National Research Cloud
Call To Action
Dear President Trump and Members of Congress:
Artificial intelligence (AI) will be one of the most consequential technologies of the 21st century, carrying enormous implications for the economy and society. The United States, along with all countries, will be profoundly affected by its deployment. Many have compared AI technology to electricity and predict social changes comparable to those brought about through electrification.
Historically, the United States has led the world in the development of AI technology, with U.S. universities at the forefront. But our continued preeminence in AI research is by no means assured. We believe the government, working with industry and academia, must devote substantially more computational, data, and human resources to maintain our leadership in this area.
For over 70 years, the United States has benefited greatly from a cooperative and distributed research and development model involving the federal government, academic institutions, and private industry. Federal funding for fundamental research has been a critical engine of the nation’s success in developing and commercializing technology. American universities, with their unique mission of coupling research with training the next generation of scientists and technologists, are free to pursue the most basic, long-term and speculative research. The private sector, with its brilliance in applied research and product development, excels at commercializing innovations generated either in their own labs or by their university partners.
These interlocking pieces have given rise to the most fertile and innovative economy the world has ever seen. From aerospace to biotech to the Internet, the vibrancy of our industry owes much to the crucial contributions of all three partners. Until recently, the budding AI industry has followed this traditional pattern. But it is increasingly difficult for academic institutions to play their traditional role in the prolific research ecosystem that has served our country so well.
Modern AI’s need for expensive computational resources — specifically, hardware tailored to the massive computational demands of large machine learning models — is growing exponentially, becoming increasingly out of reach for students and faculty. And the kinds of large, well-curated data sets required to train these models can be equally hard to access. Without these resources, the basic, long-term research and training functions performed so well by universities will be hobbled.
To combat these trends, we are calling for the creation of a “national research cloud” — a new partnership between academia, government and the private sector to provide crucial computing and support for maintaining and enhancing the strength of our academic institutions in AI research and training. This capability will provide both the computational resources and the vital, large-scale data needed to enable AI research. A national research cloud will provide crucial infrastructure — hardware, software and personnel — to sustain U.S. preeminence in AI research.
First, the national research cloud will provide academic and public interest researchers with free or substantially discounted access to the advanced hardware and software required to develop new fundamental AI technologies and applications in the public interest. Second, it will provide the expert personnel necessary to deploy these advanced technologies at universities across the country.
Currently, at list price on a commercial cloud, training a state-of-the-art neural network for a natural language processing task can cost anywhere from tens of thousands to hundreds of thousands of dollars, depending on the complexity of the application and the network model employed. Moreover, the expertise required to make effective use of these resources is not widely held, limiting progress to a small number of institutions and fields. These limitations put many research projects beyond the capabilities of most academic researchers, and seriously hamper the education of our next generation of AI practitioners. A national research cloud could ameliorate this growing problem and allow university research and teaching to remain at the forefront of the field.
Policymakers could choose to implement this aspect of the research cloud in several ways. Following on the successful model of federal user facilities operated by the DOE, NIST and NIH, the new infrastructure could be housed at existing or new national labs. Alternatively, the hardware infrastructure could be provided through contracts with commercial cloud providers, thereby benefiting from the rapid technology upgrades that are imperative in the private sector. Finally, a combination of these strategies could reduce startup costs and lead time to get the effort up and running.
Finally, government agencies should be directed to redouble their efforts to make more and better quality data available for public research at no cost. Researchers could work with agencies to develop and test new methods of preserving data confidentiality and privacy, while government data will provide the fuel for breakthroughs from healthcare to education to sustainability. For example, by analyzing longitudinal data from federal health programs such as TRICARE, Medicare and Medicaid and the Veterans Administration, we could potentially make critical discoveries that substantially improve the health and longevity of Americans, improve diagnosis and treatment and reduce costs. The government also maintains the most substantial weather dataset - collecting billions of observations per day - which, if fully unlocked, could improve our ability to forecast crop yields, predict dangerous storms and prepare affected communities.
Timeliness is imperative for this initiative, as progress in the AI field is moving rapidly in countries around the globe. Partnership with leading commercial cloud providers in the U.S. is one important component of both starting quickly and continuing to move quickly, compared to conventional government and university infrastructure timelines. Importantly, this initiative will spread opportunity in the AI economy more widely throughout the nation, supporting students and researchers at colleges and universities across the country, activating the vast potential of researchers who currently lack access to the funding or data needed to harness their talents and creating fertile ground for further commercialization through entrepreneurship and small businesses.
In order to carry out this plan, a bipartisan task force, jointly chaired by leaders in academia and government, and including representatives from industry, should be charged by the Office of Science and Technology Policy (OSTP) to develop a specific proposal within six months. The group will identify and prioritize computational needs and potential datasets to support academic AI research, and develop a plan that ramps up over the next five years. Though we recognize and support the work done by other organizations, including the OSTP, the President’s Council of Advisors on Science and Technology and the Congressionally mandated National Security Commission on Artificial Intelligence, we are calling for this separate, near-term action to address widely recognized needs, while, over the medium term, these initial steps can serve as building blocks for other efforts.