Concentration is a stack, not one market share
AI power is distributed across semiconductor design, chip fabrication, advanced packaging, data centers, cloud platforms, energy, data, foundation models, application distribution, and skilled labor. A firm can face competition at one layer while depending on a chokepoint at another. Counting chatbot brands therefore does not answer whether the underlying system is concentrated.
Today’s evidence concerns frontier generative AI and its supply chain. The claim that one laboratory will control AGI is a scenario, not an observed result. It nevertheless deserves analysis because control of unusually general, economically valuable systems could amplify existing advantages.
Why frontier development favors scale
Training and serving leading models can require large capital commitments, scarce chips, power, networking, and specialized teams. Cloud providers can spread those costs across customers and offer developers immediate infrastructure. Model developers benefit from distribution, user feedback, and revenue that can finance the next generation. These effects can create barriers to entry even when algorithms are published.
The pattern is not inevitable. Algorithmic efficiency can lower costs; smaller models can compete in specialized tasks; open weights can enable local deployment; and customers can switch among APIs if interfaces remain portable. Concentration should therefore be measured rather than assumed. Useful indicators include cloud share, accelerator supply, training-compute access, model-performance gaps, switching costs, exclusive contracts, interoperability, talent movement, and the share of downstream products dependent on a few providers.
Stanford’s 2026 AI Index reports that global corporate AI investment more than doubled in 2025 and that use spread broadly across organizations (Stanford HAI, 2026). Investment totals show scale and momentum, not by themselves monopoly power. Industry-funded surveys may also overrepresent large adopters, so national statistics and procurement data remain important.
Corporate power goes beyond prices
A general-purpose model provider can shape which languages, viewpoints, and activities its systems handle well; which users receive access; what data is retained; and which downstream developers bear liability. These choices can affect education, media, employment, health, and government procurement simultaneously.
Traditional competition harms still matter: tying model access to a cloud, preferential treatment of affiliated applications, exclusive capacity deals, acquisitions of potential rivals, discriminatory API terms, and high egress or switching costs. But price alone may miss governance power when a subsidized service becomes infrastructure.
Concentration can sometimes improve safety. A provider operating a controlled API can monitor abuse, patch safeguards, revoke access, and finance evaluations. Fragmentation can make accountability difficult. The opposing risk is that a small group writes its own rules, withholds information, or becomes too important for governments to discipline. The policy problem is to preserve accountability and redundancy without assuming that either centralization or openness is automatically safe.
Cloud, chips, and energy create separate chokepoints
Leading accelerators depend on a global supply chain with geographically concentrated fabrication and packaging. Export controls demonstrate that governments view access to these inputs as strategic. Data centers also depend on local grids, water, land, permits, and long-lead equipment. Communities may bear infrastructure and environmental costs while benefits flow elsewhere.
Compute governance can exploit physical traceability, but it may also entrench incumbents if compliance costs fall more heavily on smaller firms. Public policy should ask who can obtain lawful compute, whether universities and startups have access, and whether reporting thresholds are capability-relevant rather than a fixed moat.
State power and corporate-state interdependence
States license exports, fund research, purchase cloud services, control energy and telecommunications rules, and conduct intelligence operations. Firms possess technical expertise and infrastructure governments may lack. That creates mutual dependence rather than a simple story of companies replacing states.
National-security partnerships can improve resilience and oversight, but secrecy can weaken democratic accountability. Governments may use AI for surveillance, targeting, border control, policing, or information operations. Concentrating capability in democratic states does not by itself ensure democratic use; procurement law, judicial review, civil-rights protections, inspectors general, and legislative oversight remain necessary.
Export controls can slow competitors but may also accelerate indigenous supply chains, fragment standards, and intensify race narratives. Those are possible second-order effects, not certain outcomes. Evidence should distinguish stated policy goals from measured results.
Data and labor are sources of power too
Models depend on creative work, public information, licensed databases, user interactions, and human feedback. Disputes over consent, copyright, compensation, and privacy concern who captures value from those inputs. A competitive model market can still rest on unequal bargaining between developers and creators or workers.
Concentration affects labor inside the industry as well. A small employer set can make it harder for safety researchers to leave after a disagreement without sacrificing specialized careers or equity. Enforceable whistleblower protections, limits on retaliatory nondisclosure terms, and portable credentials increase the independence of insiders.
Open weights complicate the picture
“Open source” should not be used casually when only model weights are available. Training data, code, documentation, and license rights may remain closed. Widely available weights can reduce dependence on hosted APIs, enable independent research, support privacy-preserving local use, and broaden innovation. They can also allow safeguards to be modified and capable systems to proliferate beyond recall.
The U.S. NTIA’s 2024 review found insufficient evidence at that time to justify restricting widely available model weights categorically, while recommending ongoing monitoring and the capacity to respond as capabilities changed (NTIA, July 2024). That finding is dated and explicitly conditional, not a permanent verdict.
Governance options and tradeoffs
No single intervention addresses every layer. Competition enforcement can scrutinize exclusionary conduct and mergers. Interoperability and data portability can reduce switching costs. Public or shared compute can broaden research access. Procurement can require audit rights, incident reporting, security, and exit plans. Public-interest research funding can reduce dependence on lab agendas. Corporate-governance rules and external evaluation can constrain self-regulation.
Structural separation may be appropriate in some settings, but it can sacrifice integration efficiencies and does not solve global chip concentration. Licensing can impose safety duties, but poorly designed licensing can protect incumbents. Public ownership can align infrastructure with social goals, but it still requires competent, rights-respecting administration.
Conflicts of interest should be visible. Frontier companies have incentives to favor rules they can afford and that burden smaller competitors. Open-model companies benefit from arguments for openness. Cloud and chip firms benefit from expanding compute demand. Safety organizations, labor groups, creators, civil-rights advocates, and national-security agencies also bring distinct mandates. Disclosure does not settle the argument; it helps readers evaluate it.
What AGI could change
If a system could automate a large share of research, management, persuasion, and software production, the owner might compound advantages faster than ordinary competition or government processes can respond. That is a theoretical mechanism, not proof that recursive dominance will occur. Bottlenecks, leakage, regulation, coordination, and diminishing returns could limit it.
The prudent objective is contestable, accountable capacity: no single failure should disable essential services; no provider should be beyond lawful oversight; researchers and users should have meaningful alternatives; and the public should share in gains built on collective inputs. Power concentration is a risk because of incentives and dependency, not because every large organization is malicious. Measuring the stack and governing each chokepoint is more credible than trusting either market competition or institutional goodwill alone.