China’s rapid AI-driven modernization exposes a US vulnerability: slow procurement cycles. Speed will determine strategic readiness in the coming decade.
China’s DeepSeek recently matched OpenAI’s capabilities while claiming just $5.6 million in compute costs for their base model—training it in 55 days on 2,048 H800 graphic processing units (GPUs) acquired before export restrictions. Though the true development costs are far higher when including infrastructure and research and development (R&D), the speed of execution is what matters: they went from concept to deployment in months, not years. This same velocity advantage extends across China’s military modernization. They fielded hypersonic weapons in 2019 while we’re still years from deployment.
Our defense acquisition programs average 11-12 years from requirement to fielding. In that time, adversaries iterate through multiple generations of technology. This isn’t just bureaucratic delay—it’s a strategic vulnerability that compounds as technological change accelerates.
The Real Challenge: Beyond the Obvious
Most discussions about artificial intelligence (AI) in defense focus on natural language processing—the ChatGPT effect. Procurement seems like the obvious target for this solution because it’s document-heavy, and everyone now understands that AI can process text. But this barely scratches the surface of what’s possible.
Through my team’s work with the US Air Force on their Small Business Innovation Research (SBIR) program, we discovered a more fundamental issue: Department of Defense (DOD) personnel don’t know where to begin with AI implementation. It’s not their fault. Over the past five years, they’ve received 18 different memos and guidance documents about acquiring AI—each adding requirements, none providing a unified framework. No acquisition professional can keep up with that volume of contradictory guidance while managing actual programs.
The knowledge gap extends beyond procurement. Most acquisition professionals understand that ChatGPT can answer questions, but they’re unaware of deep learning architectures that could predict program risks, or agentic AI systems that could autonomously monitor supply chains. They don’t know about model ensembles that could combine multiple AI approaches to solve complex problems, rather than just answering natural language queries.
Technical Realities of Defense AI Implementation
The F-35 program—the DOD’s next-generation stealth fighter—is a great example of systemic US procurement challenges. Total projected acquisition and sustainment costs now exceed $2 trillion through 2088, with sustainment costs alone rising from $1.1 trillion in 2018 to $1.58 trillion in 2023, according to the US Government Accountability Office. These aren’t just management failures—they’re symptoms of a system that can’t process complexity at the speed required.
Milestone reviews for major defense programs consume 5,600 staff days over two-year periods. Program managers tell us only 10 percent of the required documentation actually supports decisions. The rest is compliance theater. For AI specifically, the problem is even worse—acquisition officers face 18 separate guidance documents just for AI procurement, with overlapping and contradictory requirements that significantly extend timelines.
Natural language processing can help with document analysis, but real transformation requires deeper AI integration. Predictive models could identify program risks before they cascade into delays. Graph neural networks could identify bottlenecks in the defense industrial base before they impact production. Reinforcement learning could determine the best way to allocate resources across program portfolios.
What We Learned from the Air Force SBIR
Our prototype revealed that the acquisition workforce needs more than just AI tools—they need intuitive platforms that don’t require data science degrees to operate. We built advanced AI tools to read and interpret regulations, but the real breakthrough was creating easy-to-use interfaces that let non-technical people actually work with that AI.
The system had to hide the complexity of the underlying models while exposing their power. Acquisition officers could query compliance requirements in plain English, but behind the scenes, the platform leveraged ensemble methods combining multiple specialized models—one for regulatory interpretation, another for precedent analysis, a third for risk assessment.
More importantly, we learned that training and upskilling are critical. With properly designed platforms and targeted training, acquisition professionals can quickly learn to leverage AI capabilities they didn’t know existed. The technology is only as good as the workforce’s ability to use it.
International Context: Speed as Strategy
China deployed 360 intelligence satellites by 2024, up from 36 in 2010. Their Jilin-1 constellation tracks objects as small as cars with 95 percent precision—a sevenfold improvement in capability. They’re not just building more; they’re iterating faster.
Ukraine transformed procurement out of necessity. They produced 2 million drones in 2024 with 96 percent domestic manufacturing. Seven hundred military units gained direct commercial purchase authority. Their acquisition timeline: weeks, not years.
These aren’t just statistics. They represent fundamentally different approaches to capability development—approaches that prioritize speed over perfection.
Security Without Paralysis
Defense AI requires air-gapped systems, classification-appropriate models, and human oversight. We can train AI models by learning from data where it already lives, without having to gather or store all the sensitive information in one place. Differential privacy protects program information while enabling pattern recognition. Adversarial testing validates model robustness.
The real risk isn’t AI compromise—it’s obsolescence. Perfect security on outdated systems provides no operational advantage.
Implementation Reality
The US government is making big moves to secure its lead in AI—spending heavily on advanced systems, putting GPT-4 to work in secure classified settings, boosting the Pentagon’s AI office, and issuing an AI Action Plan and executive orderto cut red tape so AI can be adopted faster.
But technology alone won’t solve this. Success requires cultural change: accepting iterative development, integrating commercial innovation, and recognizing that deployment speed matters more than requirement perfection.
The acquisition workforce needs platforms that make advanced AI accessible without requiring extensive retraining. They need unified guidance, not 18 contradictory documents. They need leadership that understands both AI’s potential and its limitations.
The Strategic Timeline
China is racing to build advanced AI-driven military capabilities by 2027–2030, while the United States risks falling behind because its procurement process is too slow. By the time the United States finishes writing requirements, China could already have working systems in the field.
The path forward requires moving beyond surface-level applications like chatbots to deep integration of ensemble models, predictive analytics, and autonomous systems. It requires upskilling our workforce to understand and leverage these capabilities. Most critically, it requires accepting that good enough today beats perfect tomorrow.
In modern strategic competition, acquisition speed equals strategic readiness. The question isn’t whether to adopt AI, but whether we can transform our culture fast enough to remain competitive.
About the Author: Rick Hubbard
Rick Hubbard is Chief Scientist at Core4ce and leads the Autonomy, Artificial Intelligence, and Machine Learning (AAIM) Lab.
Image: TSViPhoto/shutterstock
