2025年人工智能人才报告

JANUARY 14, 2025 AI Talent Report Executive summary The number of graduates in the United States with AI-relevant technicaltraining has risen in recent years along with major breakthroughs in AI. CEAinvestigates the capacity of the United States to meet the growing demandfor AI talent. We begin by documenting the growing demand for AI talentand develop a classification system for categorizing AI-related postsecondarydegrees into two relevant categories: AI software and AI hardware. Usingthis degree classification system, we explore the level and growth of thesupply of graduates in AI-relevant degrees from U.S. institutions of highereducation overall, and in comparison to other countries. Wefind that: The number of AI-relevant graduates (bachelor’s, master’s, anddoctorate) in the United States has increased significantly in the lastdecade, driven by growth in BA and MA degrees. Non-U.S. citizens make up a significant share of graduates, with non-U.S. citizen PhD students making up nearly half of AI-relevant PhDgraduates. Most AI-relevant PhD graduates from U.S. institutions are trained atpublic universities. Internationally, the United States produces more AI-relevant degreegraduates than most countries, with the notable exceptions of India(which produces more AI-relevant BAs) and China (which produces bothmore relevant BAs and PhDs). The number of AI-relevant BA and PhD graduates is growing fasterin China than in the United States. The United States still produces a larger share of top AI researchersand continues to be the world leader in housing top AI labs andproducing frontier AI models. Demand for AI talent (to the limited extent we can assess it) appears tobe growing at an even faster rate than the increasing supply of AI talenttrained in U.S. colleges and universities. The United States could increase its supply of AI talent in three broadways: 1) increasing the number of students training at U.S. institutions, 2)increasing post-graduate inflow of students trained abroad and reducingpost-graduate outflow of students trained in the United States, and 3)providing incentives for capable workers to switch into AI sectors. The report concludes by discussing policy interventions that could supportthese three paths to increasing the supply of AI talent. Motivation While Artificial Intelligence (AI) is still in its early stages of development, itsrapid integration into everyday life has importanteconomic implications,especially for productivity and competitiveness. The enormous potential ofAI to shape critical infrastructure and technology also makes it a key focus ofnational security policy. At the most basic level, most of today’s advanced AIsystems are produced by using computing power to train models by exposingthem to large volumes of data. Discussions of the capacity of the UnitedStates to produce top-tier AI models usually focus on the domestic capacityto scale up the physical components of these inputs (like the production ofsemiconductors, the construction of, and the electricityprovision to run them). However, human talent is also a core engine behindthe growth of the AI sector. Highly-skilled workers develop the algorithms,implement the training of models on datasets, develop cutting-edgehardware, and operate the datacenters where AI systems are trained and run.The supply of AI talent—the human capital necessary for the production ofAI systems—is therefore an important consideration when evaluating thedomestic capacity of U.S. AI industry. This report examines the currentproduction of AI talent in the United States to understand the degree towhich the United States is poised to continue to lead in AI development andimplementation.datacenters In this report, CEA considers two classes of work that require AI talent. Thefirst of these classes is the “software side:” the researchers who work directlyonlike creating and curating data, developing model architecturesand creating algorithms to train, adapt andfinetune models, and deployingthefinal product. This classmachine learning engineers, softwareengineers, and research scientists working at companies likeorthat produce AI models. The second of these two classes isthe “hardware side.” This is a broader category. CEA defines “AI hardware”work to include skilled work at all parts of the AI supply chain for computerhardware. Because this report is concerned with the capacity of the U.S.educational system to train eligible workers, CEA focuses on uniquely skilledwork in these sectors to differentiate from work that does not require highereducational background or training.[1]This class includes work at differentparts of the AI supply chain, including engineers and technicians inas well asdesign, fabrication and assembly,testing, & packaging (ATP) facilities. The distinction between software andhardware is more complex than the simple grouping outlined here: engineersdesigning semiconductors are increasinglyto help optimizechip design, and h