Center for Strategic and International Studies: The Call for a National Research Cloud and the Competition Over Compute

WASHINGTON, April 4 -- The Center for Strategic and International Studies issued the following blog on April 3:

One of the foundations for America's leadership in AI has been the university researchers whose work has driven much of the recent explosion of AI capabilities. However, some researchers are concerned that the growing cost of AI "training" and a lack of access to key datasets may prevent universities from continuing to serve as a source of progress in the future. These concerns have led the Presidents and Provosts of twenty-two top universities across the country--led by Stanford's Institute for Human-Centered AI--to submit a letter to Congress and the President calling for the establishment of a task force to institute a National Research Cloud to support American AI research.

Today, most AI systems are powered by deep learning, a form of AI that operates by passing information through a complex network of artificial neurons. The neurons are programmed to perform mathematical operations on the data they receive, and then pass along the results to other neurons which continue the process through multiple stacked layers. In this way, neural networks are able to process information in a similar way to how neurons function in the brain. These neural networks are trained on thousands or even millions of examples of the types of problems the system is likely to encounter, allowing the model to "learn" how to correctly identify patterns from the data. As models have grown more complex, this training process has come to take much longer and be far more computationally intensive than in the past.

The growing cost of training AI models is hard to overstate. In 2018, One AI research group found that the number of neural net operations (often referred to as "compute") used in the largest AI training runs doubled every 3.4 months. From 2012 to 2018, this represented a 300,000x increase in computational power. In comparison, compute would only have increased by 7x if it followed the same rate of progress as the famous Moore's Law--describing the density of transistors that can be fit on a microchip. In their letter, Stanford and the co-signers describe how training this kind of state-of-the-art AI system on the commercial cloud can cost tens or even hundreds of thousands of dollars. They note that Google's recent Meena chatbot cost $1.5 million in compute cycles to train. These costs are completely out of the question for almost every university research group. The price of cloud computing may be dropping, but it makes little difference when the computational demands of research are increasing far faster. Soon, only the largest tech companies will have the resources to support cutting-edge work, a fact that was driven home last year when the nonprofit OpenAI--co-founded by Elon Musk--was forced to open a for-profit arm and take on $1 billion in investment from Microsoft after concluding that it was simply impossible to compete with the likes of Google otherwise.

If this trend continues, it could severely damage America's leadership in AI. Cost increases could lead university researchers to become discouraged by the widening gap between academic and corporate labs and lure them to join the private sector, becoming part of the brain drain that is already threatening to hollow out our universities and deprive the next generation of AI talent of educators and mentors. As professors flee and costs mount, universities' role as the drivers of basic research will come under threat. University labs play a foundational role in the U.S.' technological leadership because of the way they enable researchers to take risky bets on long term projects with revolutionary potential. If emphasis begins to shift towards applied research in corporate labs, the United States may cease to be the source of cutting-edge technological innovations it has been in the past. Further, as the capacity to undertake AI research becomes limited to only a handful of firms, we will continue to see a concentration of power in the hands of a small number of organizations, and the formation of an innovation ecosystem increasingly unfriendly to disruption by low-resource startups.

Stanford's solution to these looming problems is to establish a National Research Cloud to support the efforts of academic and public interest researchers. This proposal would allow researchers to access advanced hardware and software for AI at free or discounted rates, and provide expert support to help deploy these technologies at universities across the country. Stanford also proposes working with the government to further open high-value public data sets for research purposes while developing new ways to preserve confidentiality and security.

These are important proposals, and the President and Congress should strongly consider taking up the letter's call to establish a task force to study them. The United States must think strategically about how the AI ecosystem continues to evolve, and what steps will be necessary to position the country for continued leadership. AI researchers are currently engaged in a vigorous debate about the staying power of the computationally-intensive deep learning techniques that currently dominate the AI landscape, but there are powerful arguments for the continuing importance of compute as a determinate of AI leadership.

Other nations are already beginning to realize the importance of compute, and have begun crafting strategies to position their countries for success. In China, concerns around computing power are a major factor driving the government's recent focus on AI chips and high-performance computing, and in Europe, policymakers recently released a Data Strategy outlining the bloc's plans to leverage data and cloud computing resources to promote Europe's global competitiveness. Policymakers in the United States must come to understand that compute will become as important a factor as talent or data in preserving national competitiveness in AI, and ensure that America's advantages in cloud services are able to be leveraged by the university labs that our AI future depends on. Establishing a National Research Cloud would be an excellent place to start.