AI Economy: Invisible Growth in the United States

A Rapid Yet Discreet Economic Growth

The artificial intelligence economy in the United States is experiencing spectacular expansion, with an annual growth rate of 2,000%. However, this progress remains difficult to discern in the overall gross domestic product (GDP) statistics. A report authored by economists from the University of Virginia, Anthropic, and the Bank of Canada sheds light on this dynamic while questioning the reasons for its invisibility.

Estimating AI GDP

According to forecasts, the nominal GDP of AI could reach approximately $250 billion by 2025, with a real annual growth rate adjusted for quality of 2,600%. Despite this, several factors contribute to its low visibility in global economic statistics.

Factors of Invisibility

Although the construction of data centers is booming, it has not yet had a significant enough impact to meaningfully influence GDP. Additionally, AI inference, which constitutes the main economic impact, sees its nominal revenues increase moderately, as unit prices decrease almost as quickly as quality-adjusted production increases.

Measuring the AI Economy

Three main methods are used to measure the AI economy:

1. Nominal Computing Expenditures: Computing expenditures in the United States rose from $37 billion in 2023 to $90 billion in 2024, and then to $219 billion in 2025.

2. Raw Computing Capacity: Thanks to innovations in chip technology, actual capacity is increasing by over 200% per year.

3. Quality-Adjusted AI Production: This has seen growth of 2,290% in 2024 and 2,271% in 2025.

The AI economy is therefore much larger than conventional measures suggest.

Recommendations for Better Measurement

The authors of the report suggest several avenues for improving the measurement of the AI economy:

• AI Satellite Accounts: Create AI-specific measures to better inform GDP calculations.
• Data Improvement: Encourage collaboration between statistical agencies, businesses, and universities to enrich available data.
• Integration into Economic Projections: Policymakers should include measures of AI productive capacity in their medium-term economic forecasts.

Working on AI while observing current economic data gives the impression that something significant is lurking beneath the surface, akin to a scene from the movie Jaws where the shark approaches.

Challenges in AI Oversight

Oversight of AI systems, particularly regarding automated alignment, presents complex challenges. An article from the UK AI Security Institute highlights that this research, while promising, is far from straightforward.

Complexities of Automated Alignment

• Optimization Pressure: AI is often optimized for human approval, which can lead to non-intuitive errors.
• Volume of Research: Automated systems could process datasets much larger than those used by humans.

Proposed Interventions

To improve oversight, several interventions are suggested:

• Measurement: Reproduce research projects and evaluate agent performance.
• Generalization: Test various training proxies to understand generalization.
• Evolving Oversight: Develop and test new oversight protocols.

The ability to oversee systems that are more intelligent than humans raises fundamental questions about controlling the future.

Giant Permissive Image Corpus: A Valuable Resource

Researchers from Stanford University, Radical Numerics, the University of Michigan, and Salesforce Research have launched the Giant Permissive Image Corpus (GPIC), a dataset comprising 100 million images with captions. All these images are under permissive licenses, making them available for research and commercial use.

GPIC Details

The GPIC consists of 100 million training images, 200,000 validations, and 1 million test examples. This corpus represents an invaluable resource for researchers and startups, facilitating development and innovation in the field of AI.