Issue Brief

Key Takeaways

➜ Academia is falling behind industry in frontier AI research. Today, no university in the world can build a frontier AI system on par with industry. 

➜ Industry is dominating AI development due to its massive datasets, unprecedented computational power, and top-tier talent. Companies have over 1000x more compute than universities, and they produce AI models that are 50x larger.

➜ Governments should continue investing in public sector AI. Academia must be at the forefront of training the next generation of innovators and advancing cutting-edge scientific research in the public interest.

 Introduction

Academic researchers invented nearly all of the core technologies underpinning AI. From the 1963 founding of Stanford’s AI Lab, the home of many foundational AI breakthroughs, to the creation of the groundbreaking image classification architecture AlexNet at the University of Toronto in 2012, the academy has been at the heart of the field from the very beginning. For AI to be developed responsibly and in the public interest, academia must continue to play a central role.

In the last decade, however, the field has been increasingly dominated by the private sector. Building and deploying AI systems has become hugely resource intensive, often requiring billions of dollars in investment, custom supercomputing clusters, and enormous datasets containing much of the available data on the internet. This shift has created a significant power imbalance, where academic talent and government support flows to private companies that now produce the vast majority of the world’s most powerful AI systems.

As AI systems have grown more capable, the costs of developing foundation models has become a substantial barrier to entry. In 2017, Google spent less than $1,000 to build its first transformer-based AI model—by 2023, it cost Google almost $200 million worth of computational resources to develop its state-of-the-art model Gemini Ultra. In a world where the cost of building a single foundation model is equivalent to the annual operating budget of an entire university, few academics can meaningfully participate in the development of state-of-the-art AI models.

This disparity undermines not only the future of academic research but also the potential for a public sector AI ecosystem that serves the public interest. Unlike industry, academic research is driven not by profit but by the pursuit of scientific knowledge. Time and time again, pathbreaking AI innovation has come from the curiosity-driven research of academics who have the freedom to pursue ideas that are not immediately commercializable. Academia must play a leading role in developing frontier AI to ensure that we can understand and safely deploy the technology.

 The Industry-Academia Divide in AI

Until 2014, academia produced the largest number of notable machine learning models each year. Since then, academia has continued to advance, while industry has raced ahead. In 2023, industry produced 51 notable machine learning models, far outstripping academia’s 15; that same year, there were just two notable machine learning models produced by governments and none from nonprofits or government collaborations with academia.

The gap between academia and industry can be measured across three dimensions: funding, compute, and talent. In each of the three, the resource divide between researchers in academia and industry continues to grow.

The funding disparity between academia and industry is long-standing, but it has expanded significantly in the last decade. Since 2014, aggregate private investment in American AI companies has exceeded $300 billion, whereas many universities have seen a decline in funding in real terms as government support has not recovered following the 2008 global financial crisis. A decade after the Great Recession, U.S. states spent on average 16 percent less on higher education, and many states made further reductions as part of their budget cuts associated with the COVID-19 pandemic. Academic institutions do not have the resources to keep pace with Big Tech or many startups, and the trend lines are heading in a troubling direction.

Computational resources are an essential building block of AI. Building large AI models requires advanced chips capable of carrying out trillions of operations in parallel. Without chips and the energy to run them, it is impossible to build AI systems like ChatGPT. In recent months, top American universities have announced sizable purchases of the graphics processing units (GPUs) needed to build foundation models. Consider access to Nvidia H100s, the current best-in-class GPU. In 2023, Harvard announced its purchase of 384 H100s, while this year UT Austin announced it would purchase 600 and Princeton 300. These universities are now among the world’s top academic institutions in terms of computational power.

Nevertheless, these purchases pale in comparison to those of major AI companies. The same week that Princeton announced it would purchase 300 H100s, Meta announced it would buy 350,000. Microsoft plans to have 1.8 million H100s by the end of this year, and the startup xAI is using a supercomputing cluster of 100,000 H100s to train its latest Grok model. Among academics with access to compute, it is typical for them to have access to between 1 and 8 GPUs, whereas industry researchers may have access to thousands. This gap in raw computational power has led to a situation where many foundation models from industry are more than 50 times larger than those from academics.

Academia must also contend with issues related to attracting and retaining talent. In 2011, PhD graduates in AI were equally likely to pursue careers in academia or industry. By 2020, the balance had shifted decisively in favor of industry, with nearly 70 percent of new AI PhDs pursuing careers in the private sector. The compute divide paired with funding disparities has contributed to an increase in top talent moving from academia to industry. Access to computational power determines the scale and complexity of experiments that researchers can conduct, leading top researchers to gravitate to organizations that can offer them sufficient compute to carry out cutting-edge research. Without access to adequate computational resources, academics cannot complete even basic research projects that involve building small foundation models.

In addition to extra compute, industry positions pay substantially more. Whereas the average computer science professor in the United States earns $113,000 per year, research scientists in AI at companies such as Meta earn $335,000 in total compensation—nearly triple their academic counterparts. The earnings differential has led many faculty members to leave the academy for the private sector. Nationwide, the net flow of talent from academia to industry has doubled in recent years. The brain drain from academia to industry is noticeable across the country, with computer science departments losing many of their brightest researchers and professors to industry.

Read the full brief    View all Policy Publications

Authors

Subscribe to Policy Program Updates