Redistributed

Paper: Anton Korinek and Jai Vipra. "Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence." Economic Policy, January 2025. Earlier versions: "Market Concentration Implications of Foundation Models: The Invisible Hand of ChatGPT," Brookings Institution, 2023. Links: Economic Policy | NBER Working Paper | Brookings

Why This Paper Matters

Five companies control the development of frontier AI. Why five? Why not fifty, or two, or one? And will that number grow or shrink?

Korinek and Vipra provide the economic framework to answer these questions. Where much AI discourse treats concentration as background fact or assumes tech markets naturally consolidate, this paper shows the mechanisms — the specific cost structures, scale advantages, and strategic behaviors that push AI toward oligopoly.

This matters for redistribution directly. Market structure determines who captures value when AI creates it. An industry with many competitors distributes gains differently than an industry dominated by a handful of vertically integrated giants. Understanding why AI concentrates tells us where to intervene.

The paper is also notable for what it doesn't do: declare concentration inevitable. The authors identify structural pressures toward monopoly, then map the policy choices that could maintain competition. Economics is not destiny — it's a field of forces that policy can redirect.

The Argument

The Cost Structure of Foundation Models

The fundamental economics: training frontier AI models requires massive fixed costs and near-zero marginal costs.

Training GPT-4 or Claude 3 Opus costs somewhere between $50 million and $100 million in compute alone — before researcher salaries, failed experiments, or data acquisition. These are fixed costs: you pay them once, regardless of how many users eventually query the model.

Serving those queries is cheap. Once trained, running inference on a foundation model costs fractions of a cent per request, and that cost falls as hardware improves. The marginal cost of serving one more user approaches zero.

This cost structure creates powerful economies of scale. Average costs decline as deployment grows. A company that serves 100 million users has lower per-user costs than one serving 10 million. The larger player can offer lower prices, capture more users, and reduce costs further. The advantage compounds.

Korinek and Vipra identify a parallel effect: economies of scope. It's cheaper for one company to produce multiple foundation models — different sizes, different modalities, different domains — than for multiple companies to specialize. The infrastructure, talent, and data pipelines built for one model serve others. A company with a foundation model for text can extend to image and audio more cheaply than a newcomer can build any of them from scratch.

The Barriers That Compound

Beyond cost structure, the authors identify resource constraints that reinforce concentration:

Compute access is controlled by a handful of cloud providers — AWS, Azure, Google Cloud — who are themselves investors in the leading AI labs. Nvidia supplies roughly 95% of frontier training chips (see Hardware Bottleneck). These chokepoints mean that even companies with capital face allocation constraints. Compute isn't something you buy off the shelf; it's something you negotiate access to, often from a competitor. AI Chip Export Controls add a geopolitical layer to this bottleneck.

Data advantages compound over time. The labs that trained on early web data captured something irreproducible — you cannot retrain on 2015's internet. Proprietary data from user interactions (queries, corrections, preferences) flows back to model producers, improving each generation. The more users a model has, the more feedback it collects, the better the next model becomes.

Talent concentration limits competition. Perhaps a few hundred researchers worldwide have led frontier training runs. They command million-dollar salaries and concentrate in a handful of organizations. Knowledge about what works and what doesn't transfers through people, not papers.

First-Mover Dynamics and Market Tipping

Korinek and Vipra distinguish frontier models from behind-the-frontier models. The dynamics differ substantially.

For the most capable models at any given moment, first-mover advantages are severe. Being first to market allows capturing users, generating feedback data, building distribution relationships, and amortizing fixed costs across a larger base. Followers face a moving target: by the time they match capabilities, the leader has moved on.

This creates risk of market tipping — the possibility that early advantages lock in permanent dominance. If one company achieves sufficient lead in capability and scale, competitors may find the gap uncloseable. The market tips to monopoly, not through conspiracy, but through compound advantages.

Behind-the-frontier markets are different. Competition among open-weight models, among smaller specialized systems, among fine-tuning services — here, barriers are lower and competition more robust. The Rise of Small Language Models and projects like EleutherAI and Ollama illustrate this more competitive layer. The authors note this distinction matters for policy: frontier markets need active intervention; behind-the-frontier markets may self-correct.

Vertical Integration Risks

The paper's sharpest policy concern: vertical integration threatens competition in downstream applications even if the model market itself has multiple players.

The five frontier labs don't just train models. OpenAI builds ChatGPT; Anthropic builds Claude applications; Google integrates Gemini into Search, Workspace, Cloud. Meta deploys AI across Facebook and Instagram. The model producers are also the dominant applications.

Worse, the cloud providers who sell access to compute are investors in the labs who need that compute, and competitors in applications built on those models. Microsoft invested $13 billion in OpenAI, provides OpenAI's compute through Azure, integrates OpenAI's models into Office and Bing, and competes with products built on OpenAI's API. The layers blur.

Korinek and Vipra warn that producers could "expand their market power vertically to downstream uses, in which competition would otherwise ensure lower prices and better services." A company that controls both the foundation model and the leading applications can preference its own products, access competitors' usage data, adjust pricing to squeeze independent developers, and set standards that entrench its architecture.

This is Lina Khan's Amazon playbook applied to AI. In "Amazon's Antitrust Paradox" (2017), Khan showed how Amazon's control of marketplace infrastructure gave it information and leverage over merchants who were also competitors. The dual role — platform and participant — created conflicts that consumer welfare analysis couldn't capture. The same logic applies to foundation model providers who also build applications.

The Cloud Provider Nexus

The paper pays particular attention to cloud providers as concentration amplifiers.

AWS, Azure, and Google Cloud control the infrastructure where models train and run. They are also the primary investors in AI labs: Microsoft in OpenAI ($13 billion), Google in Anthropic ($2 billion), Amazon in Anthropic ($4 billion). And they are competitors in AI applications.

This creates layered dependencies. An independent lab needs cloud compute. The cloud provider may also be an investor, a competitor, and a potential acquirer. Negotiating compute access occurs in shadow of these relationships. The cloud providers can observe what's being built on their platforms and decide where to compete.

Korinek and Vipra don't call this structure anticompetitive by design, but they note it creates conditions where anticompetitive behavior is easy and detection difficult. Information flows that would be illegal if explicit can occur implicitly through infrastructure access patterns.

How AI Differs from Previous Tech Markets

The paper implicitly contrasts AI with earlier tech concentration — search, social media, e-commerce. Several differences emerge:

Scope is broader. Google dominates search; Facebook dominates social networking; Amazon dominates e-commerce. Each is a specific market. Foundation models, by contrast, are general-purpose infrastructure applicable across the entire economy. The stakes of concentration are correspondingly higher.

Capital requirements exceed prior waves. You could start a search engine or social network in a dorm room. You cannot train a frontier AI model without hundreds of millions in compute. The entry barrier is not just technical capability but capital access at nation-state scale.

Speed of capability growth. AI capability has grown faster than previous technologies, compressing the timeline for establishing dominance. By the time regulation catches up to current capabilities, the frontier has moved.

Safety as a competitive variable. Unlike prior tech markets, AI development involves genuine safety considerations. A company that invests heavily in safety testing moves slower than one that doesn't. The competitive dynamic can pressure everyone toward less caution. Korinek and Vipra note that monopolists may actually invest more in safety than competitive markets would, because they internalize downstream risks — a rare case where concentration might have beneficial effects.

Policy Interventions

The authors propose a two-pronged regulatory strategy.

Maintaining Contestability

The first prong addresses market structure directly:

Prevent vertical foreclosure. Model providers should not be allowed to disadvantage downstream competitors through preferential access, discriminatory pricing, or information advantages. The vertical integration between cloud providers, model developers, and applications requires scrutiny.

Monitor strategic behavior. Predatory pricing, anticompetitive lobbying, acquisition of nascent competitors — these require active enforcement attention. The authors note that competition authorities should watch for strategies that deter entry rather than waiting for monopoly to consolidate.

Ensure downstream competition. Even if the model layer concentrates, competition in applications built on models can distribute value. This requires preventing model providers from foreclosing that competition.

Quality Regulation Where Competition Fails

The second prong acknowledges that frontier models may concentrate despite intervention. If monopoly or tight oligopoly is structural, regulation should address the consequences:

Treat leading model providers like public utilities. This means mandated standards for safety, privacy, non-discrimination, reliability, and interoperability. Reasonable pricing, quality guarantees, and access requirements would constrain what concentrated producers can do.

Require interoperability. If switching costs create lock-in, mandated interoperability reduces them. Standards for API compatibility, data formats, and model outputs would allow users to move between providers.

Mandate disclosure. Training data composition, evaluation results, incident reports, capability assessments — transparency requirements enable oversight when competition doesn't provide discipline.

Level the regulatory field. AI and non-AI solutions should face identical requirements. If discrimination in hiring is illegal, it should be illegal whether a human or an algorithm does it. Regulatory asymmetry that advantages AI applications should be eliminated.

What the Paper Doesn't Propose

Notably absent: breakup remedies or prohibitions on large-scale development. Korinek and Vipra don't argue that concentration should be prevented by fragmenting existing players or capping investment. They accept that economies of scale are real and that frontier development may require large organizations.

This reflects a pragmatic assessment. The genie is out of the bottle. The question isn't whether to have concentrated AI development but how to constrain its harms while preserving its benefits. Whether this pragmatism is realistic or accommodation depends on your theory of change.

The Redistribution Lens

The paper's framework maps directly to redistribution questions.

Who captures value: Market structure determines how AI-generated value divides among producers, users, workers, and society. Concentrated markets give producers pricing power; competitive markets push value to users. Vertical integration lets producers capture value across the stack rather than sharing with independent developers.

Where interventions work: The paper identifies specific chokepoints — compute, data, talent, vertical integration — where policy could shift distribution. These aren't the only points, but they're economically grounded rather than speculative.

What shapes outcomes: Korinek and Vipra treat concentration as emergent from choices about investment, policy, and corporate strategy — not as technologically determined. This matters because it identifies agency. If concentration is inevitable, we can only manage consequences. If it results from choices, different choices remain possible.

International dimensions: The paper touches only lightly on geopolitics, but the framework extends. AI concentration isn't just a domestic competition issue. It's also about which countries develop frontier capabilities, who controls critical infrastructure, and how export controls and national strategies interact with commercial dynamics.

Engaging the Debate

The Khan-Wu Framework

The paper resonates with the "New Brandeis" approach to antitrust associated with Lina Khan and Tim Wu. Khan's work on Amazon showed how platform power creates conflicts that consumer welfare standards miss. Wu's The Curse of Bigness (2018) argued that concentration threatens democratic governance, not just prices.

Korinek and Vipra's analysis supports this view. Foundation model markets exhibit exactly the dynamics Khan identified: infrastructural control, information asymmetry, dual roles as platform and competitor. The consumer welfare framework — focused on prices — cannot capture harms from foreclosed competition, reduced innovation, or accumulated leverage.

But the paper also reflects limits of the New Brandeis moment. Khan's FTC brought ambitious cases but achieved mixed results. The courts haven't fully adopted the expanded view of antitrust harm. Korinek and Vipra's policy recommendations require either judicial evolution or new statutory authority — neither of which is assured.

The EU Approach

European competition authorities have moved more aggressively. The Digital Markets Act (DMA) designates large platforms as "gatekeepers" subject to structural requirements — interoperability mandates, prohibitions on self-preferencing, data access rules. The European Commission is examining whether to extend DMA to AI.

Korinek and Vipra's analysis suggests AI fits the gatekeeper framework. Foundation model providers control access to capabilities that downstream services depend on. The vertical integration concerns the paper identifies are precisely what DMA was designed to address.

The practical question is enforcement. The DMA designated gatekeepers in 2023; compliance and enforcement continue. Whether the framework can extend to AI's faster-moving, technically complex domain remains uncertain.

The Innovation Counterargument

Not everyone accepts that concentration harms innovation. A contrary view: large-scale AI development requires large organizations with access to capital, talent, and infrastructure. Fragmenting that capacity would slow progress, reduce safety investment, and leave development to less responsible actors or foreign competitors.

Korinek and Vipra don't dismiss this entirely. They acknowledge that some concentration may be efficient and that monopolists can internalize safety externalities that competitive firms cannot. Their argument is not for fragmentation but for contestability and regulation — accepting concentration's benefits while constraining its harms.

Whether this balance can be achieved is the central policy uncertainty. Effective antitrust requires identifying anticompetitive behavior in technically complex, fast-moving markets where regulators have less expertise than participants. Utility regulation requires capacity that most governments don't currently have. The framework is sound; implementation is hard.

What This Means for You

If you're building on AI platforms: Understand the dependency. Your business relies on infrastructure controlled by potential competitors. Watch for behavior changes — pricing, API terms, capability access — that squeeze independent developers. Build with portability in mind where possible.

If you're setting policy: The paper provides economically grounded rationale for intervention. The concentration isn't speculative; it's structural. The harms aren't theoretical; they follow from identified mechanisms. The question is not whether to act but how.

If you're tracking power in AI: The paper maps where control concentrates: compute, data, talent, vertical integration, cloud provider relationships. These are the chokepoints to watch. Who controls them determines who shapes AI development.

If you're working toward alternatives: The paper identifies what alternatives must overcome: not just technical capability but the compound advantages of scale, data, and infrastructure. Open-weight models help but don't address training compute. Public infrastructure could matter but requires sustained investment governments haven't committed. The Open (For Business) paper by Widder et al. examines how corporate "openness" can reinforce rather than challenge concentration. But projects like BLOOM, Hugging Face, and Lelapa AI suggest that alternative development models, while constrained, are not impossible.

Source

Korinek, Anton and Jai Vipra. "Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence." Economic Policy, January 2025. Economic Policy

Earlier versions:

"Market Concentration Implications of Foundation Models: The Invisible Hand of ChatGPT." Brookings Institution, September 2023. Brookings
NBER Working Paper #33139, November 2024. NBER

General

Paper Review: AI Market Concentration