All major news outlets are covering the launch of DeepSeek R1 and how it compares to the latest releases from established AI companies—like OpenAI’s GPT o1—while spending far less than these companies have on their own flagship models.
Adding fuel to the conversation, Nvidia’s stock price declined by 15% in yesterday’s trading compared to last week, hinting that investors may believe Nvidia and other American companies could be losing the race.
Geopolitics
At first glance, it might seem that this Chinese company has successfully bridged the gap separating it from well-known American firms in the field. However, after reading the research paper accompanying the DeepSeek R1 announcement, and looking into the matter, it appears that this is not the case.
Major players such as OpenAI, Anthropic AI, Google, Meta, and other American tech companies still comfortably lead the pack, leaving their global competitors—including the Chinese company behind DeepSeek R1—well behind.
Here’s are some arguments supporting this opinion:
- Semiconductor and Microchip Technology: The latest semiconductor tech from companies like Samsung and TSMC remains several years ahead of anything China can currently produce.
- Export Restrictions: The United States is further tightening restrictions on exporting semiconductor technology to China and is quietly discouraging its citizen scientists and tech experts from working there, reminiscent of Cold War-era practices. For more insights, I recommend Chip War: The Fight for the World’s Most Critical Technology (2022), which discusses this undeclared conflict and draws parallels to the Cold War of the 1970s and 1980s.
- Reproducing the same results will not be difficult. Companies like Google, Anthropic, and OpenAI can adopt the same techniques used by DeepSeek R1 to boost their own products, allowing them to surpass DeepSeek R1 swiftly at least in training or inference time.
- DeepSeek R1 doesn’t offer anything fundamentally new. Meanwhile, genuine scientific breakthroughs —often pioneered by American companies— continue to emerge. A prime example is the paper “Titans: Learning to Memorize at Test Time”, published the 31st of December 2024… Perhaps to end the year on a high note?
Supply and Demand
And about the viewpoint suggesting a potential drop in demand for more advanced chips. The reasoning is that DeepSeek R1 has demonstrated how older hardware can achieve better results, potentially reducing the need for cutting-edge processors.
I, however, do not believe that demand for advanced chips will wane. On the contrary, with the lowered barrier to market entry, demand for high-end chips will grow both vertically and horizontally, for several reasons:
1- Medium and large companies that previously hesitated to enter the AI race—largely due to high investment costs without a guaranteed return—can now operate their own AI servers at a significantly lower cost than before.
2- As the hardware requirements for running AI technologies decrease, the prospect of moving inference tasks to edge devices (such as smartphones) becomes more realistic. This situation is reminiscent of the early 1990s, when computing began shifting from centralized mainframes and terminals to personal computers, a transition that drove massive growth in the chip market and continues to do so today. Which is known as the Jevons Paradox.
Further reading:
- Behrouz, A., Zhong, P., & Mirrokni, V. (2024). Titans: Learning to Memorize at Test Time.
- Miller, C. (2022). Chip War: The Fight for the World’s Most Critical Technology.