Korinek & Vipra — Concentrating Intelligence: Scaling and Market Structure in AI

Summary

"Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence" is Anton Korinek (University of Virginia, Brookings, NBER) and Jai Vipra's (Centre for the Governance of AI) analysis of competition dynamics in the foundation-model market — the market for large language models such as GPT-4, Gemini, and Claude.[1] It circulated first as an INET working paper (No. 228, October 2024, under the title "Market Concentration Implications of Foundation Models: The Invisible Hand of ChatGPT" in its earliest draft) and an NBER working paper (w33139), and has since been peer-reviewed and published in Economic Policy 40(121), 225–256 (2025).[1] It is the direct successor, on the diagnosis side, to Korinek & Ng's earlier digital-superstars model and to the Furman Review's 2019 warning that if the next technology wave centers on AI, "the firms best placed to exploit it may well be the existing large companies because of the importance of data" — a prediction this paper checks against 2023–24 evidence.

The paper's contribution to the wiki's rent gradient is a fresh, granular answer to what specifically would make AI profit a rent rather than a competitive return: extreme economies of scale in compute, a near-monopoly supplier of the physical input those economies run on, and data/talent bottlenecks that compound first-mover advantage. But the paper is explicit that, as of late 2024, none of this had yet produced a durable winner — the market it describes is fiercely, even destructively, competitive.

What It Establishes — the Diagnosis

The cost structure is a classical natural-monopoly setup. Foundation models combine a large fixed cost for pre-training (Epoch (2024) estimates Gemini Ultra's 2023 training run at $130 million) with near-zero marginal cost per query — "since fixed costs are high and variable costs relatively low, foundation models offer a classical example of economies of scale," compounded by economies of scope because one model serves many industries at once.[1] Training compute for frontier systems has grown 4.1x per year for fifteen years (Epoch AI data reproduced in the paper), a trend the authors say could produce "trillion-dollar foundation models" by the end of the decade if it continues.[1]

The compute layer is genuinely, not just metaphorically, concentrated. The chips that train and run foundation models are supplied almost entirely by one company: the paper's own Figure 3 puts Nvidia at 92% of the GPU market, and it cites a February 2024 Wells Fargo estimate of 98% of the data-center GPU market specifically.[1] This is the paper's clearest, least-contestable rent-adjacent fact — a bottleneck input controlled by a single firm sits upstream of every foundation-model producer's cost base.

Talent is price-inelastic in the short run, compounding entry costs. Because expertise in frontier models takes years to acquire and because engineers cluster where the compute is, "new entrants to the market often have to hire people who already work with market leaders and pay a premium to induce them to switch" — a labor-market echo of the same increasing-returns story.[1]

"Monopoly rents" is the authors' own term for the tipping scenario. Describing the historical playbook of digital platforms (heavy early losses, a shakeout, a durable winner), the paper writes: "the losses became unsustainable and led to a shake-out whereby a single platform or a small number of platforms survived, became the dominant players, and started to earn significant monopoly rents. The expectation of these profits was what led to the large number of entrants."[1] The paper argues foundation models share the mechanisms that produced that outcome for platforms — economies of scale and scope, data feedback loops, user inertia — plus one AI-specific channel with no platform analogue: an "intelligence feedback loop," in which a lab with better internal models becomes faster at building the next model, potentially "leave[ing] the competition ever further behind" and, in the limit, producing what I.J. Good (1965) called an "intelligence explosion."[1]

Vertical integration is already underway and compounds the concentration risk. The paper documents Google DeepMind's in-house TPU chips, Microsoft's and Amazon's in-house AI chip programs, Microsoft's cloud exclusivity over OpenAI, and Google Cloud's preferred-provider status for Anthropic, plus 2024 "acquisitions-by-proxy" — Microsoft hiring Inflection AI's founders and staff via a $650 million licensing deal rather than a formal acquisition, which the authors flag as "expected to lower [Inflection's] valuation and has drawn attention from antitrust authorities since it may represent a take-over in disguise."[1] Four antitrust authorities — the US FTC and DOJ, the UK CMA, and the EU Competition Commissioner — issued a rare joint statement in July 2024 naming concentrated control of compute/data/talent inputs, market-power extension, and anticompetitive partnerships as shared concerns, explicitly motivated by "a sense that competition authorities were not sufficiently proactive" the last time this happened with digital platforms.[1]

But the Paper Is Equally Clear the Market Was Not Yet Tipped

The honest complication for the rent reading is that the paper's own snapshot of the market — September 2024 — describes fierce, ongoing competition, not a settled winner. Fourteen companies had matched or exceeded the original GPT-4's capability within a year of its release; leading models were "clustered quite closely together" on the LMSYS benchmark (a 52% win probability for the top model against the second-ranked one); and "prices charged by the leading AI labs barely seem to allow them to cover their variable costs... the competition dynamics seem close to Bertrand competition" — the textbook zero-economic-profit outcome, the opposite of a rent.[1] The paper's own framing is conditional and forward-looking: it identifies the economic forces that could produce concentration and monopoly rents, and recommends proactive policy precisely because the authors think regulators moved too late last time — not because they conclude the rent has already arrived.

Policy remedies target the mechanism, not the profit. The paper's own menu is dissolve-the-moat, in the same register as the Furman Review and the DMA: mandated data sharing to blunt data feedback loops (with an explicit warning that this can cut against privacy regulation such as GDPR and can itself increase incumbents' data power if not designed carefully); policies that reduce switching costs (aggregator subscriptions, common API standards); disclosure of model architectures and public research investment to slow the intelligence feedback loop; stronger ex-ante merger review given how fast intellectual property can move before an acquisition can be unwound; and, as foundation models become more utility-like, non-discrimination requirements modeled on public-utility law.[1] No rent tax, data dividend, or ACE/DBCFT-style capture instrument appears in the paper's own remedy list — this is a pure-dissolve diagnosis paper, structurally the AI-market analogue of the Furman Review.

Relation to the Georgist Case — Both Ways

• The rent reading gets a genuinely new mechanism, not a restatement of the platform case. Unlike ordinary platform network effects, the paper's "intelligence feedback loop" describes a channel where the first-mover advantage compounds through the technology itself — better models make better models faster — which is a stronger tipping mechanism than anything in the older superstar-firms literature. And the near-monopoly compute layer (Nvidia at 92–98%) is a rent sitting upstream of every foundation-model firm, whether or not any individual AI lab ever earns one downstream — a distinct, verifiable rent claim that does not depend on resolving whether OpenAI's or Anthropic's own margins are rent or quasi-rent.
• The efficiency/competition counter is carried by the same paper, not a rival source. The Bertrand-competition observation — labs pricing near variable cost, a proliferation of near-substitute models, six-monthly leadership churn — is the paper's own evidence that as of the data available, the tipping the authors worry about had not yet happened. That is an unusually clean instance of the wiki's rent-gradient discipline: the same authors who supply the strongest mechanism for AI rent also supply the strongest evidence it has not yet materialized as realized profit.
• The gradient-honest synthesis. Read together with Korinek & Stiglitz (2017) — which theorizes that AI's long-run distributional gravity pulls toward whatever factor stays irreproducible — this paper supplies the short-run, empirical companion: a specific, checkable account of which inputs (compute above all) are concentrated today, and a call for proactive policy precisely because market tipping, once it happens, is expected to be hard to reverse. Neither paper claims current AI-lab profit already is rent at scale; both supply reasons a reader should expect the question to get more decidable, not less, over the coming years.

Limits

• A working paper turned journal article, not a competition-authority finding. Although now peer-reviewed (Economic Policy 2025), the paper is an academic analysis, not a market investigation with subpoena power or access to firms' internal cost data; its cost and market-share figures are drawn from public estimates (Epoch AI, Wells Fargo, Fernandez et al.) that the authors themselves flag as approximate.
• The 92%/98% Nvidia figures come from two different sources measuring slightly different things (the paper's own Figure 3, sourced to Fernandez et al. 2023, versus a Wells Fargo estimate specifically for data-center GPUs) — both point the same direction but should not be quoted interchangeably as a single precise number.
• The paper is explicitly a snapshot of September 2024 in a fast-moving market; the authors flag that "the technological landscape may change quickly," and the market structure should be re-checked against later evidence (e.g., subsequent DeepSeek, Grok, and open-source model releases) before this page's "not yet tipped" reading is treated as current.
• Does not itself evaluate tax instruments. This page covers the diagnosis only; for the authors' own view on how to tax AI rents if and when they materialize, see the companion paper Korinek & Lockwood, "Public Finance in the Age of AI", read separately for this wiki.
• Provenance. Findings and quotations verified against the INET working-paper PDF (No. 228, October 2, 2024 revision), fetched and read in full this session; the published citation (Economic Policy 40(121), 225–256, 2025) is taken from the paper's own reference list in the companion Korinek & Lockwood (2026) manuscript, which cites it as published.

See Also

• Taxing Tech Rents — Instrument Comparison — the graded synthesis this diagnosis feeds
• Korinek & Lockwood — Public Finance in the Age of AI — the same research program's instrument-design companion (compute/token/robot taxes, sovereign wealth funds, windfall clauses)
• Korinek & Stiglitz — AI and income distribution — the long-run theory this paper's short-run market evidence complements
• Korinek & Ng — digital superstars — the predecessor model of digitization-driven superstar rents
• Furman Review — Unlocking Digital Competition — the platform-era diagnosis this paper explicitly checks against the AI market
• Rent Dissolution vs. Rent Capture: the Enforcement Record — the dissolve-pole remedies this paper's own policy menu belongs to
• Platform and Data Rents · Economic Rent · Quasi-Rent
• Objection: Taxing quasi-rents kills innovation — the incentive question this paper's Bertrand-competition evidence bears on directly
• Geoism — the rent-domain program and its gradient

Sources

1. Anton Korinek & Jai Vipra (2024, rev. October 2, 2024), "Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence," Institute for New Economic Thinking Working Paper No. 228; published as Korinek & Vipra (2025), Economic Policy 40(121), 225–256; also circulated as NBER Working Paper 33139. INET PDF · DOI · NBER — used for the cost-structure and economies-of-scale analysis; the Nvidia 92%/98% GPU-market figures; the talent-bottleneck mechanism; the "monopoly rents" quotation and the platform-shakeout analogy; the "intelligence feedback loop" mechanism; the vertical-integration examples (Microsoft/OpenAI, Google DeepMind TPUs, the Inflection AI "acquisition-by-proxy"); the July 2024 four-authority joint antitrust statement; the Bertrand-competition observation and LMSYS clustering evidence; and the full policy-remedy menu (B/C-claims — a peer-reviewed working paper offering diagnosis and policy analysis, not a randomized or quasi-experimental empirical estimate; fetched and read in full this session).