Sam Altman Warns AI Power Could Consolidate Among a Small Few

Key Takeaways

• Sam Altman argues that concentrating advanced AI among a handful of companies, models, or individuals could weaken public control over the technology.
• The debate increasingly pits centralized safety oversight against open-weight models that businesses can modify and deploy independently.
• Enterprises can reduce dependency through model diversification, portable infrastructure, open standards, and formal AI governance.

OpenAI CEO Sam Altman is drawing attention to a risk that accompanies the technical dangers usually associated with artificial intelligence: the possibility that control over advanced systems becomes concentrated among a small group of companies, models, or people.

Speaking on David Senra's podcast, released Sunday, Altman said the preferred path should keep individuals involved in determining how AI develops and affects society.

“The right approach is to say we want people deeply in control of the future,” Altman said. “Society and the models are going to co-evolve.”

His concern reflects a practical issue for enterprises. Access to advanced AI already depends on a relatively narrow collection of model developers, cloud providers, semiconductor suppliers, and data center operators. A business might choose between several AI applications while still relying, several layers down, on the same infrastructure and foundation-model providers.

The concentration is measurable. Global cloud infrastructure spending reached about $330 billion in 2024, while Amazon Web Services held roughly 30% of the market, Microsoft Azure 21%, and Google Cloud 12%, according to Synergy Research Group figures reported by CIO Dive. Together, the three providers controlled more than 60% of the enterprise cloud market used to train and deploy many advanced AI systems.

Model development shows a similar pattern. Gartner reported that the generative AI models market expanded by more than 300% in 2024, with a small number of foundation-model providers capturing much of the commercial spending and adoption. OpenAI, Anthropic, and Google DeepMind are among the prominent developers operating on top of these infrastructure stacks.

Concentration does not automatically mean harmful conduct. Large-scale AI requires expensive computing capacity, specialized expertise, security controls, and extensive testing. Those demands naturally favor well-capitalized organizations. But dependence on a limited set of suppliers can still create risks involving pricing, service availability, data portability, contractual terms, and sudden changes to model behavior or access policies.

Altman also argued that fear of powerful AI could encourage society to accept excessive central control in the name of safety.

“There are a lot of people who are so nervous about the magnitude of those risks and get so taken by that and feel a need to protect the world,” he said. “Like, you know, ‘we should trade off a lot of liberty for safety.’”

That position contrasts with the more cautionary posture associated with Anthropic CEO Dario Amodei. In a June essay, Amodei said Anthropic's new Mythos models presented “very real risks” to cybersecurity, finance, critical infrastructure, and national security. Anthropic has advocated for a larger government role in AI oversight and was the only major AI company last month not to sign a letter supporting open-weight models.

The disagreement is not simply about whether AI presents risks. It is about who should manage them, and how much authority users, businesses, model developers, and governments should each hold.

Open-weight models are central to that argument because developers can download, modify, and deploy them on infrastructure of their choosing. China has embraced that approach, helping its developers compete with Silicon Valley at a fraction of the cost. In the United States, Nvidia, Meta, and Microsoft have argued that broader support for open-weight technology could help the country remain competitive. OpenAI released gpt-oss-120b and gpt-oss-20b last year, although it has not released the source code for its frontier models.

Would broader model access genuinely distribute power, or simply move control to the organizations that own the chips and data centers? In practice, it may do some of both. Open weights can improve portability and customization, but running large models still requires capital, technical staff, and computing resources.

Governance offers another route between unrestricted deployment and tight central control. The OECD has outlined common guideposts for accountability, transparency, fairness, and interoperability across national AI risk-management approaches. Enterprises can apply similar principles internally while maintaining multiple model options and testing whether workloads can move between providers.

Altman ultimately framed AI as a tool for distributing economic agency. “We are about to see the greatest boom in people starting smaller businesses than we have ever seen,” he said. “I think AI is empowering that.” Whether that boom materializes may depend on more than model capability. Affordable access, supplier choice, portable data, and room for businesses to experiment could determine whether AI broadens opportunity or concentrates it further.