Concentration vs Distribution: Who Benefits, Who Decides, Who Owns

The Question

Technologies have different distributional signatures. Some (like the internal combustion engine) mostly distribute: widely produced, widely used, widely owned. Some (like the atomic bomb) mostly concentrate: tightly held, rarely used, highly regulated.

AI's signature is being determined right now. The factors pushing in each direction are unusually strong. Which way the balance tips matters enormously for what kind of transition happens.

If benefits concentrate: a small number of people and firms get rich; most of the population experiences AI as something done to them, not with them.

If benefits distribute: the productivity gains reach most people; AI functions like past infrastructure, extending human capability broadly.

Both outcomes are consistent with current trajectories. Which one happens depends on choices being made in the next few years.

What Concentrates

Several forces push toward concentration:

Compute

Training frontier models requires enormous compute. A GPT-4-class model costs hundreds of millions to train; GPT-5-class and beyond are in the billions. Only a small number of firms can afford this.

The cost isn't declining as fast as capabilities grow. Even if inference costs drop (they have), training costs for the frontier keep rising. Frontier AI is a specialised industrial activity, not a garage workshop product.

Data

Large high-quality datasets are themselves valuable and concentrated. Data licensing deals (OpenAI with News Corp, various labs with publishers) reflect this. Public data scraping is increasingly contested in courts.

Talent

The number of researchers who can design and train frontier models is small. They're concentrated at a small number of labs, which bid up their salaries. This makes frontier research expensive and slow to diversify.

Regulatory attention

As AI becomes consequential, governments pay attention. Regulatory compliance is expensive, which favours incumbents who can afford legal teams. The EU AI Act, various proposed US federal rules, and emerging regulation in other countries all advantage large firms.

Network effects

The more people use a model, the better the signals for improvement, the more valuable the model. Leading labs' advantages compound through user interaction.

Integration

Large firms can integrate AI into existing products (Microsoft with Office, Google with Workspace). Integration produces stickier usage. Upstart competitors have to build everything.

What Distributes

Several countervailing forces:

Open source

Open-source models (Llama, Mistral, DeepSeek, various) have narrowed the capability gap with closed frontier models. A company or researcher can download a highly capable model and run it on their own hardware.

The gap between open and closed is debated. Open models lag the frontier, but not by much at many common tasks. The lag may be narrowing or widening depending on which metric you pick.

Inference cost

Running (inferencing) a model is much cheaper than training it. As inference costs drop, the barrier to using AI falls dramatically. A startup can build on top of OpenAI, Anthropic, or open-source models without training their own.

Tooling

Frameworks (Hugging Face, LangChain, LlamaIndex, Ollama), developer tools, and API ecosystems make AI easier to apply. This enables smaller actors to build products.

Geographic diversification

Labs are emerging outside the US: DeepSeek, Zhipu, and 01.AI in China; Mistral and Aleph Alpha in Europe; various in Canada and UK. The frontier is less geographically concentrated than it briefly was.

Regulatory divergence

Different jurisdictions are taking different approaches. The EU's AI Act, China's AI regulations, various US states, the UK's post-Brexit approach: diversity here prevents any single regulatory regime from locking in a particular concentration pattern.

User agency

Some uses of AI distribute power directly: education, health literacy, legal aid, access to professional knowledge previously gated by cost. A person in a poor country with a phone and a free-tier AI has access to advice previously unavailable.

Historical Comparisons

Useful to consider how past general-purpose technologies shook out:

Electrification

Started concentrated (a few industrialists built the grid). Eventually distributed widely (everyone has electricity). Took decades. Required regulation, public investment, and political choices.

The internet

Started as military, then academic, then broadly consumer. Distributed enormously in its first two decades, enabling many small actors. Then concentrated again in the 2010s around platforms (Google, Meta, Amazon). The current concentration is the result of specific economic dynamics (data, network effects) and specific regulatory choices.

Previous AI waves

Narrower deep-learning-era AI (2012-2022) produced substantial concentration: a few firms dominated translation, search, recommendation. Current wave is similar in structure, maybe larger in stakes.

The pattern: distribution doesn't happen automatically. It requires specific conditions: open standards, antitrust enforcement, public investment, and technical features that enable many participants.

The AI Labs

The most-discussed concentration: the labs actually building frontier AI.

As of 2026, the short list includes:

• OpenAI: leading commercial frontier, closely tied to Microsoft, capped-profit structure
• Anthropic: Claude family, public-benefit corporation, positions itself as safety-focused
• Google DeepMind: substantial internal investment, Gemini family
• Meta: Llama family; bets heavily on open-source releases
• xAI: Musk's lab; Grok family; less transparent than peers
• DeepSeek, Zhipu, 01.AI, Moonshot: Chinese labs at or near frontier
• Mistral: French lab, mixture of open and closed models
• Smaller labs: Cohere, Inflection (absorbed into Microsoft), AI21, others

A handful of firms account for most frontier training. If you want to understand who's making decisions, this is roughly the list.

Power Alongside Benefits

Distribution isn't only about economic benefits. It's also about:

Decision power

Who gets to set the norms for AI use? Labs decide what their models will and won't do. Governments decide what's legal. Academic and civil-society voices weigh in. The balance among these is contested.

Data power

Who owns the data AI is trained on? Authors, publishers, artists are asserting rights over training data. Courts haven't fully resolved this. Outcomes will shape whether future training concentrates further or spreads.

Governance power

Who decides what AI is developed? Currently: the labs and their investors. Increasingly: governments, especially in safety-critical applications. Potentially: international bodies, though these are weak.

Epistemic power

AI mediates what people know. If most people use a small number of AI systems to answer questions, those systems shape shared understanding. This is a form of power that previously wasn't concentrated; current AI use is concentrating it. Newspapers had this once; search engines have it now; AI chat may have it next.

Arguments About Openness

A specific concentration-related debate: should frontier AI be open or closed?

For openness

• Distributes capabilities to many users
• Enables academic research, scrutiny, and development of safety tools
• Reduces reliance on a small number of firms
• Allows pluralistic development of applications

Advocates: Meta, many academics, Mistral, open-source community.

For closure

• Reduces proliferation of dangerous capabilities to bad actors
• Supports commercial investment in expensive training
• Enables safety interventions (refusing harmful requests)
• Allows coordinated governance

Advocates: OpenAI (partly), Anthropic (partly), some governments, some AI safety researchers.

This debate won't be resolved abstractly. It will be resolved through specific cases: what specific models get released with what specific capabilities under what specific conditions.

Both sides have serious arguments. The right answer may vary by capability threshold: fine to release current models openly, not fine to release hypothetical future ones that could meaningfully accelerate bioweapons development, for instance.

Capital and Equity

A distinct question from technical concentration: the financial returns of AI.

If AI massively boosts productivity, the returns go somewhere. To whom?

• To the workers using AI? (Some, depending on bargaining power)
• To the firms deploying AI? (Often, especially if productivity gains are captured as lower prices or higher wages)
• To the labs building AI? (Depends on moat; may be narrow)
• To owners of data used to train AI? (Some legal claims; mostly not yet)
• To owners of compute and infrastructure? (Significant; Nvidia is an example)

Distribution of AI's economic gains depends on labour market power, antitrust enforcement, tax policy, corporate structures, and the robustness of competition between AI providers. Many of these are active policy questions.

What Would Good Distribution Look Like

Serious thinkers have proposed various approaches:

• Open-source frontier models (Yann LeCun, Meta): continued capability diffusion
• Public-option AI (some policy proposals): government-funded AI as a commons
• Competition policy: prevent monopoly in the AI stack via antitrust
• Data rights: clear legal framework for training data compensation
• Windfall clauses: labs commit to share profits if they achieve transformative AI (OpenAI had a version; Anthropic has discussed)
• Progressive deployment: gate capabilities by demonstrated safety, widely accessible rather than narrowly held

No single proposal resolves distribution. A combination, if implemented, would push toward more distributed outcomes.

What Bad Distribution Looks Like

The failure mode to watch:

• One or two firms dominate frontier AI indefinitely
• Regulatory capture entrenches incumbents
• Benefits flow to capital and skilled workers; large majorities see stagnant or declining wages
• Governance concentrates in a small number of states
• Epistemic dependence deepens: most people route their thinking through a small number of systems

This isn't a prediction; it's a failure mode worth naming to notice whether we're heading toward it.

Common Pitfalls

"Concentration is inevitable." It depends on choices. Antitrust, open source, public investment, and consumer behaviour all affect the outcome

"Open source is always good." At some capability level, unrestricted release becomes dangerous. The right answer depends on capability, use case, and available safeguards. Neither blanket position is obviously right

"If the US doesn't do it, China will." This argument is used to justify avoiding safety precautions. It has real force and real limits. It shouldn't override every other concern; it shouldn't be dismissed either

"The market will sort it out." Sometimes. Network effects, data moats, and capital barriers can produce stable concentration without outside intervention. Assuming the market solves everything is a conclusion, not an argument

"Distribution means everyone owns GPUs." Not necessarily. Distribution can mean many people benefit through services built on top of shared infrastructure. Who benefits is what matters, not who owns what

Next Steps

Continue to 05-institutions.md for whether the institutions we have can absorb this.