third-party evaluations. Organizations such as Model Evaluation and Threat Research (METR)—a spin-off from the Alignment Research Center—conduct pre-deployment capability elicitation for frontier labs. METR evaluates the "time horizon" of AI agents: the maximum duration a model can autonomously operate before its success rate on a complex task drops to fifty percent.

Recent METR analyses, utilizing evaluation suites like RE-Bench and HCAST, indicate that the time horizon for autonomous AI agents has been doubling roughly every seven months. Extrapolating this trend suggests that within a few years, AI systems could seamlessly execute month-long autonomous projects. The primary argument for a slowdown is rooted directly in these evaluations: if time horizons soon exceed humanity's ability to monitor, interpret, or control agentic models, a pause is required to develop advanced interpretability tools and verifiable alignment techniques before crossing the threshold of catastrophic risk.

5. United States: the innovator's dilemma

The United States currently dominates the global frontier of artificial intelligence, with compute dominance, concentrated AI talent, and massive venture capital deployment that tolerates high-risk innovation. For the US, pacing AI development balances existential risks of uncontrolled superintelligence against geopolitical and economic imperatives of maintaining global hegemony.

6. Why the US might welcome pacing

Mitigation of catastrophic risks and cyber vulnerabilities is the most immediate benefit. Rapidly advancing models are beginning to demonstrate automated vulnerability discovery and dual-use scientific reasoning. A deliberate slowdown allows the US AI Safety Institute and independent evaluators to establish rigorous testing protocols without an arms-race deployment cycle. Pacing also provides time for the US cybersecurity apparatus to harden critical infrastructure against AI-enabled agentic attacks. The Center for a New American Security has warned that emerging AI capabilities threaten to disrupt the traditional cyber offense-defense balance, disproportionately empowering attackers through agentic autonomy.

A second-order benefit involves relief on domestic physical infrastructure. The massive CapEx boom driven by AI is colliding with hard physical constraints, primarily the US electrical grid. AI data centers require immense, continuous, carbon-free base-load power. The US interconnection queue features wait times averaging five years, with nearly 2,300 gigawatts of generation and storage capacity stalled in regulatory review—more than the country's entire installed power capacity.

To bypass this "electron gap," US hyperscalers are securing nuclear power at unprecedented scale: Microsoft has committed $16 billion to restart Three Mile Island (Crane Clean Energy Center) by 2028; Google has partnered with Kairos Power for small modular reactors by 2030; Amazon has invested over $20 billion converting the Susquehanna site into a nuclear-powered AI data center campus. Concurrently, the US is experiencing a 76% increase in gas-fired power generation construction, threatening long-term climate goals and raising consumer electricity rates. Decelerating model scaling would reduce frantic demand for gigawatt-scale data centers, allowing the energy supply chain, nuclear approvals, and grid modernization to catch up sequentially rather than destructively.

7. Why the US might reject unilateral pacing

The most severe risk is ceding geopolitical and military superiority. The Trump administration, reflecting broad consensus within the US defense establishment, has explicitly rejected unilateral US slowdown proposals, framing AI development as a zero-sum race where "whoever wins AI wins." AI is central to future military command and control, autonomous drone swarms, and cognitive warfare. Integration into military systems is expected to transform the speed and scale of warfare, enabling drone swarms that overwhelm human cognitive capacity. If the US caps capability advancements, it risks adversaries equipped with superior agentic AI operating at machine speed. RAND notes that while AI could help cyber defense at scale, attackers utilizing frontier models will likely retain a structural edge in penetrating battle networks.

Slowing development could also stifle anticipated macroeconomic growth. Daron Acemoglu, using a task-based macroeconomic model, argues early enthusiasm is overblown: AI will primarily replace "easy-to-learn" tasks while struggling with context-dependent "hard-to-learn" tasks, estimating a modest 0.71% increase in total factor productivity and roughly 1% contribution to global GDP over ten years. Conversely, Erik Brynjolfsson suggests the economy is in the initial dip of a "productivity J-curve"—general-purpose technologies require vast complementary investments before gains appear in national statistics. Stanford Canaries Dashboard data shows employment for early-career workers (ages 22–25) in highly AI-exposed occupations has shrunk by 2.7%, while less exposed fields continue to grow. Artificially pacing AI could stall the J-curve, delaying projected GDP gains needed to offset rising national debts and demographic stagnation.

A US slowdown does not freeze global capabilities. Chinese laboratories utilize industrial-scale knowledge distillation—using outputs, reasoning traces, and synthetic data from powerful "teacher" models to train smaller "student" models. US cybersecurity advisories describe China-based actors using API proxy "transfer stations" to bypass geographic restrictions and terms of service, extracting proprietary capabilities across millions of exchanges. Copyright law does not strongly protect raw model outputs, and trade secret claims are often unviable when information is accessed via public interfaces. If the US slows proprietary advancements but fails to secure API endpoints, it effectively subsidizes China's catch-up at a fraction of safety and R&D cost. Some researchers advocate mandating release of scaled-down "analog" open-weight models to democratize safety testing without exposing full frontier architecture.

• ·National security: delays US military integration of agentic AI and autonomous C2—severe risk; cedes cognitive warfare and cyber-offense advantage.
• ·Infrastructure and energy: eases gigawatt-scale power demands—positive; runway for nuclear PPAs and SMRs to reach commercial operation.
• ·Macroeconomics: delays the productivity J-curve—negative; stalls the technological dividend needed for GDP growth.
• ·IP protection: extends shelf-life of current frontier models without advancing the frontier—highly vulnerable to adversarial distillation via API transfer stations.

8. China: state control and open-weight strategy

The People's Republic of China approaches artificial intelligence as a pillar of national rejuvenation and technological sovereignty independent of Western supply chains. Unlike the US ecosystem's speculative AGI pursuit, China's strategy weights "general AI"—broadly capable systems for industrial execution, advanced manufacturing, and embodied robotics.

Beijing publicly rebukes US slowdown proposals as hypocritical "Cold War-style" tactics to lock in American hegemony while suppressing Chinese innovation under the guise of safety. Functionally, if the US genuinely decelerated frontier scaling, it would ease competitive pressure on Chinese labs. Despite stringent US export controls on advanced NVIDIA GPUs, startups like DeepSeek and firms like Alibaba (Qwen) have achieved remarkable architectural efficiency. DeepSeek-R1 in early 2025 demonstrated reasoning rivaling Western frontier models at a publicly quoted training cost of $5.6 million—a figure that notably excludes value extracted via distillation from US models, but still proves China can engineer around compute scarcity.

A global slowdown in raw algorithmic scaling would shift competition toward energy and infrastructure—areas where China holds structural advantage. China has deployed 46 ultra-high voltage transmission lines spanning 60,000 kilometers; 1100kV lines can transmit up to 12 GW with minimal loss, connecting western clean energy to eastern industrial hubs. In 2024 alone, China added an estimated 429 gigawatts of new power capacity, compared to 51 gigawatts in the US. A pause in Western scaling would let China maximize hardware-infrastructure integration before the race resumes.

China has no domestic incentive to unilaterally slow down. The CCP views AI through cycles of economic growth versus ideological control. Following DeepSeek-R1, Beijing prioritizes AI as a macroeconomic growth engine to offset a sluggish property sector, local government debt, and demographic decline. China's open-weight strategy (Qwen, DeepSeek under permissive licenses) builds a global developer ecosystem, erodes dependency on US APIs, standardizes Chinese models across the Global South, and forces domestic startups onto Huawei Ascend silicon—insulating against future export controls. Domestic regulation under the Cyberspace Administration of China focuses on censorship and ideological alignment rather than Western-style existential risk. Adhering to a Western-defined slowdown based on rogue replication or bioweapons is viewed as incompatible with sovereign technological goals.

9. European Union: regulatory vanguard at a crossroads

The European Union has positioned itself as the global superpower of digital regulation. Lacking US-scale venture capital and China's state-directed industrial capacity, Europe has sought to shape AI through normative power, culminating in the EU AI Act.

A global slowdown would validate the EU's precautionary approach. The AI Act imposes stringent obligations on general-purpose AI models trained above 10²⁵ FLOPs, designating them as presumptively possessing high-impact capabilities and systemic risk. Pacing would give the European AI Office, national authorities, and external auditors time to build technical expertise—transitioning from theoretical framework to enforceable oversight. It would also give European startups such as Mistral and Aleph Alpha breathing room; if the capability ceiling is temporarily fixed, competition shifts toward efficiency, privacy, and compliance.

The downside is competitiveness. Mario Draghi's landmark report highlighted the EU falling behind the US and China due to overlapping regulations, a fragmented single market, and lack of compute capital and STEM talent. The EU produces only 203 ICT graduates per million inhabitants versus 335 in the US. Seventy percent of foundation models since 2017 originated in the United States, and US chatbots capture over 95% of European web traffic. The AI Act's fines—up to 7% of global annual turnover—create asymmetrical compliance burdens that may deter US developers from releasing frontier models in Europe. Industry leaders including Siemens warn that draft GPAI Code of Practice will delay innovation. The Commission's InvestAI initiative aims to mobilize €200 billion and upgrade Euro-HPC, but translating ambition into industrial integration remains a bottleneck. A slowdown could lock Europe into permanent dependency on foreign infrastructure.

10. Enforcing a slowdown: hardware, cryptography, verification

The most profound argument against attempting to slow AI development is the difficulty of verifiable enforcement. An international agreement cannot rely on self-reporting or mutual trust when incentives to defect are existential. Regulators cannot govern what they cannot independently measure.

Multilateral pacing requires verification architectures often compared to the IAEA nuclear inspection regime—but verifying software proliferation differs fundamentally from verifying fissile material. Software has zero marginal duplication cost; once weights leak, safety interventions cannot be revoked. Verification must occur at the hardware level—the physical, excludable substrate of AI.

Advanced semiconductors may need built-in governance: a hardware root of trust (immutable cryptographic identity in silicon via physical unclonable functions or write-once memory), integrated with standards such as IEEE 1012 and IEEE 3864, so regulators can attest to a chip's location, firmware state, and operational topology.

• ·Location verification: cryptographic ping-response times to detect smuggling to sanctioned entities.
• ·Offline licensing: chips requiring a cryptographically signed lease from an AI Safety Institute, expiring after set computational cycles.
• ·Network topology verification: enclave-gated allocations throttling interconnects if chips form unauthorized clusters above FLOP caps.
• ·Zero-knowledge proofs (ZKPs): prove training stayed below capability thresholds or safety filters were applied without revealing weights or raw data—historically costly, but recent work suggests frontier dense pre-training verification within ~36 months at roughly 0.5–1.5% overhead via deterministic GPU kernels (FlashAttention, pinned cuBLAS, Poseidon hashes, Merkle roots on concurrent streams).

11. Confidence-building measures and the TSMC bottleneck

Until hardware and ZKP architectures mature commercially, the global community relies on confidence-building measures—diplomatic tools adapted from Cold War arms control. The International Network of AI Safety Institutes (US, UK, Japan, EU, and others) spearheads shared evaluation standards and incident reporting. Without China's full integration into a verifiable hardware-backed treaty, Western CBMs remain asymmetrical and cannot prevent covert scaling in unmonitored data centers.

Any analysis of pacing must account for the semiconductor supply chain. Frontier AI rests on Taiwan Semiconductor Manufacturing Company (TSMC), which produces over 90% of the world's most advanced logic chips (sub-7nm). This concentration creates a precarious "silicon shield." Analysts project that a military blockade of the Taiwan Strait would halt global AI infrastructure expansion. TSMC foundries rely on uninterrupted specialized chemicals and ASML lithography; kinetic disruption or activation of internal kill switches would sever advanced GPU supply.

The macroeconomic shock could wipe out an estimated $5–10 trillion in global equity value within days and trigger a 5–10% contraction in global GDP. Duplicating supply chains in the US and Europe would require Apollo-program-scale capital—upward of $1.5 trillion and years to parity. The most realistic mechanism for a global, involuntary AI slowdown may not be negotiated regulation or corporate RSPs, but geopolitical conflict in the Indo-Pacific.

12. Strategic synthesis

Pacing frontier AI is driven by legitimate, empirically backed concerns regarding existential risks, autonomous replication, and automated discovery of cyber and biological weapons. Artificial intelligence is the ultimate dual-use technology embedded in fierce tripolar competition.

For the United States, unilateral slowdown is fraught with peril: it may stall macroeconomic productivity gains and cede the military-technological high ground in cyber warfare and autonomous systems to China. Unless paired with strict restrictions on API access and open-weight release, Chinese state-backed labs will continue adversarial distillation to achieve parity at a fraction of the cost.

For China, AI development is an uncompromisable vector for economic sovereignty and bypassing Western embargoes, backed by superior ultra-high voltage grid capacity and state-directed infrastructure speed. Beijing views American pause calls as hypocritical containment.

For the European Union, pacing aligns with regulatory ethos but deepens a competitiveness crisis Draghi articulated: burdensome compliance may relegate Europe to consuming foreign AI, exacerbating brain drain and stifling domestic capital formation.

Ultimately, meaningful slowdown is impossible without ironclad, mathematically verifiable international agreements. Until hardware-enabled mechanisms, zero-knowledge proofs, and on-chip cryptographic attestations standardize across global semiconductor supply chains—functioning as a digital IAEA—any voluntary pause by Western democracies may only accelerate the shifting balance of global power.