← TRACKER
THREAT VECTOR // HIGH

POWER CONCENTRATION

High  ·  CURRENT VALUE 79/100

Superintelligent capabilities concentrated in the hands of a few entities. Democratic oversight mechanisms inadequate for the pace of development.

Current status as of 2026-09-04

The AI industry's compute-scaling dynamic has produced the most rapid concentration of technological power in the post-antitrust era. Frontier model training now requires clusters costing $500M+ (Meta's Llama 4, xAI's Colossus, OpenAI/Microsoft's Stargate). The number of organizations capable of training a frontier model shrank from ~15 in 2020 to ~6 in 2025 (OpenAI, Anthropic, Google DeepMind, Meta, xAI, DeepSeek — with tight coupling between OpenAI/Microsoft, Anthropic/Amazon, and Google/DeepMind). Nvidia's near-monopoly on training accelerators (>90% market share for H100/H200-class chips) creates a single upstream chokepoint that touches every frontier lab.

The 2026 signal is that the “AI is a natural monopoly” position (Suleyman, Andreessen) and the “AI compounds existing monopolies” position (Khan, Lynn) are both partially correct. Vertical integration of the stack — chips, clouds, models, applications — is producing companies with unprecedented cross-market leverage. The DOJ Google search ruling (2024) suggests judicial willingness to unwind such stacks, but the specific antitrust theory for AI (whether it is foreclosure, exclusive dealing, or something new like “model-capability leverage”) has not been articulated in a filed case yet. Watch for the first AI-specific antitrust action rather than any specific merger.

Historical trajectory

DateLevel
2020-06
40
2021-01
45
2021-06
50
2022-01
55
2022-06
58
2022-11
62
2023-03
65
2023-06
68
2023-11
70
2024-03
72
2024-08
75
2025-02
77
2025-06
78
2026-01
79

Key papers

  1. Amazon's Antitrust Paradox
    Lina Khan · 2017 · Yale Law Journal 126:710
    The paper that reopened US antitrust theory for platform companies. Not about AI directly, but the intellectual framework any AI antitrust action will draw on. Read to understand why the “consumer welfare” standard is now contested.
  2. The Foundation Model Transparency Index
    Bommasani et al. (Stanford HAI/CRFM) · 2024 · Stanford HAI
    Quantifies how little disclosure frontier labs provide on training data, model behavior, and downstream harms. Baseline for any regulatory argument about market power in models — you cannot regulate what you cannot see.
  3. Market concentration implications of foundation models
    Vipra & Korinek · 2023 · arXiv:2311.01550
    Economic analysis of AI market structure with predictions about which segments concentrate and which remain competitive. Cited in EU Commission Digital Markets Act preparatory work.
  4. Investigation of Competition in Digital Markets (Majority Staff Report)
    US House Judiciary Antitrust Subcommittee · 2020 · US House of Representatives
    The underlying evidentiary record for current tech-monopoly cases. The AI-specific successor investigation is still being assembled but will draw on the same investigative methodology.

Strongest counterargument

The steelman: what looks like concentration is the normal J-curve of any capital-intensive general-purpose technology. Railroads, electricity, telecom, and cloud computing all showed identical patterns — a small number of massive first-movers, then broad diffusion once the underlying capability commoditized. Open-weight models (Llama 3.1 405B, DeepSeek-V3, Mistral Large 2) are already competitive with the closed frontier, and the inference cost curve is dropping ~10× per year, which means the “moat” of frontier training will erode before regulatory action could take effect. Aggressive AI antitrust in the current geopolitical climate would specifically disadvantage US companies against Chinese national champions who face no comparable scrutiny. The pattern of applying existing antitrust categories to a novel technology has failed before (Microsoft in the 1990s, which regulators fought to break up despite the market already routing around Windows).

Related events (1)

Events from the tracker's timeline whose tags, title, or description match this vector. Heuristic auto-match; some may be tangential.

Dead Hand systems that amplify this vector (7)

Revision history