AI Safety in an Age of Geopolitical Competition: Whose Rules, Whose Market?

image courtesy of Gen AI - prompt author

Introduction

The conversation about AI safety has moved from research labs and philosophy seminars into boardrooms, parliaments, and dinner-table arguments. What was once a relatively contained technical debate about alignment and control has become entangled with industrial policy, national security, and great-power competition. This entanglement matters because it changes the problem's shape. A purely technical question - how do we ensure advanced AI systems behave as intended - is one thing. A question about how to govern a technology that is simultaneously a commercial product, a strategic asset, and a potential existential risk, developed by companies and states with divergent interests, is quite another.

This essay examines that second, messier question. It asks what happens to safety discourse once geopolitics, market economics, and infrastructure dynamics are folded in. It argues that the resolution of the "AI safety problem" will likely be determined less by consensus on risk than by the structural realities of who can afford to build frontier systems, who controls the infrastructure those systems run on, and what consumers are actually willing to pay for.

The Safety Discourse: Existential Risk or Strategic Rhetoric?

Public statements from prominent AI researchers warning that the technology is advancing too quickly and could pose catastrophic or existential risks have become a familiar feature of the discourse. These warnings deserve to be taken seriously; the people making them include some of the field's most technically credentialed figures, and the underlying concern - that systems could become powerful enough to act in ways their creators neither intend nor can correct - is a coherent one, even if there is no scientific consensus on its likelihood or timeline.

At the same time, it would be naïve to treat every safety statement as disinterested. Critics have offered at least two alternative readings of the same public warnings. One is that researchers and executives are, in effect, "stoking their own fire" - that dramatic warnings about uncontrollable AI generate attention, talent interest, and a sense of inevitability that benefits the firms making them. A second, more cynical reading holds that safety rhetoric functions as market positioning, particularly for companies approaching public offerings, where a narrative of being the "responsible" lab can be commercially useful.

Neither of these readings should be accepted uncritically either. Motivated reasoning is not the same as false reasoning, and it is entirely possible for a statement to be both self-serving and substantively correct. The more useful posture - and the one this essay adopts - is to treat the object-level claims about risk on their merits, while remaining alert to the incentives of whoever is making them. For an academic or consultant audience, the discipline here is familiar: separate the argument from the arguer, weigh the evidence, and resist the temptation to resolve uncertainty through cynicism in either direction.

The Geopolitical Complication

Layer geopolitics onto this picture and the situation becomes considerably harder to reason about. A significant share of low-cost, openly available AI models now originates from China. Even setting aside questions about the extent to which these models are distilled from or trained using outputs of Western systems - itself a contested and technically difficult question to adjudicate - the more basic point stands: these models are, on many benchmarks, competitive with their American counterparts, and they are frequently cheaper or freely available. This creates an asymmetry that safety-focused regulation in any single jurisdiction cannot easily resolve. A regulatory regime built in Washington, Brussels, or London constrains the companies operating under its authority. It does not constrain a lab operating under a different political and legal system, with different incentives around transparency, disclosure, and openness. Analysts and policymakers disagree sharply about how much this matters in practice. Some argue that safety and capability are sufficiently intertwined that a lab cutting corners on safety would also cut corners on reliability and performance, making unsafe systems less commercially competitive over time. Others argue the opposite: that safety investment is a cost with no immediate payoff in benchmark performance, and that a regulatory gap simply hands a competitive advantage to whichever actor is willing to forgo it.

It is also worth resisting a simple frame in which "the West" is straightforwardly open and market-driven while China is straightforwardly closed and state-directed. The reality is more layered: China's AI sector includes highly competitive private firms operating in a market environment, even as that environment remains subject to significant state involvement and coordination in ways that differ meaningfully from Western regulatory traditions. This difference is real and relevant to how governance proposals will land in practice, but it is a difference of degree and structure rather than a clean binary, and analysts should be cautious about overstating the contrast in either direction. The practical upshot is that any Western safety framework has to be evaluated not only on its own technical merits, but on how it performs in an environment where a meaningful share of global AI capability sits outside its jurisdiction. A framework that meaningfully slows down five American and European labs while leaving global capability diffusion largely unaffected is a very different policy instrument than one which genuinely constrains the global frontier.

What Consumers Actually Optimise For

Amid the geopolitical and existential framing, it is easy to lose sight of a more mundane but arguably more decisive variable: what end users actually choose. The evidence available to date suggests that for the large majority of consumers and businesses selecting AI tools, price and capability dominate the purchasing decision, while safety - understood broadly as alignment, transparency, or provenance - receives comparatively little weight, despite its prominence in media coverage and policy debate.

This is not a novel pattern; it echoes long-standing findings in consumer behaviour research across other domains, from data privacy to food safety to financial products, where stated preferences for "ethical" or "safe" options diverge substantially from revealed preferences at the point of purchase. There is little reason to expect AI tooling to be an exception, particularly in enterprise and developer contexts where procurement decisions are often driven by cost-per-token, latency, and integration ease rather than by an assessment of a model provider's safety practices.

If this pattern holds, it has a significant implication for safety-focused regulation: rules that raise costs for compliant providers without a corresponding consumer willingness to pay for safety risk simply shifting demand toward cheaper, less-regulated alternatives - whichever jurisdiction they originate from. This is not an argument against safety regulation; it is an argument that safety regulation which relies on consumer preference to reinforce it, rather than on structural or infrastructural leverage, may be building on weak foundations.

Regulation, Cost, and the Risk of Entrenching Incumbents

This brings the discussion to a structural question that deserves more attention than it typically receives in public debate: who actually benefits from broad safety regulation applied uniformly across the industry? If comprehensive regulatory requirements - extensive testing regimes, disclosure obligations, liability frameworks, compute reporting thresholds - are applied to all companies building frontier-scale models, the compliance cost is unlikely to be distributed evenly. Well-resourced incumbents, with existing legal, policy, and safety-research infrastructure, are comparatively well positioned to absorb these costs. Smaller labs, startups, and academic or open-source initiatives are not. There is a real possibility, worth taking seriously rather than dismissing as a libertarian talking point, that broad-based regulation - however well-intentioned - functions in practice as a moat, entrenching whichever handful of companies currently sit at the frontier.

This is not a new dynamic in regulated industries; it has well-documented precedents in banking, pharmaceuticals, and telecommunications, where compliance costs have historically favoured incumbents capable of amortising fixed regulatory costs across larger revenue bases. Applied to AI, the question becomes: would broad regulation targeting frontier model development simply preserve the current standing of the five or six companies - OpenAI, Anthropic, Google DeepMind, Meta, and xAI, among the most frequently cited - that currently define the frontier, at the expense of would-be challengers, including academic institutions and public-interest initiatives that might otherwise contribute meaningfully to the field?

This is a genuine tension rather than a rhetorical one. Safety advocates are correct that some minimum floor of testing and disclosure is desirable for systems with significant potential for harm. Competition advocates are correct that regulatory capture by incumbents is a well-documented risk. Reasonable policy design tries to thread this needle — for instance, by scaling requirements to compute thresholds or deployment scale rather than applying flat rules to all developers — but it is not obvious that such calibration will survive the political process intact, and it is worth academics and consultants tracking this space treating incumbent-preservation as a plausible outcome of regulation, not merely an unintended side effect to be waved away.

AI as Infrastructure: The Cloud Analogy

A useful lens for thinking through where this settles is the analogy, increasingly common among industry observers, between AI providers and cloud infrastructure providers. Amazon, Microsoft, and Google collectively host a substantial share of the internet's computing and storage infrastructure. Few users interact directly with AWS, Azure, or Google Cloud; instead, they use applications built on top of them, largely indifferent to which underlying provider is doing the work, so long as the application performs reliably and affordably.

There is a plausible case that advanced AI models are heading toward a similar structural position: not a product consumers evaluate directly on grounds of "safety," but an infrastructural layer selected by application developers and enterprises largely based on cost, latency, and fitness for a given task. In such a world, the meaningful competitive and regulatory battleground shifts away from public perception of any individual model's safety credentials, and toward the economics of the infrastructure layer itself - compute costs, energy availability, chip supply chains, and data centre geography.

If this analogy holds, the incorporation of safety requirements into that infrastructure layer - model cards, red-teaming disclosures, usage restrictions - would likely be absorbed into the overall cost structure much as security certifications and compliance standards have been absorbed into cloud computing pricing. The outcome would not be that safety disappears from consideration, but that it becomes one input among several - alongside price, reliability, and performance - in what is ultimately an economically driven allocation decision, made largely by intermediary developers and enterprises rather than by end consumers weighing safety directly.

Conclusion

None of this is to suggest that safety concerns are unimportant, or that the researchers raising them should be dismissed as self-interested actors dressed in existential-risk language. The technical case for caution around increasingly capable systems is substantive, and merits continued attention from policymakers, technologists, and the public alike. But it is equally important to recognise that safety will not be resolved in a vacuum. It will be resolved - or more likely, continuously renegotiated - within a landscape shaped by geopolitical competition between jurisdictions with different regulatory philosophies, by consumer behaviour that consistently privileges cost over stated ethical preference, by the risk that well-meaning regulation entrenches the incumbents it aims to constrain, and by the gradual consolidation of AI into an infrastructural layer governed as much by the economics of compute and energy as by the ethics of alignment.

For academics and consultants engaging with this space, the practical implication is a call for analytical humility and structural thinking. Debates that treat AI safety as primarily a matter of persuading the public or credentialing the right experts risk missing the forces that will actually determine outcomes: capital expenditure, compute access, regulatory design, and market structure. A more productive research and advisory agenda would ask not only "what does safe AI look like?" but "under what economic and geopolitical conditions does safety become something the market, and the state, are actually equipped to enforce?" The answer to that second question may end up determining the fate of the first.