The National Science Foundation is working to alleviate barriers to AI adoption — such as lack of GPUs and cost constraints — as it aims to play a supporting role in the ongoing national effort, according to a senior agency official.
The NSF has long been a cornerstone of national technology interests, from democratizing biotechnology research infrastructure to creating open, multi-modal AI models for the scientific community and investing in AI-ready testbeds. With the Department of Energy leading the AI-focused Genesis mission, the NSF is refining its strategy to champion these efforts.
“The short answer is it’s a work in progress,” Ellen Zegura, NSF’s acting deputy director for computer and information science and engineering, said of her agency’s relationship with the DOE at a news conference. public information gathering session last week. “We see elements that are both complementary and collaborative.”
The Genesis mission pursues several broad objectives, such as creating a national AI platform and improving national productivity linked to the research and development budget. Zegura said the NSF wants to find ways to strengthen the mission and identify areas of opportunity that are not necessarily covered by the executive order to launch the Genesis Initiative.
Strengthening the national AI workforce and contributing to skills development, for example, is an area that the NSF continues to monitor. The agency published a roadmap last year, with the aim of outlining the key skills and investments needed to facilitate the necessary STEM skills. NSF sought input from organizations and individuals across sectors on the identified avenues and recommendations.
Lawmakers have also tried to expand the agency’s staffing and training missions. In September, a bipartisan trio of House members reaffirmed legislation that would authorize the NSF to award AI-related scholarships and fellowship opportunities.
Despite its ambitions, the NSF faces the same obstacles to AI adoption as many agencies and private sector organizations: constraints on the chips that power AI and exorbitant costs associated with the technology.
Demand for AI chips continues to grow at an unprecedented rate. Nvidia, a leading supplier of ruggedized processors, has seen revenue soar and supply dwindle.
“Clouds are exhausted and our installed base of new and previous generation GPUs, including Blackwell, Hopper and Ampere, are fully utilized,” Colette Kress, Nvidia’s executive vice president and chief financial officer, said at the company’s conference. Third Quarter 2026 Earnings Call in November.
Controlling costs is also a priority for most organizations trying to advance adoption. The analyst firm Gartner estimates that there is 10 hidden costs for every AI tool organizations purchase, as well as the transition costs associated with training and change management.
NSF strives to find resource-conscious common ground.
“There will be capabilities and capabilities that will be at the highest level for which the NSF price will be excluded,” Zegura said. “It’s really important that there is broad capacity to get to the next level, because that’s where we provide more opportunities for new or disruptive ideas.”
The agency plans to facilitate access to a wide range of researchers in the “one tier down” range, Zegura added. If the NSF price is further away, the potential positive impact could diminish.
“It doesn’t have to be at the highest level, but you have to be within firing range,” Zegura said. “Otherwise, you don’t have a trajectory, a path between the testing of an idea and that of a large playing field.”
Reorganization, refocusing
Like other agencies, the NSF has experienced palpable changes over the past year.
The agency lost more than 500 workers in 2025 and has already lost a handful this year, according to the Office of Personnel Management’s new federal workforce data site. Of these workers, 30 were classified as IT managers.
The required reconfiguration at NSF resulted in an office bringing together advanced cyberinfrastructure, research infrastructure and other major projects, including AI institutes, Zegura said.
“We have made changes that bring AI Institute program leaders closer to [Office of Advanced Cyberinfrastructure] program leaders,” Zegura said. The evolved structure could enable experiments around alternative AI architectures that might not have happened when the groups were further apart, she added.
The AI institutes have been widely seen as a success story and a model for future initiatives, according to Zegura. AI Institutes date back to 2019 but have evolved since their first iterations.
“It’s really been a huge success, so we’re inclined to continue down this path,” Zegura said. “It’s led to some really interesting research, education and outreach efforts – sort of hits on everything we like to see.”
AI institutes are powered by partnerships, which “is not without challenges,” according to Zegura. Partner rewards are often long-term commitments, which can become thorny due to turnover or rate of innovation. The agency is learning from past mistakes, Zegura said, and is reworking partnership structures.
Still, AI institutes have managed to fill critical gaps, Zegura said.
“At first it was all just a gap,” she says. “Now we want to be as strategic as possible about where we make the additional investments and we’re really interested in getting input…from communities.” »
