Nvidia’s stock could fall as Deepseek’s ‘amazing’ AI model destroys Openai

American policy to restrict China’s access to Nvidia’s most advanced AI chips unintentionally helps Chinese AI developers, who have full access to the company’s latest chips, leapfrog American rivals I did.

This proves the fundamental reason why startups are often more successful than large companies. Scarcity breeds innovation.

One example is the Chinese AI model DeepSeek R1. It’s a complex problem-solving model that competes with Openai’s O1, which has “zoomed into the global top 10 for performance,” but was also built much faster and with fewer, more powerful AI chips. According to the Wall Street Journal, at a much lower cost.

R1’s success should benefit the company. That’s because companies see no reason to pay more for effective AI models when cheaper models are available and are likely to improve more quickly.

“Openai’s model is great for performance, but we also don’t want to pay for capabilities we don’t need,” says Anthony Poo, co-founder of a Silicon Valley-based startup that uses generative AI to generate financial returns. Predict. journal.

Last September, Poo’s company moved from Humanity Claude to Deepseek after tests showed it “works as well at about a quarter of the cost.”

When my book, Brain Rush, was published last summer, I was worried that the future of generative AI in the United States was becoming too dependent on the largest technology companies. This was in contrast to the creativity of US startups during the DOT-COM boom. This resulted in 2,888 initial public offerings (compared to zero IPOs for US generative AI startups).

Deepseek’s success could spur a new rival to large language model developers in the US. As these startups build powerful AI models with fewer chips and market improvements faster, NVIDIA’s revenue will increase as LLM developers replicate DeepSeek’s strategy of using fewer, more advanced AI chips. may grow more slowly.

“We decline comment,” an Nvidia spokesperson wrote in a Jan. 26 email.

Deepseek’s R1: superior performance, low cost, and short development time

Deepseek was impressed by a major US venture capitalist. “Deepseek R1 is one of the most amazingly impressive breakthroughs I’ve ever seen,” Silicon Valley venture capitalist Marc Andreessen wrote in a January 24 mailer.

To be fair, Deepseek’s technology lags behind our rivals like Openai and Google. However, the company’s R1 model, launched on January 20th, is a close rival, even though it uses “fewer and less sensible chips, and in some cases, US developers say it is essential.” Skip the steps you’re thinking about,” says the journal.

The high cost of deploying generative AI has left businesses wondering whether they can positively capture a return on investment. As I wrote last April, more than $1 trillion could be invested in technology, and the killer app has yet to emerge.

Companies are therefore excited about the prospect of reducing the required investment. Enterprises are very interested in R1’s open source model because it works very well and is much cheaper than Openai or Google’s.

Why? R1 is the top trending model downloaded on Huggingface (109,000) according to VentureBeat, matching “Openai’s O1 at just 3%-5% of the cost.” R1 also offers a search function that users judges will find better than Openai or Puzzle.

Deepseek developed R1 faster and at a much lower cost. Deepseek said it trained one of its latest models for $5.6 million. This is far below the $100 million to $1 billion range quoted by CEO Dario Amodei in 2024 for the cost to train the model, he said.

To train the V3 model, DeepSeek used a cluster of over 2,000 Nvidia chips.

Independent analysts at Chatbot Arena, a platform hosted by UC Berkeley researchers, rated the V3 and R1 models in the Top 10 for Chatbot Performance on January 25th.

The CEO behind Deepseek is Liang Wenfeng, who manages an $8 billion hedge fund. His hedge fund, called Highflyer, uses AI chips to build algorithms to identify “patterns that can influence stock prices,” the Financial Times says.

Lian’s outsider status helped him succeed. In 2023, he launched DeepSeek to develop human-level AI. “Liang has built an exceptional infrastructure team that really understands how chips work,” one founder of a rival LLM company told the Financial Times. “He brought the best people with him from hedge funds to Deep Seek.”

Deepseek benefited when Washington banned Nvidia from exporting Nvidia’s most powerful chips to China. According to CNBC, local AI companies were forced to engineer the limited computing power scarcity of less powerful local chips (NVIDIA H800S). Liang’s team “already knew how to solve this problem,” the Financial Times said.

Microsoft is very impressed with Deepseek’s work. “To see DeepSeek’s new model is very impressive, both in terms of how effectively they have actually done this open source model that does the inference time calculations,” World Economic Forum. “We should take development out of China very seriously.”

Will Deepseek’s breakthrough slow demand growth for Nvidia chips?

Deepseek’s success should spur changes to US AI policy while making Nvidia investors more cautious.

US export restrictions on Nvidia have put pressure on startups like DeepSeek to prioritize efficiency, resource pooling, and collaboration. To create R1, DeepSeek Reeingineer created a training process to take advantage of the NVIDIA H800S’s lower processing speed. One half of H100S, former Deepseek employee and current PhD student, spoke to MIT Technology Review at Northwestern University Zihan Wang.

An Nvidia researcher was enthusiastic about Deepseek’s work. The Deepseek paper reports that the results “bring up memories of pioneering AI programs that mastered board games like chess built ‘from the ground up’ without first imitating a human grandmaster.” Senior Nvidia research scientist Jim Fan spoke about X as featured in the journal.

Will Deepseek’s success moderate Nvidia’s growth rate? I don’t know. But based on my research, companies want powerful generative AI models that clearly pay off. When companies seek high-wage generating AI applications, they can experiment more if the cost and time to build those applications is low.

As such, R1’s lower cost and lower performance should continue to attract more commercial interest. Key to DeepSeek’s ability to deliver what businesses want is its skill in optimizing for more powerful GPUs. This is lower than state-of-the-art chips.

If more startups can replicate what DeepSeek has accomplished, there could be less demand for Nvidia’s most expensive chip.

I don’t know how Nvidia would react if this happened. However, in the short term, DeepSeek’s strategy could mean less revenue growth as startups build models with lower-priced chips.