DeepSeek: An inflection point in the new Cold War

Source: New York Times

Introduction

On January 20 2025, DeepSeek, a relatively unknown Chinese AI company, launched a chatbot based on its DeepSeek-R1 model, available for free on iOS and Android. Within two days, it surpassed ChatGPT as the most downloaded free app on the App Store, triggering an 18% fall in NVIDIA’s share price. To grasp the scale of this event: NVIDIA lost $589 billion in market capitalization in a single day—the largest one-day loss in US stock market history.

DeepSeek’s sudden rise not only alarmed US investors but also drew the attention of the US government. The day after NVIDIA’s record-breaking loss, the White House confirmed an intensifying National Security Council review of DeepSeek’s AI model.Immediately after, Senator Josh Hawley introduced a comprehensive US-China AI Decoupling Bill that, if passed, could ban the use of DeepSeek’s AI models in the US.  

While arguably necessary, these reactive measures highlight the US government’s lack of foresight—especially considering DeepSeek was founded in July 2023 and released its first model, DeepSeek Coder, on November 2, 2023. More concerning is that individual agencies and lawmakers responded more swiftly than the federal government, which is still struggling to craft a unified AI strategy. 

Behind the eightball 

A key issue  is the ineffectiveness of US export control on AI chips. With chip smuggling and stockpiling on the rise, export controls have failed to keep pace. For instance, NVIDIA developed the H800 chip, used by DeepSeek to train its latest model, specifically for the Chinese market, circumventing Biden-era regulations from 2022.

All NVIDIA had to do to get around the rules was manufacture downgraded chips that dipped just under the threshold the government had set. Although technically legal for Chinese firms, the chips delivered nearly equivalent performance. By the time the US government discovered this loophole, Chinese companies had already accumulated a hoard of NVIDIA’s most advanced chips. 

Although it’s unclear exactly how and when DeepSeek got their hands on NVIDIA H800s, there’s nothing that can be done now. Jimmy Goodrich, a senior adviser for technology analysis at the RAND Corporation, noted “Had the Biden administration more quickly responded and limited the H800 to China, there’s no doubt DeepSeek would have faced greater challenges in launching this model.”

The risks of escalation 

With the US government playing catch up on AI’s rapid development and China pouring billions into AI research, the likelihood of an escalation in the rivalry between the two nations is increasing.While it may seem exaggerated to liken the current geopolitical climate to the US–USSR nuclear arms race, the parallels are increasingly apparent.

The clearest similarity lies in the militarization of AI by both nations. Last November, the People’s Liberation Army (PLA) unveiled ChatBIT, built using Meta’s open-source Llama framework, to collect and analyze intelligence. Meta released a statement asserting that the PLA’s deployment of the Llama model was “unauthorized” and violated its acceptable use policy.  

The US military is also starting to ramp up its AI applications. Last December, the Pentagon's Chief Digital and AI Office launched Rapid Capabilities Cell (RCC) to collaborate with the Defense Innovation Unit on four initial frontier AI projects involving warfighting and enterprise management. The RCC is also tasked with building digital sandboxes to test AI integration into government networks.

As the two major powers clash for technological supremacy, like the US and USSR did in the past, the stakes keep getting higher. AI remains fragile; machine learning systems frequently err due to poor training data, algorithmic flaws, or unpredictable interactions.Since 2017, the Global Partnership on AI has identified “more than 1,200 reports of intelligent systems causing safety, fairness, or other real-world problems.”

Most of these errors are minor, but imagine what may occur if AI weapon systems malfunction or—worse— get hacked. For now, nuclear command and control remain in human hands; however, a day may come when AI maliciously overrides that authority. With cyberattacks surging 56 percent year-over-year, it wouldn’t be surprising if that day comes sooner than we think. 

Conclusion and what’s next 

Some experts predict that by 2030, AI could potentially add $600 billion to China's economy annually. Even if highly optimistic, such projections underscore China’s ambition to lead globally in AI. This leaves the US in a precarious position: either adopt a hard-nosed approach that limits China’s access to chips and risk provoking an all-out conflict or continue to act reactively and risk being overtaken.  

The US government should acknowledge that it cannot halt China’s AI progress. Thus, they should focus more on helping expand the global market share of American AI companies than getting sucked into a tit-for-tat clash with China. One approach is fostering cross-border partnerships among investors, entrepreneurs, and policymakers to encourage inclusive development. Microsoft has already taken action in this respect by investing $35 billion across 14 countries. Another strategy is to attract top AI scientists around the world by offering specialized visas and providing them with funding to conduct critical research. 

To end on a positive note, China also appears to recognize the need for cooperation over unchecked competition to avoid the pitfalls of AI. Last March, leading global AI experts convened in Beijing for the second International Dialogue on AI Safety. The consensus was clear: just as diplomacy averted thermonuclear disaster during the Cold War, global cooperation is vital to managing today’s transformative AI innovations.

---

Jeff Zhou possesses a strong track record in the venture capital and technology sectors. As Principal at Untapped Ventures, he has facilitated $20M+ in venture funding and oversees a diverse portfolio of 32 AI startups, driving strategic investment and go-to-market execution. His expertise extends to public-private partnerships, having co-designed the Untapped Future of Work Accelerator with the City of Burbank, a program that supports AI-driven workforce innovation. 

Outside of work, Jeff contributes to AI governance by advocating for responsible AI deployment. Fluent in Mandarin and with global experience across North America, Asia, and Europe, he brings a multinational perspective to AI’s role in shaping the future of industries and infrastructure.