To Counter China, America’s Military AI Needs an Open-Source Reboot

A merica cannot afford a misstep in the AI arms race — and the Pentagon’s rapidly deteriorating relationship with Anthropic is a wake-up call that we need to rethink our AI strategy before we lose our innovation edge.

Anthropic leadership and Secretary Pete Hegseth have been locked in a months-long stand-off over the use of Anthropic’s Claude by the U.S. military. The negotiations continued to escalate until things came to a head at the end of February, when President Trump directed not just the Pentagon but every federal agency to cease all use of the company’s technology.

The dispute reflects a deeper tension between Washington and Silicon Valley. While there are important debates that should be had about the proper use of AI in a war-fighting context, those discussions should be led by our elected officials and military leaders, not dictated by the private companies the Pentagon relies upon for services. Washington needs complete command of secure, customizable models capable of supporting classified missions. But that demand clashes with how most American AI firms are structured.

Thanks in large part to President Trump’s leadership, the defense enterprise had forged forward-looking partnerships with AI firms like Google, OpenAI, xAI, and Anthropic. Eyeing lucrative government contracts, each firm rolled out a government-specific model and partnered with the likes of Microsoft, Palantir, and Amazon Web Services to deploy their proprietary technology in secure cloud environments capable of handling classified information.

This system works — until it doesn’t. Anthropic’s fall from grace is a warning that America cannot rely on a handful of walled-garden companies to safeguard our national interests and guarantee AI dominance. Especially as China rapidly improves and exports its own AI stack.

For Anthropic, OpenAI, and other closed labs that refuse to disclose their training methods, secrecy is the default protocol. These firms test, train, and fine-tune models without independent external oversight before selling the finished product to customers, including military and intelligence agencies.

The Pentagon licenses use of this technology but depends on the model provider to refine and customize the backend for mission needs. It also relies on cloud partners to ensure the models can run in secure, air-gapped environments. In other words, the Department of Defense purchases the capabilities but outsources much of the testing, training, and enhancement. If even one player in this fragile ecosystem — as in the case of Anthropic — is deemed untrustworthy or, in one Pentagon official’s words, “undemocratic,” the system unravels, and our AI dominance is imperiled.

Meanwhile, China has taken a different path. Beijing has integrated DeepSeek — an open Chinese model engineered to run on less powerful hardware at a lower cost — into its military-civil fusion strategy. A Reuters review of patent and procurement notices filed by the People’s Liberation Army documented aggressive investment in DeepSeek-powered autonomous target recognition and real-time battlefield decision-making tools last year.

China is not only building everything from AI-enabled robot dogs to war-game simulators atop DeepSeek; it is also distributing the model globally. Unlike OpenAI or Anthropic, DeepSeek is free to download — no license requirements or usage constraints — and is quickly gaining traction in Russia, Iran, Cuba, Belarus, and Africa, strengthening China’s geopolitical posture.

Unburdened by a moral or legal compass, the Chinese are taking brazen shortcuts to run laps around American AI models. On separate occasions, both OpenAI and Anthropic warned regulators that DeepSeek breached terms of service and regional access restrictions to extract model secrets from Claude and ChatGPT using 24,000 fake accounts. According to Anthropic, these distillation attacks, “reinforce the rationale for export controls: restricted chip access limits both direct model training and the scale of illicit distillation.”

But tougher export controls won’t necessarily solve the problem. Reports that DeepSeek used an illegal supply of Nvidia’s most advanced Blackwell chips clustered in an Inner Mongolian data center to train their latest model confirm that even the strictest export restrictions are no obstacle to an adversary as determined as the Chinese. These findings also reveal that DeepSeek’s computing requirements are steeper and its cost advantage lower than previously thought. This gives the United States an opening.

With China rapidly improving and exporting its own cheap, ill-gotten models overseas, America must beat the Chinese at their own game by developing and exporting its own high-performing open-source models that we and our allies can use to counter these concerning trends.

The benefits would be immediate. Warfighters would gain reliable access to efficient, transparent systems designed to function in secure environments. The White House would gain a sharper edge in the AI arms race and new leverage at the negotiating table. And allies would see a clear signal that the United States, not China, is the dominant player and the country they should partner with in the global AI ecosystem.

America has two options: Cling to an inflexible closed architecture vulnerable to concentrated power and corporate whims, or reclaim strategic control through an open, resilient AI ecosystem built to serve national and sovereign interests. It’s time for America to get serious about open-source artificial intelligence or risk forfeiting the future of military AI to China.