On a Monday in late October 2023, the White House released an executive order that ran to nearly 20,000 words. It mandated safety testing for powerful AI models, established reporting requirements for frontier AI developers, and instructed dozens of federal agencies to examine their own exposure to the technology. The Biden administration called it the most sweeping action on AI safety in history. Critics called it a dress rehearsal. Within fifteen months, a new administration had revoked it outright.
That whiplash captures the central tension of the AI era: the technology is advancing faster than any regulatory framework has ever moved, while the question of who should be allowed to slow it down — and for what reasons — remains genuinely, sometimes dangerously, unresolved. The answers to that question will determine not just the shape of the artificial intelligence industry, but perhaps the shape of everything else.
The Voluntary Slowdown That Wasn’t
The most dramatic deceleration attempt in AI history arrived not from a government but from a think tank email list. In March 2023, the Future of Life Institute published an open letter calling for a six-month pause on training AI systems more powerful than GPT-4. More than 30,000 people eventually signed it, including Elon Musk, Apple co-founder Steve Wozniak, and a cavalcade of AI researchers. The letter warned of “profound risks to society and humanity” and asked: “Should we let machines flood our information channels with propaganda and untruth? Should we automate away all the jobs, including the fulfilling ones? Should we develop nonhuman minds that might eventually outnumber, outsmart, obsolete and replace us?”
The pause never happened. Not for a single day. The major AI labs — OpenAI, Google DeepMind, Anthropic, Meta — continued training at full throttle. The episode was revealing: in the absence of binding legal mechanisms, voluntary restraint in competitive industries tends to evaporate under market pressure.
The underlying structure resembles a prisoner’s dilemma: every actor that slows down unilaterally creates space for a competitor to advance, which is why many observers argue that only external enforcement can change the dynamic.
The uncomfortable corollary is that external enforcement — government regulation — brings its own distortions. Regulatory frameworks can be captured by incumbents, can stifle legitimate innovation, and have historically moved at a pace that makes technology obsolete before rules are even finalized. The EU’s AI Act, for example, took over three years to negotiate and pass, by which point several of its risk-tier classifications were already being questioned as outdated.
The Regulatory Patchwork and Its Gatekeeping Effect
There is still no comprehensive federal AI law in the United States. Instead, a proliferating patchwork has emerged: sector-specific guidance from the FTC on deceptive AI practices, requirements from the SEC on AI-related disclosures for public companies, FDA pathways for AI-enabled medical devices, and a growing number of state laws — California, Colorado, and Texas among them — addressing algorithmic discrimination, deepfakes, and automated decision-making.
The European Union’s AI Act, which entered into force in August 2024 and is being phased in over several years, is the most ambitious attempt so far at a unified framework. It classifies AI applications into risk tiers — unacceptable, high, limited, and minimal — and imposes corresponding requirements. High-risk applications, including those used in credit scoring, hiring, and critical infrastructure, face mandatory conformity assessments, human oversight requirements, and detailed documentation obligations before deployment.
The compliance costs are substantial. A 2021 analysis by the Center for Data Innovation, drawing on the European Commission’s own impact assessment, estimated that a small business could face compliance costs of up to €400,000 for a single high-risk AI product. For smaller companies, that is often prohibitive. For large incumbents with legal teams already fluent in European regulatory frameworks, it is manageable — and, crucially, it raises barriers to entry that protect existing market positions.
This is the gatekeeping paradox of AI regulation: the entities best positioned to navigate costly compliance regimes are precisely the entities that critics argue should face the most scrutiny. The risk is that regulation ends up cementing the position of a small number of large players rather than creating a genuinely safer or more competitive ecosystem — a tension that sharpens whenever the largest labs themselves lobby for the rules they would be best placed to meet.
Antitrust as a Backdoor Throttle
The competition angle of AI development may ultimately matter as much as explicit safety regulation. The AI industry has consolidated around a handful of extremely well-capitalized companies with access to the three scarce resources that define competitive advantage: massive datasets, the computing infrastructure to train on them, and the talent to do so effectively.
Microsoft’s investment in OpenAI, which now totals over $13 billion, gave it a relationship so intimate that the two companies share infrastructure and distribute products through integrated channels. Google DeepMind merges two of the most advanced AI research organizations in history under Alphabet’s umbrella. Meta has constructed what is arguably the world’s largest AI research operation as a cost center attached to an advertising business — meaning it can afford to release powerful models like the Llama series for free, a competitive dynamic that is economically devastating for standalone AI companies trying to charge for similar capabilities.
Antitrust regulators on both sides of the Atlantic have taken notice. In 2023, the UK’s Competition and Markets Authority opened a review of AI foundation model markets, publishing updated findings in 2024. In January 2024, the FTC under Lina Khan ordered Microsoft, OpenAI, Amazon, Google, and Anthropic to hand over information about their AI investments and partnerships. The European Commission opened preliminary inquiries into AI partnerships. The investigations share a common concern: that the integration of AI capabilities into dominant platforms could extend those platforms’ market power into adjacent markets in ways that harm competition and, over time, innovation.
Put bluntly, when one company controls the model, the distribution channel, and the underlying compute infrastructure, the market starts to look less like a market and more like a tollbooth.
For AI development specifically, the antitrust dimension cuts in complicated directions. More concentrated markets may move faster in one sense — large incumbents can invest at scales that smaller competitors cannot match, and the compute costs of training frontier models now run into hundreds of millions of dollars. But concentration can also slow innovation by reducing competitive pressure, suppressing heterodox research directions that don’t fit incumbents’ commercial roadmaps, and creating monocultures of approach.
The compute bottleneck is particularly instructive. NVIDIA’s data-center GPUs — the A100 and H100, and now their Blackwell successors — power the majority of serious AI training workloads, and demand has so dramatically outstripped supply that access to compute has become a significant determinant of which organizations can participate in frontier AI research at all. NVIDIA’s gross margins reached 74% in 2024 — a figure that reflects something close to monopoly pricing power. Meaningful AI regulation that increases demand for compute testing (which many safety frameworks require) could, paradoxically, further entrench NVIDIA’s dominant position.
The Safety Debate: Genuine Concern or Competitive Moat?
Few debates in technology have generated more heat with less light than the question of AI safety — specifically, who is genuinely motivated by safety concerns and who is using safety rhetoric instrumentally to constrain competitors.
The concern is not paranoid. Elon Musk co-founded OpenAI in 2015, fell out with its leadership, departed, and subsequently sued the organization — while simultaneously founding his own AI company, xAI, and arguing publicly that OpenAI’s safety concerns were either insufficient or pretextual. Sam Altman and OpenAI, meanwhile, have oscillated between presenting the company as a safety-first nonprofit mission and a commercial enterprise that restructured into a for-profit public benefit corporation in 2025 and was valued at $852 billion in 2026. These positions are not necessarily incompatible, but they are in significant tension.
The people who study AI safety as a serious academic discipline — researchers at institutions like Oxford’s Future of Humanity Institute (closed by the university in 2024), the Machine Intelligence Research Institute, and university AI safety labs — tend to be more cautious about assigning bad faith, but are often equally frustrated by the performative quality of big tech safety discourse.
“The problem is that ‘safety’ has become almost a meaningless term because it’s been stretched to cover everything from avoiding racist chatbot outputs to preventing existential catastrophe,” says Stuart Russell, professor of computer science at UC Berkeley and author of Human Compatible, one of the definitive texts on AI alignment. “Those are real concerns, but they’re on completely different timescales and require completely different interventions. Conflating them serves no one.”
There is a strand of safety argumentation that does function as a competitive moat. When well-resourced incumbents argue that only organizations with existing safety infrastructure should be permitted to train powerful models — and that smaller, open-source developers lack the capacity for adequate safety measures — the effect is to recommend regulatory frameworks that would exclude potential competitors. This is not to say the argument is wrong. It may be correct. But the alignment between the policy position and the commercial interest should invite scrutiny rather than deference.
The open-source dimension is particularly contentious. Meta’s release of the Llama model family, now downloaded hundreds of millions of times, has made powerful language model capabilities available to researchers, startups, and governments around the world. Proponents argue this democratizes AI and prevents dangerous concentrations of capability. Critics — including some prominent AI safety researchers — argue that unrestricted release of powerful models makes it trivially easy for malicious actors to fine-tune them for harmful purposes, removing safety guardrails that closed systems maintain.
Geopolitics as the Ultimate Regulator
Any discussion of who controls the pace of AI development that omits China is fatally incomplete. The geopolitical dimension of AI may be the most powerful throttle mechanism of all, because it creates pressures that operate in both directions simultaneously: accelerating development in the name of national security while simultaneously motivating restrictions on technology transfer that constrain global progress.
The United States has imposed sweeping semiconductor export controls targeting China’s ability to access advanced AI chips. The October 2022 rules, significantly tightened in 2023, prohibited export of NVIDIA’s most advanced chips to Chinese entities without licenses, and imposed strict controls on semiconductor manufacturing equipment. The intended effect was to create a persistent technological deficit in Chinese AI capability relative to American frontier models — though in 2025 the Trump administration began licensing some NVIDIA chip sales to China in exchange for a share of the revenue.
The policy is working — imperfectly. Chinese companies including Baidu, Alibaba, and Huawei have accelerated domestic chip development, and while China’s best chips remain roughly two to three generations behind NVIDIA’s frontier offerings, the gap is narrower than American policymakers initially projected. In January 2025, the release of DeepSeek R1 — a Chinese reasoning model that matched or exceeded leading American models on several benchmarks while reportedly requiring a fraction of the compute — sent shockwaves through the American AI industry and validated fears that chip restrictions alone cannot contain Chinese AI advancement indefinitely.
DeepSeek’s emergence also complicated the domestic policy debate in the United States. If Chinese researchers can achieve comparable results with dramatically less compute, the assumption that raw computational scale is the primary determinant of AI capability is weakened — which in turn weakens arguments for regulatory frameworks built around compute thresholds. Biden-era measures, including the 2023 executive order and a Commerce Department rule proposed in 2024, had tied mandatory government reporting requirements to model training runs above certain compute levels.
Looking Forward: Who Gets the Final Vote
The regulatory landscape for AI remains characterized less by settled frameworks than by active contestation. The Trump administration revoked the Biden order on its first day in January 2025, emphasized American AI dominance over safety caution, and in December 2025 directed federal agencies to challenge state AI laws. In Europe, the AI Act’s rules for general-purpose models took effect in August 2025, but a simplification package that entered into force in July 2026 pushed its main high-risk obligations back to December 2027 and August 2028. State-level legislation continues in the absence of federal action — California’s and Texas’s AI laws took effect in January 2026 — creating compliance complexity that some analysts expect will eventually force congressional action simply to reduce the operational burden on companies operating across multiple jurisdictions.
What seems increasingly clear is that the meaningful decisions about AI’s pace are being made through a combination of mechanisms that traditional regulatory analysis struggles to capture: the investment decisions of a small number of sovereign wealth funds and tech giants; the export control policies of two competing superpowers; the chip allocation priorities of a semiconductor company with de facto monopoly power; and the internal governance processes of a handful of AI labs whose leadership structures have proven, as the OpenAI board crisis of November 2023 vividly demonstrated, to be fragile and contested.
The five-day saga in which OpenAI’s board fired and then rehired Sam Altman was a stress test of AI governance conducted in public, and the lesson many drew from it was dispiriting: when a nonprofit board attempted to exercise precisely the kind of independent oversight over a powerful AI developer that safety advocates have been demanding, it was overwhelmed within a week by the combined pressure of Microsoft, most of OpenAI’s employees, and the commercial momentum of a $13 billion investment relationship.
That may be the most honest summary of where AI governance stands today. The mechanisms that could meaningfully modulate the pace of AI development exist, at least in embryonic form. But the political economy surrounding them — the competitive pressures, the national security imperatives, the investment flows, the revolving door between industry and regulators — consistently favors acceleration over restraint. The gatekeepers are real. Whether they are actually guarding anything is a question that no executive order, however lengthy, has yet answered.