The AI Future May Be Decided by Who Gets Access

We used to imagine that AI development might be slowed by a single, catastrophic event: a Chernobyl-sized failure that would force governments and companies to stop and reconsider. That remains possible. But the more consequential intervention may be quieter. A powerful model could fail a safety test, be withheld from public release, or be withdrawn from a widely used service. Its developer could keep building more capable systems while offering the public a narrower version.

That would not be a rollback of AI itself. It would be a rollback of access. For the people and businesses building on these systems, the distinction could matter more than a temporary pause in research.

This is no longer purely speculative as a governing approach. The European Union requires providers of general-purpose models that pose systemic risk to evaluate and mitigate those risks and report serious incidents. California has enacted frontier-model transparency and incident-reporting requirements. These rules do not order a broad model rollback, but they establish a framework in which deployment decisions can be scrutinized and changed. AI developers’ own safety policies also contemplate delaying deployment until safeguards are adequate. European Commission^[1], California Governor’s Office^[2], OpenAI^[3]

A rollback has limits. A provider can restrict an API or remove a hosted model. It cannot reliably recall model weights that have already been downloaded and copied. And if a capability is genuinely dangerous, reserving it for a small circle of “trusted” institutions creates another question: who decides whom to trust? A safety rule that only the largest firms can afford to satisfy may protect the public while also protecting those firms from competition.

That competition risk deserves more precision than the claim that OpenAI and Anthropic are becoming an inevitable duopoly. Other companies are near the frontier, and open models remain part of the market. Stanford’s 2026 AI Index places several developers in a closely grouped top tier. Still, competition is about more than model scores. Compute, distribution, cloud contracts, talent, and the cost of complying with rules can determine who gets to build and who merely rents access. A 2025 Federal Trade Commission staff report identified ways major cloud partnerships could raise switching costs or restrict access to essential inputs. Stanford HAI^[4], FTC^[5]

That is the stronger antitrust concern. A breakup is a possible remedy only if evidence and law justify it; declaring one necessary now gets ahead of the case. Governments should first examine the actual bottlenecks: exclusive arrangements, access to compute, restrictions that prevent customers from switching models, and rules that make independent developers dependent on a handful of gatekeepers.

The same discipline is needed on jobs. One imaginable response to agentic AI is a cap on the share of employees a company may lay off because of automation. Its aim is understandable: give workers and communities time to adjust. But a fixed cap would be hard to administer. Firms rarely replace a neat percentage of a workforce for one stated reason. They could avoid hiring, outsource work, or label an AI-driven restructuring as something else. It could also freeze people in jobs that are changing instead of helping them move into better ones.

The need for a response is real even if that particular rule is weak. The International Labour Organization estimated in 2025 that one in four workers globally held a job with some exposure to generative AI, while finding that transformation was more likely than outright replacement for most jobs. Policy should make that transition less punishing: portable benefits, retraining tied to real opportunities, support for displaced workers, and ways for employees to share in productivity gains. ILO^[6]

It should also help people create new work. A grant for a small AI-native business will not compensate every person who loses a job, and entrepreneurship should never be prescribed as a universal cure for unemployment. But access to capital, practical training, public procurement, and affordable AI tools could let more people build companies around problems they know firsthand.

This is where we remain bullish. AI can give a solo founder or a small team capabilities that once required departments: research, prototyping, software development, design, translation, and operations. That does not make judgment, expertise, customer trust, or execution free. It makes them more valuable. A founder who understands a particular industry can use AI to test an idea faster and serve a narrow need that a giant platform may overlook.

The future is unlikely to divide neatly between unrestrained progress and a total stop. Governments may regulate dangerous capabilities, intervene in concentrated markets, and cushion changes to work, all while supporting AI development for economic and strategic reasons. The test is whether those choices keep the benefits open to new builders.

If the most capable AI becomes accessible only to governments and a few dominant companies, we will have managed one set of risks by creating another. The better outcome is safety rules proportionate to risk, markets that remain contestable, and a broad path for people to turn these tools into useful businesses. That is how AI’s innovation potential becomes more than a promise made by its largest developers.