Defense Budget Forces Tough Compute Decisions as AI Evolves

Defense officials are grappling with the challenges of finding more financial and power resources for AI implementation as the technology becomes more advanced and integral to government operations.

“The processing and compute can get expensive very quickly,” Marine Corps Cyberspace Command Cyber Technology Officer Shery Thomas told GovCIO Media & Research. “We have to make a deterministic business case analysis on which efforts are worthwhile to pursue and which increase user productivity in the long run.”

Agencies are spending more on AI technology each year. According to a 2024 Brookings Institution report, the potential value of AI-related federal contracts increased by almost 1,200%, from $355 million in August 2022, to $4.6 billion in August 2023. Most of that increase came from the Defense Department and its components. The potential value of the Pentagon’s AI-related contracts rose from under $200 million in August 2022 to more than $550 million in August 2023, according to Brookings.

Even with that jump, budgets for new technology are strained.

“It is clear to me that more resources will be required,” said Air Force Secretary Frank Kendall at the Air & Space Forces Association’s Air, Space & Cyber Conference. “The Space Force is beginning a transformation that must be executed quickly and at scale. That takes resources. The Air Force must move to a new generation of more competitive capabilities as quickly as possible.”

With limited budgets, Defense components need to make decisions about which AI systems actually augment mission, Air Force National Laboratory CIO Alexis Bonnell said at the Air, Space & Cyber Conference.

“[We are] recognizing that, when we navigate a new technology or emerging technology, it isn’t a question of having to redo everything,” said Bonnell. “It’s really asking ourselves ‘what is additive? What is actually different and unique?'”

The Marine Corps is increasingly using AI. Decisions about different AI systems rely on use cases and return on investment, Thomas told GovCIO Media & Research.

“We have to get to the specific use case on where artificial intelligence or machines can augment the user in their tasks. Additionally, it should provide value added ROI and these become the factors in judging if a business or warfighting mission area should be supplemented with AI,” said Thomas. “The processing and compute can get expensive very quickly and therefore we have to make a deterministic business case analysis on which efforts are worthwhile to pursue increases on user productivity in the long run.”

Military services use risk-based approaches to make budget decisions about AI system acquisition and implementation, Army Principal Deputy Assistant Secretary for Acquisition, Logistics and Technology Young Bang said recently. Those tradeoffs need to be explained and examined by all involved.

“We need to inform our consumers, our operation operators, our commanders, to say, ‘this is the technology, help us understand the mission or the use case that you want resolved,’“ said Bang during the NVIDIA AI Summit in Washington, D.C. earlier this month. ” We can make a risk-based decision. [The Army wants] to enforce and push and run around algorithms and generative AI. But we want to do it carefully and very deliberately.”

Thomas said that AI’s ubiquity creates conundrums for IT officials, including the question of which systems to buy and where to add the systems.

“As AI technology proliferates, we have to make careful decisions on which systems are capable to add this capability,” said Thomas. “Legacy system data will need to be tagged, analyzed, synchronized, and federated so it can processed with AI tools… We have to keep pace with the new technology to stay on top of it – all of which require additional budget.”

Processing Power Needed

Money is not the only scarce resource. AI uses more processing power than traditional applications and government organizations have to find the computing to run AI programs. According to a Feb. 2024 report by the Centre for the Governance of AI, “the amount of compute used to train leading AI systems has increased by a factor of 350 million.” The scarcity of compute is a challenge that faces AI implementation as much as funding does, according to officials.

“We’ve got to answer a lot of questions about [boosting] infrastructure,” said Bonnell at the NVIDIA event. “We had to pull a couple miracles with some folks in this room, when we started hitting compute demands.”

Agencies cannot count on miracles, however, to solve compute problems. At Space Force, the budget issue and the compute issue dovetail. Space Force Data and Artificial Intelligence Officer Chandra Donelson said at the NVIDIA event.

“We need industry’s help…understanding what our compute and infrastructure needs are, and then helping us map out our strategy to be able to scale that across the department,” said Donelson. “We do want to invest heavily into compute this year for fiscal year 2025. It’s something that is a top priority for me.”

National security uses AI more each year, bringing more data to train systems. More compute power is needed daily, Thomas said, however it be summoned.

“The budget concerns have increased as the data increases each day and therefore the processing and compute required increases exponentially to be able to have machines make decisions on data sets,” said Thomas.

Creative thinking is integral to solving the compute resource problem, national security leaders said. At Defense Information Systems Agency, CTO Stephen Wallace told GovCIO Media & Research that cloud computing and lessons learned can help solve problems with computer power scarcity.

“[Cloud computing] is helping also to bring the resources though to the AI game that we wouldn’t have if we were trying to do all of this on our own,” Wallace said. “[DISA has] thousands and tens of thousands of [ultrapowerful processors] that are out there actively training models and working through data sets and the ability to store the data in mass, you know, within the cloud environments. It is an absolute game changer.”

As the National Security Agency (NSA), sharing resources at speed is critical for AI implementation. NSA Director Gen. Timothy Haugh said that collaborative work in AI implementation can help compute power stretch across organizations.

“We’ll also look to share compute resources to allow us to scale and then do so faster. So I think those opportunities for us can be unique in the department, with our workforce, and with the way that we apply both our experience in AI and ML,” Haugh said during an Intelligence and National Security Alliance leadership dinner in August. “Certainly NSA has done that for decades. And but now how do we bring that to other partners in the government? We think there’s still a lot more we can do.”

Ultimately, each agency or organization in the national security ecosystem needs to take the fullness of the mission into any decision about AI and resources like compute power and budget.

“The basis of any AI informed system is the ability of the data to be visible, accessible, understandable, linked, transparent, interoperable and secure,” Thomas said.