Evaluating Natural Monopoly Conditions in the AI Foundation Model Market

Chapter 1
Introduction

up to 15 percent of occupations are highly exposed to AI

Foundation models are a core technology, which can be fine-tuned:

chatbots

medical image analysis

autonomous driving

To assess whether foundation models are a natural monopoly:

First
establish a set of generic criteria

Second
describe the production process for foundation models

Third
adapt the criteria to the production process for foundation models

Finally
apply the adapted criteria to the status quo foundation model market

The focal market of this study is that of pre-trained foundation models

OpenAI’s GPT

Anthropic’s Claude

Google’s Gemini

Chapter 2
Natural Monopoly Criterion

Monopoly

a market characterized by the
absence of competition

Natural Monopoly

a single firm can provide a homogeneous good or service for the full market at a lower total cost

Cost Subadditivity

Economies of Scale

Economies of Scope

Criterion

Product Homogeneity

Economies of scale

Sunk cost

Network effects

Economies of scope

spreading fixed cost over a larger number of units

spreading fixed cost over multiple distinct product lines

Social Cost

high prices

low product quality

limited product availability

low levels of innovation

Chapter 3
The Market for and
Development of Foundation Models

Foundation Models

Definetion

any model trained on a broad
set of data that can be adapted for various downstream uses

Modalities

Natural language

Visual models

Tactile models

Multimodal models

Development

Compute Acquisition and Training

Compute for Inference

Data

Algorithms

Training

Labor

Requirement

GPU

TPU

CPU

RAM

Vertical integration between large foundation model developers and compute owners will likely raise barriers to
entry

GPU demand has
outpaced supply

training takes significantly longer than inference, resulting in greater cost

Non-rival input

Transformer Model

Self-Attention

Multi-Head Attention

Model performance is primarily influenced by its size

Attracting and retaining workers for these roles can be expensive

Cost Classification

Fixed Cost

Variable Cost

Compute acquisition

Training data acquisition

R&D

Inference compute
(less significant)