Days after China’s DeepSeek detailed an approach to generative AI that needs just a fraction of the computing power used to build prominent U.S. tools, the global conversation around AI and national security—from how the Pentagon buys and uses AI to how foreign powers might disrupt American life, including privacy—is changing.
DeepSeek’s announcement drew a collective wail from the White House, Wall Street and Silicon Valley. In Washington, D.C., President Trump called it a “wake-up for our industries that we need to be laser focused on competing” against China. White House press secretary Karoline Leavitt said the National Security Council is currently reviewing the app. The Navy has already banned it. On Wall Street, chip maker Nvidia’s stock tumbled. OpenAI, DeepSeek’s closest U.S. competitor, is crying foul and claiming the app essentially distills their own model.
If you believe the United States “must win the AI competition that is intensifying strategic competition with China,” as former Google chairman Eric Schmidt and former Deputy Defense Secretary Robert Work wrote in 2021, then DeepSeek is a big deal.
Why is DeepSeek so significant? For one thing, it’s much more open-source than other models. But the defining technical innovation lies in the model’s ability to distill advanced reasoning capabilities from massive models into smaller, more efficient counterparts. One DeepSeek model often outperforms larger open-source alternatives, setting a new standard (or at least a very public one) for compact AI performance.
DeepSeek relies heavily on reinforcement learning to develop reasoning skills, sidestepping the supervised fine-tuning typically used in the initial phases by competitors like OpenAI. This approach is a deliberate divergence from the hybrid training strategies employed by U.S.-based AI giants.
Benchmark results described in the paper reveal that DeepSeek’s models are highly competitive in reasoning-intensive tasks, consistently achieving top-tier performance in areas like mathematics and coding. However, the research highlights some vulnerabilities as well, particularly in non-reasoning tasks and factual query accuracy, where it falls short of OpenAI’s most advanced offerings.
No one has independently verified that DeepSeek isn’t using large compute resources to achieve its benchmark results (or has not essentially copied OpenAI), but U.S. controls on highly advanced microchips would limit the resources available to China.
Alex Wang, CEO of Scale AI, whose company also evaluates AI models, described DeepSeek as comparable to OpenAI in a CNBC interview. He also said China has obtained roughly 50,000 of Nvidia’s H100 chips despite export controls.
An Nvidia spokesperson didn’t address the claim directly. He told Defense One: “DeepSeek is an excellent AI advancement and a perfect example of Test Time Scaling,” a technique that increases computing power when the model is taking in data to produce a new result. The extra compute power allows the model to explore different options and improve their answers, thus reaching better answers with less training (less compute.) The model can then focus its computational energy more effectively. It’s sort of like exercise: At first, working out depletes energy, but in the longer term it helps the body build the capacity to store and more effectively use energy.
“DeepSeek’s work illustrates how new models can be created using that technique, leveraging widely-available models and compute that is fully export-control compliant. Inference requires significant numbers of NVIDIA GPUs and high-performance networking. We now have three scaling laws: pre-training and post-training, which continue, and new test-time scaling,” the Nvidia spokesperson said.
The development represents a fundamental shift in the discussion of how to build AI dominance. While companies like OpenAI achieved their results based on huge data sets, very large models, and ever-expanding computer resources, the next phase of AI will likely usher in smaller models that need fewer compute resources.
That might bode poorly for large enterprise cloud providers, including many of the tech giants whose leaders attended Trump’s inauguration. Many companies were counting on huge demand for resource-hungry generative AI products—and were squeezing out alternative approaches. But the change in discussion around how to build AI could be good news for troops who want to tap into the most robust tools in places where power and connectivity to big cloud resources are patchy. And it could also be helpful for a Defense Department tasked with capturing the best AI capabilities while simultaneously reigning in spending.
A new, smaller future for artificial intelligence
AI researchers who were attempting to chart a very different path than that of OpenAI and the big enterprise cloud providers were not surprised by the DeepSeek breakthrough.
Data scientist Drew Breunig told Defense One, “If there’s a lesson from DeepSeek’s triumph, it’s this: be wary when the route to progress is simply spending more money. This path fosters no innovation and your poorer competitors will be forced to get creative, work within their constraints, and eventually…they’ll win. Spending is not innovating.”
In a recent blog post, he described how synthetic data can reduce the amount of raw data—and compute power—needed to produce high-performing models. “This tactic benefits smaller models at the same rate as large ones,” he said.
Pete Warden, CEO of AI startup Useful Sensors, told Defense One, “DeepSeek demonstrates that spending more and more money on larger and larger models isn’t the only approach to improving AI. TinyML is based around the idea that using smaller models that are cheaper to train, we can build applications that have a big impact, despite their size.”
But Stanford AI researcher Ritwik Gupta, who with several colleagues wrote one of the seminal papers on building smaller AI models that produce big results, cautioned that much of the hype around DeepSeek shows a misreading of exactly what it is, which he described as “still a big model,” with 671 billion parameters.
“However, it is very notable that the DeepSeek-R1 team offers first-party ‘distilled’ versions of their models,” Gupta told Defense One. “What DeepSeek has done is take smaller versions of Llama and Qwen ranging from 1.5-70 billion parameters and trained them on the outputs of DeepSeek-R1. This allows an ‘R1-like’ model to work on smaller devices, like laptops or phones.”
DeepSeek’s performance—insofar as it shows what is possible—will give the Defense Department more leverage in its discussions with industry, and allow the department to find more competitors.
“I would not be surprised to see the DOD embrace open-source American reproductions of DeepSeek and Qwen,” Gupta said. “The DOD has always had the pull to ask for special, on-premise versions of otherwise cloud-only service offerings. I would not be surprised if they make this ask of OpenAI and Claude.”
Heidy Khlaaf, the chief AI scientist at the AI Now Institute, focuses her research on AI safety in weapons systems and national security. She told Defense One that the breakthrough, if it’s real, could open up the use of generative AI to smaller players, including potentially small manufacturers. But such models will never be suitable for combat, she said, despite an eagerness to employ them in such contexts.
“In general, LLMs or foundation models are not suited for safety-critical tasks given how error-prone they are with applications requiring dependability and precision. However, the size and capabilities of DeepSeek does open up the use of foundation models to smaller actors who previously may have not had access, and that may include car manufacturers who may be interested in using foundation models in a non-safety critical way,” Khlaaf said.
Andrew Redding, who leads the technology and cybersecurity portfolios at Berkeley’s Center for Security in Politics, told Defense One, “The performance of DeepSeek is entirely unsurprising for those of us who have been tracking how AI researchers are able to develop models with decreasing amounts of compute.”
American companies should see the breakthrough as an opportunity to pursue innovation in a different direction, he said. “Interestingly, the compute challenges faced by Chinese researchers (in light of U.S. export control on NVIDIA GPUs) are not dissimilar to those that U.S. academics are facing given that we are increasingly compute-constrained compared to the players in private industry.”
The United States military is already spending significantly on edge capabilities to get computing power as close to warfighters as possible. The smaller model performance breakthrough suggests that those edge-computing investments have increased in value, Redding said.
“There’s also a really interesting question as to the use of open as opposed to closed models within the military context,” he said. “The advantage of the former is that they are easy to move inside of government networks to leverage gov/mil data, but there are the obvious risks of adversary states getting their hands on the training data, model weights, etc.”
But perhaps the most important take-away from DeepSeek’s announcement is not what it means for the competition between the United States and China, but for individuals, public institutions, and anyone skeptical of the growing influence of an ever-smaller group of technology players. It’s good news if you want to build your own generative AI tool, with data you control, rather than rely on a tool from a big company that may or may not have your best interests at heart.
”The internet has historically thrived as a decentralized set of services,” Gupta said. If the goal is to get everyone to have their own ‘personal AI’, then it will be necessary for small models to run on people’s personal devices. I expect companies like Apple, who have a privacy-first model, to continue to push for offline, disconnected algorithms.”
But Khlaaf warns that substituting big models for distilled ones poses individual privacy risks that apply to troops, too, as exposure of personal data affects them just as it does civilians, making them vulnerable to adversarial targeting, coercion, etc.
And the broad exposure of Americans’ personal data is in itself a national vulnerability that adversaries could use in the event of conflict, as military leaders have pointed out. Without comprehensive reform to help individuals better safeguard their own data, a proliferation of robust small models like DeepSeek could make a bad trend worse.
“DeepSeek challenges the idea that larger scale models are always more performative, which has important implications given the security and privacy vulnerabilities that come with building AI models at scale,” Khlaaf said.
For personal privacy, “distillation techniques allow the compression of larger models into smaller ones while preserving many of the properties of the larger model. For citizens who had foundation models train on their data, all of the same privacy issues would be perpetuated into DeepSeek’s distilled models—only now not under U.S. jurisdiction. That’s why we’ve warned that training AI models on sensitive data poses a national security risk.”
Jesus Christ is King
