AI Competition Policy

Summary: Policy framework addressing market concentration risks in the AI stack. Argues that open-weight models are essential competitive infrastructure — preventing monopolistic control of foundation models, enabling competition at every layer (cloud, chips, applications, services), and distributing economic benefits broadly rather than concentrating them in a few closed-model providers.

Core Thesis (from Coalition Statement)

"Open weights also strengthen competition and competition is what keeps the gains of AI broadly shared rather than concentrated in a few hands. By allowing many organizations to build, adapt, and deploy advanced models, open weights create rivalry not only among model developers but across cloud chips, applications, and services. That competition spurs innovation, drives down costs, and distributes the benefits of AI broadly across our economy."

"Concentrating advanced AI capabilities behind a small number of closed models compounds that risk. It results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers."

AI Stack Market Structure Analysis

Current Concentration Risks (2026)

Layer Incumbents Concentration Risk Open Weights Impact
Foundation Models OpenAI, Anthropic, Google, xAI, (Meta open) 3-4 closed providers control frontier Multiple open frontiers (Llama, Mistral, Qwen, Gemma, Phi) = competitive constraint
Model Hosting/Inference Azure OpenAI, AWS Bedrock, GCP Vertex, Anthropic API Cloud lock-in + model lock-in Open models portable across clouds; multi-cloud inference (Together, Fireworks, Anyscale, vLLM)
Training Compute NVIDIA (GPUs), cloud hyperscalers NVIDIA ~90% AI GPU; cloud oligopoly Open models enable on-prem, specialty clouds, sovereign clusters — diversifies compute demand
Developer Tools OpenAI SDK, LangChain, HF Transformers Ecosystem lock-in to closed APIs Open models work with any framework; standards (OpenAI-compatible APIs) reduce switching cost
Application Layer Emerging vertical AI (Harvey, Cursor, Glean, Abridge) Currently competitive; risk of platform capture Open weights = lower barrier to entry; more vertical startups

Herfindahl-Hirschman Index (HHI) Projections

Scenario Foundation Model Layer HHI Downstream Competition
Closed-only frontier >2500 (highly concentrated) Oligopoly pricing; limited negotiation leverage
Open + closed frontier <1500 (moderate) Price discipline; feature competition; negotiation leverage
Pluralistic open frontier <1000 (competitive) Innovation race; cost optimization; specialization

Competition Mechanisms Enabled by Open Weights

Mechanism Description Economic Effect
Multi-homing Customers run Llama on Azure, Mistral on AWS, Qwen on GCP, Phi on-prem Reduces cloud switching costs; forces cloud price/feature competition
Model arbitrage Switch base models for different tasks (coding → Code Llama; reasoning → Mistral; chat → Llama) Best-model-for-task; prevents single-model monopoly rents
Fine-tuning competition 1000s of orgs create specialized derivatives; compete on quality/cost Drives down fine-tuning cost; raises quality floor
Inference provider competition Together, Fireworks, Anyscale, Replicate, Baseten, vLLM self-host Price competition (0.30-3/MTok); SLA differentiation; specialization
Hardware diversification Open models optimized for AMD, Intel, Apple Silicon, TPU, custom ASICs Reduces NVIDIA pricing power; enables sovereign hardware stacks

Antitrust / Competition Law Relevance

US Context (2026)

Development Relevance to Open Weights
DOJ/FTC AI Competition Inquiry (2024-2025) Examining foundation model market structure; open weights as competitive constraint
Executive Order 14110 (AI EO) Section 5: "Promoting Competition" — calls for fair competition in AI markets
Merger Reviews (Microsoft/Inflection, Amazon/Adept, Google/Character.ai) "Acqui-hires" of model talent; open weights reduce talent concentration value
State AG Investigations California, NY, others examining AI market power

EU Context

Instrument Open Weights Angle
Digital Markets Act (DMA) Foundation models as "core platform services"? Open weights = interoperability remedy
AI Act Art. 53: GPAI open source carve-outs; competition assessment for systemic models
EU Merger Control Open weights as competitive constraint in market definition

Key Competition Theories

  1. Foreclosure — Closed model + cloud bundle forecloses rival model hosting
  2. Tying — API access tied to cloud consumption commitments
  3. Data Advantage — Closed models accumulate usage data for improvement; open models break this loop
  4. Innovation Competition — Open weights enable "innovation at the edges" (fine-tuning, distillation, merging)

Policy Toolkit for AI Competition

Tool Target Open Weights Role
Merger Scrutiny Vertical integration (cloud + model) Open weights reduce merged entity's market power
Interoperability Mandates API standardization, model portability Open weights natively portable; mandate OpenAI-compatible APIs
Data Access / Portability Training data, user interaction data Open weights don't solve data; need separate data portability
Compute Access NAIRR, cloud credits, GPU marketplaces Open weights increase compute demand diversity
Open Source Procurement Gov purchasing preferences Mandate open weight options in federal AI contracts
Standard Essential Patents (SEPs) Model architecture patents FRAND commitments for foundational architectures

Coalition's Competition Argument Structure

  1. Premise: AI gains should be "broadly shared rather than concentrated in a few hands"
  2. Mechanism: Open weights → competition across full stack (cloud, chips, apps, services)
  3. Outcome: "Spurs innovation, drives down costs, distributes benefits broadly"
  4. Risk of alternative: "Concentrating advanced AI capabilities behind a small number of closed models... results in a small number of single points of failure, weakens competition"
  5. Policy ask: "Avoid premature restrictions on open models that stifle competition or drive innovation overseas"

Tensions & Counterarguments

Tension Counterargument / Resolution
Open models may be lower quality Gap narrowing (Llama 3.1 405B ≈ GPT-4o); competition drives all forward; specialization > single metric
Training cost = natural monopoly Compute costs falling (scaling laws, efficiency); open models amortize cost across ecosystem
Safety requires centralization Coalition: "Openness may be one of most important paths to AI safety" — distributed red teaming
National security requires control Coalition: "Defenders need access to models with comparable capabilities" — open enables defense
Foreign adversaries benefit Allies need sovereign capability; closed models also leak (theft, insiders); open = faster defense innovation

Metrics for Competition Health (Proposed Dashboard)

Metric Target Measurement
Frontier model count (open + closed) ≥6 credible frontier families Model cards, benchmarks, release tracking
Inference price ($/M tokens) Declining 2x/year Provider APIs, self-host cost models
Fine-tuning diversity >10K public derivatives/year HF Hub, ModelScope, GitHub tracking
Cloud/model multi-homing rate >50% enterprises use ≥2 model providers Procurement surveys
Startup formation (vertical AI) Growing YC batch AI % YC, a16z, PitchBook data
Hardware diversity ≥3 viable AI accelerator vendors MLPerf submissions, cloud offerings

Related Wiki Pages

  • open-weights-policy — Parent framework
  • american-ai-leadership — Strategic frame
  • market-concentration-ai — Measurement & analysis
  • compute-access-policy — NAIRR, credits, diversity
  • vertical-ai-applications — Downstream competition
  • platform-economics-ai — Two-sided markets, network effects
  • antitrust-ai — Legal framework

Version History

Version Date Changes
1 2026-07-24 Initial creation from coalition statement + competition economics