Last weekend, DeepSeek launched its new large language model, the DeepSeek R1. Right away, it shot to the top of the iPhone downloads on the Apple App Store, challenging ChatGPT. What makes R1 stand out? It’s open source and leverages something called “cold start data.” Instead of scavenging the internet for data, it uses reinforced learning to enhance its accuracy.
On their GitHub page, DeepSeek’s team explains that R1 applies reinforcement learning to the base model directly. This lets the model explore problems in a more complex way. Over the weekend, DeepSeek garnered about 2.1 million searches, with 1.6 million just on Sunday—12.3% of ChatGPT’s traffic during the same period. This surge in interest is fueled not just by its unique approach but also by competitive pricing and open-source access.
To put it in perspective, while OpenAI charges $2.50 per million tokens for its GPT-4o, DeepSeek comes in at just $0.14 when tapping into cached info. Non-cached data costs $0.55 per million tokens. This price difference has shaken tech stocks in the U.S., notably Nvidia, which saw a 17% drop in share price—effectively reducing its market cap by $593 billion.
In a speech on Monday, former President Donald Trump described DeepSeek as a wake-up call for American tech. He talked about revoking AI regulations from the Biden era, emphasizing the need for fewer restrictions so companies can thrive. Trump pointed to DeepSeek as a sign that the U.S. must ramp up competition in AI, particularly against Chinese firms moving quickly and economically.
DeepSeek’s innovative algorithms have significantly cut energy demands for training and using AI. Trump lauded this achievement, arguing that it allows firms to develop applications without hefty expenses, calling it a positive asset. The development process was quick, taking just two months and under $6 million, using Nvidia’s lower-tier H800 chips, especially notable given the recent export bans on high-end Nvidia GPUs to China.
Charu Chanana, chief investment strategist at Saxo, highlighted how U.S. tech stocks, including Nvidia, Microsoft, and Alphabet, are priced at premium valuations. Any disruption, like DeepSeek’s emergence proving advanced AI can be built on less capable hardware, could impact these stock prices.
Nigel Green, CEO of DeVere Group, noted that DeepSeek’s achievements signal a pivotal shift toward energy-efficient AI, potentially redefining both the AI and energy markets. This could challenge assumptions about the growth of AI being linked to skyrocketing energy use. Despite short-term market uncertainties, Green believes that efficiency-driven models will broaden adoption across various industries, leading to sustained demand for energy solutions.
Ultimately, DeepSeek’s ability to deliver results with less expensive hardware and competitive pricing might reshape the AI landscape. Kjell Carlsson, head of AI strategy at Domino Data Lab, argued that this shows that strategy can outpace mere computational might. This scenario underscores how open-source innovation is leveling the playing field, allowing newcomers to challenge established players in the generative AI space.
DeepSeek’s approach certainly highlights the rising competition to Silicon Valley’s AI dominance. Michael Guan, CEO of Final Round AI, pointed out that DeepSeek’s profitable pricing model and open-source code are unconventional in a landscape where companies often keep their innovations closely guarded.