HeinOnline is a subscription-based resource containing thousands of academic and legal journals from inception; complete coverage of government documents such as U.S. Statutes at Large, U.S. Code, Federal Register, Code of Federal Regulations, U.S. Reports, and much more. Documents are image-based, fully searchable PDFs with the authority of print combined with the accessibility of a user-friendly and powerful database. For more information, request a quote or trial for your organization below.
![handle is hein.crs/cmnadantru0001 and id is 1 raw text is:
Congressional Research Service
Inforrning the legislative debate since 1914
S
April 16, 2025
Competition and Antitrust Concerns Related to Generative AT
Artificial intelligence (AI) technologies used to generate
synthetic content, such as text, images, audio, video, and
computer code, are broadly referred to as generative AI
(GenAI). One type of GenAI is a large language model
(LLM), which can generate responses to prompts in natural
language format once it has been trained on a massive
amount of text (e.g., text from millions of web pages) and
defined by billions of model parameters.
Developing and operating a large-scale GenAI model may
require significant computing resources, such as hardware,
software, and other information technology (IT)
infrastructure, which can be costly. This has raised concerns
about who may be able to develop such models, whether
companies with substantial resources may have a
competitive advantage over smaller competitors and start-
ups, and whether some companies that own IT
infrastructure might be engaging in anticompetitive
conduct. Some stakeholders have argued that competition is
thriving in each component of GenAI development.
Some U.S. companies-including Amazon, Google, Meta,
and Microsoft-own IT infrastructure used to train and
deploy GenAI models and also invest billions of dollars to
develop their own models. Other companies are exploring
methods to reduce computing power needed to develop AI
models. For example, the Chinese company DeepSeek
stated that it used cost-efficient model training to develop
its GenAI model. Some Members of Congress are
considering whether congressional action could help
promote competition to ensure America's global leadership
in the development of GenAI.
This In Focus discusses some potential concerns about
competition in the development of GenAI in U.S. markets
and potential issues raised by antitrust enforcers. It also
provides some considerations for Congress.
Development of GenAl Models
Although the terminology of AI is still evolving, an AI
model generally refers to a computer program that uses
algorithms and mathematical functions to process inputs
(e.g., data and prompts) into outputs (e.g., predictions and
decisions). The AI model is trained on sample datasets to
recognize certain types of patterns and learn from the
data, optimizing its performance over time to create outputs
such as text that mimic human language. For an overview
of GenAI models, see CRS In Focus IF12426, Generative
Artificial Intelligence: Overview, Issues, and
Considerationsfor Congress, by Laurie Harris.
Foundation models are general-purpose AI models
pretrained on large datasets. Foundation models can be used
for various downstream purposes, such as user-facing AI
applications and services (e.g., OpenAI's ChatGPT model,
which supports its AI chatbot), or further fine-tuned using
smaller datasets of specialized knowledge for work in
specific domains. If a limited number of foundation models
are available, the companies that deploy these models might
have market power and significant influence over GenAI
developers, particularly if it is difficult for others to enter
the market by developing their own foundation models.
Foundational Model Costs and Competition
Developing a GenAI model, especially a foundation model,
demands significant time, effort, and computing resources.
To train a foundation LLM, for example, an AI developer
may acquire a large dataset from a third party or create one
by scraping text from publicly available web pages. Before
training, the developer cleans the data (e.g., removes
incorrect, inaccurate, or duplicate texts) and preprocesses
the data into tokens-the basic units of text that the LLM
can process. The training can involve trillions of tokens to
develop foundation models with billions of parameters,
which require millions of hours of computer chip
processing time (e.g., DeepSeek claimed that one of its
LLMs required 2.8 million graphics processing unit [GPU]
hours-equal to about 57 days-to complete its training on
a cluster of 2,048 high-performance chips). To conduct AI
training at this scale, the developer typically either
purchases high-performance chips or leases computing
resources from a cloud computing service provider.
According to the annual AI Index Report 2025, costs of
training large foundation models alone are widely
estimated to reach into the millions of dollars-and
continue to rise. While AI developers rarely disclose exact
cost information, the AI Index Report published estimated
training costs based on rental prices charged by major cloud
service providers for accessing computing hardware and
processing data. For example, the estimated cost of Llama
3.1 LLM, which has 405 billion parameters and was
released by Meta in 2024, could be around $170 million
(excluding other costs, such as data acquisition and labor).
The high costs and resources needed to develop large
foundation models, particularly during the pretraining
phase, may limit the number of companies that are able to
develop these models. Some companies may also use their
resources to obtain a competitive advantage. For example,
some companies might be able to leverage resources that
they have been investing in for years (e.g., data, cloud
computing services) to develop GenAI models. This might
make it difficult for start-ups and smaller companies to
compete if they do not have access to these resources. Some
companies, such as DeepSeek, claim that they have
developed AI models cost-effectively, which raises
questions about the necessity of large AI investments and
https://crsreport](ViewImageLocal?handle=hein.crs/cmnadantru0001&div=1&collection=&method=preview&ext=.png&size=3)