Using Commercial Cloud Computing for Artificial Intelligence Research and Development

Published: July 27, 2022

Federal Market AnalysisCloud ComputingInformation TechnologyPolicy and LegislationResearch and Development

The White House National Science and Technology Council publishes lessons learned.

Back in February 2019, the White House published an Executive Order (#13859) on Maintaining American Leadership in Artificial Intelligence (AI). This order resulted in the November 2020 publication by the White House National Science and Technology Council (NSTC) of a series of recommendations concerning commercial cloud use for AI development. The NSTC recommendations included collecting lessons learned and best practices established from agency experience using cloud for AI R&D. To recount these, the recommendations called for federal R&D agencies to do the following:

  1. Launch pilot projects intended to explore the advantages and challenges of using commercial clouds for AI R&D.
  2. Improve the training of researchers to help them better leverage commercial cloud for AI R&D.
  3. Establish identity management best practices that optimize the effective use of commercial cloud resources for AI R&D.
  4. Publish best practices for the seamless use of multiple commercial cloud platforms for AI R&D.

Now, in July 2022, the NSTC’s Machine Learning and AI Sub-Committee has published a list of the best practices and challenges learned from early agency initiatives called for in the November 2020 recommendations. These include

  1. Establishing dedicated teams to manage researcher training and access to commercial cloud computing resources and services.
  2. Requiring two-factor authentication or other identity-based access to establish baseline security.
  3. Offering training and education for researchers to address skill gaps, advance access to cloud-based data sets, and build expertise across the researcher user base.
  4. Providing pre-computed resources for mission focused research to reduce duplicative work and create baseline analytical starting points.

These best practices, while addressing some of the concerns surrounding user access, data privacy, and research efficiency are probably just the beginning as the experiences from early agency initiatives also uncovered a host of common challenges that have yet to be overcome. These challenges include:

  • Setting appropriate levels of governance and administration for researcher access to and use of commercial computing resources provided by the government.
  • Developing authoritative, government-wide guidance to approved commercial cloud services.
  • Reducing the costs of storage and cloud services that are in some cases limiting access to shared data for multiple-party teams.
  • Protecting privacy and security while determining the best way to host data and make it accessible to the research community.
  • Integrating commercial cloud and government non-cloud resources.

These challenges specific to AI R&D are also common to other areas of cloud use by federal agencies. Agencies will eventually correct the governance and administration boondoggle. Meanwhile, the General Services Administration and Department of Defense are each in their own way groping toward the creation of catalogs of commercially-available cloud services. For example, the DOD’s Joint Warfighter Cloud Capability will eventually evolve into a Defense Cloud Marketplace with services categorized by security access level, capability type, and consumption based cost.

Cost will remain a challenge in an inflationary environment. Although competition will put some pressure on commercial cloud providers to reduce access costs, the reality is that the price of energy, to name just one factor, will continue to rise, forcing commercial providers to pass on those expenses to their government customers.

As for protecting privacy and security, these remain ongoing issues. The spread of zero trust architecture ought to help secure cloud resources, but ensuring privacy is a stickier issue.

Lastly, integrating commercial cloud and government non-cloud resources is the key to establishing efficient hybrid cloud environments. All agencies, R&D and non-R&D alike, are on this path. Experience and evolution should sort them out eventually.