The Inference Bottleneck: Antitrust and Neutrality Duties in the Age of Cognitive Infrastructure

Based on the paper "The Inference Bottleneck: Antitrust and Neutrality Duties in the Age of Cognitive Infrastructure" by Gaston Besanson and Marcelo Celani, here is a detailed technical summary.

1. Problem Statement

The paper addresses a shifting competitive landscape in the generative AI sector. As the industry matures, the primary source of competitive advantage is moving from one-time model training to continuous inference, distribution, and routing.

• The Core Issue: Large-scale inference is becoming "cognitive infrastructure"—a bottleneck input essential for downstream applications (e.g., AI assistants, productivity suites).
• The Risk: Firms controlling this infrastructure are often vertically integrated, competing downstream with their own API customers. This creates a "referee-and-player" conflict.
• The Mechanism of Harm: Unlike traditional antitrust concerns focused on predatory pricing, foreclosure in AI inference is executed through non-price discrimination and opaque steering. Providers can degrade the quality of service (QoS) for rivals or steer users toward their own services without changing posted prices. These actions are difficult to observe, hard to litigate, and often plausibly justified by technical constraints (e.g., safety, capacity).

2. Methodology and Conceptual Framework

The authors employ a theoretical and policy-oriented methodology, moving away from traditional market definition (which is ambiguous in AI) toward measurable conduct and functional gatekeeper status.

• Conceptual Definition: They propose defining "Cognitive Infrastructure" not by market share, but by three falsifiable conditions:
  1. Material Reliance: Downstream products rely on the service such that switching causes non-trivial re-engineering or measurable degradation in cost/quality.
  2. Vertical Incentives: The provider controls access terms and is vertically integrated into downstream complements that compete with API customers.
  3. Discrimination Capacity: The ability to raise rivals' costs or degrade quality via non-price levers (latency, throughput, feature gating) while maintaining neutral headline pricing.
• Theoretical Lens: The paper utilizes Raising-Rivals'-Costs (RRC) theory. It argues that a vertically integrated firm can profit by increasing rivals' effective costs or degrading their quality, even without below-cost pricing. This allows the firm to keep rivals "living but not thriving," preserving ecosystem breadth while neutralizing direct competition.
• Audit Protocol: The authors propose a technical audit framework that does not require access to private user data or model weights. Instead, it relies on:
  • Repeated measurements over a sampling window.
  • Controls for region, load, and contracted tiers.
  • Detection of sustained "QoS wedges" (e.g., >15% difference in p95/p99 latency or error rates) between first-party and third-party users.

3. Key Contributions

The paper makes three primary contributions to the fields of antitrust and AI governance:

1. Defining "Cognitive Infrastructure": It establishes a falsifiable, conduct-based definition of infrastructure that bypasses the need for clean market definitions. It focuses on reliance and discrimination capacity rather than static market share.
2. Reframing Theories of Harm: It shifts the antitrust focus from price-based predation to non-price foreclosure. It identifies three specific mechanisms of harm:
   • QoS Discrimination: Systematic differences in latency, throughput, and error rates.
   • Feature Gating: Withholding advanced capabilities (e.g., tool use, context windows) from third parties.
   • Steering/Default Routing: Opaque selection in agentic flows where models route tasks to the provider's own tools or preferred partners.
3. Proposing "Neutral Inference": A targeted, auditable conduct remedy designed for the AI era. It is positioned as an operational design pattern rather than broad ex-ante regulation.

4. Proposed Solution: "Neutral Inference"

The authors propose a framework of Neutral Inference duties, applicable only when observable evidence indicates functional gatekeeper status. This approach aligns with existing frameworks like the EU's Digital Markets Act (DMA) and the AI Act.

Three Core Obligations:

1. QoS Parity: Providers must ensure no materially better inference QoS (latency, error rates, timeouts) for first-party complements compared to similarly situated third parties, after controlling for objective cost and risk drivers.
2. Routing Transparency: Providers must maintain auditable routing records that capture:
   • Eligible tool sets.
   • Selected tools and their ranking.
   • Whether commercial relationships influenced eligibility or ranking.
   • Note: This focuses on selection outcomes, not chain-of-thought disclosure.
3. FRAND-style Non-Discrimination: Access terms for similarly situated buyers must be non-discriminatory, supported by fast dispute resolution mechanisms.

Enforcement Model:
The paper advocates for a compliance auditing model rather than decade-long litigation. This includes:

• Periodic audits using standardized test suites.
• Interim measures upon detection of discrimination.
• Penalties tied to inference revenue.

5. Results and Implications

While the paper is a theoretical proposal rather than an empirical study with experimental results, its "results" are the formulation of a viable policy path:

• Feasibility: The proposed audit protocol is technically feasible without requiring the disclosure of proprietary model weights, addressing privacy and trade secret concerns.
• Targeted Intervention: By focusing on "observable evidence" of gatekeeper status, the framework avoids punishing scale or innovation, focusing instead on preserving contestability in downstream markets.
• Legal Alignment: The approach fits within the constraints of U.S. antitrust law (which is reluctant to impose broad access mandates) and complements EU regulations (DMA/AI Act).

6. Significance

The significance of this work lies in its timely identification of the inference bottleneck as the next critical frontier for antitrust enforcement.

• Paradigm Shift: It moves the conversation from "who owns the model weights" to "who controls the delivery and routing of cognition."
• Preventing Quiet Monopolization: It addresses the risk that infrastructure layers could acquire the power to determine which applications can exist through subtle, non-price degradation, a threat that traditional price-based antitrust tools often miss.
• Actionable Policy: It provides regulators and enforcers with a concrete, technical toolkit (QoS parity tests, routing logs) to detect and remedy anti-competitive behavior in the age of cognitive infrastructure, ensuring that the AI ecosystem remains contestable.