DeepSeek sparks AI debate: High performance on a budget

Chinese AI company DeepSeek coming out of nowhere and shaking the cores of Silicon Valley and Wall Street was something no one anticipated. Commentators had previously placed China’s AI scene 2–3 years behind that of the US - words they are now eating.

What has been widely highlighted about DeepSeek and its AI model R1 is that it was allegedly built with only US$5.6 million in two months, using old Nvidia chipsets. In contrast, OpenAI, Google, and Meta collectively pumped US$200 billion into AI development in 2024 alone, seeing around US$25 billion in revenues, according to Counterpoint Research.

OpenAI CEO Sam Altman described DeepSeek’s R1 as an “impressive model,” acknowledging its rival’s tighter budget and welcoming the entry of a new competitor.

There are many challengers for OpenAI to contend with, but only a handful pose a credible threat. DeepSeek may not surpass OpenAI in the long run due to embargoes on China, but it has demonstrated that there is another way to develop high-performing AI models without throwing billions at the problem.

Counterpoint Research director and AI/IoT lead Mohit Agrawal pointed this out, stating: “DeepSeek has shown a path wherein you actually train a model in a much more frugal way,” which will have a widespread positive effect on various sectors (just not Nvidia, for now).

This shift is already evident, as Nvidia’s stock price plummeted, wiping around US$593 billion—17% of its market cap—on Monday.

“Investors will start asking questions, and there will be a change in mindset now. Instead of saying, ‘let’s put more computing power’ and brute-force the desired improvement in performance, they will demand efficiency.

“For example, if this year Microsoft sets a budget of US$80 billion for its data centres but Meta decides on US$65 billion, the question will arise—are they investing at the right level?

“What we should be taking away from DeepSeek is that the industry needs to explore better and more efficient ways of running AI,” said Agrawal.

What does this mean for Nvidia?

For the last two years, as AI momentum surged, some analysts warned that investing in the technology was a money trap, given that only one company (rhymes with Lydia) was making significant profits across the ecosystem. Meanwhile, hyperscalers such as Google, Meta, and Microsoft poured billions into the AI arms race.

Agrawal argued that this was not “healthy,” but as the new trend of efficiency and frugality gains traction, he predicts it will drive down the cost of AI technology, enabling industries such as telecoms to adopt AI and unlock new revenue-generating use cases.

“Hyperscalers were losing big on AI, and further down the enterprise chain, companies were cautious about AI but recognised its potential. As AI costs continue to fall—something we were already seeing before DeepSeek—smaller companies will be able to adopt it more widely.

“If adoption rises while the need for excessive compute power decreases, then more companies in the value chain will start making money. AI will become much more distributed—and if it doesn’t, that’s a problem,” said Agrawal.

He added that while Nvidia is taking a financial hit in the short term, growth will return in the long run as AI adoption spreads further down the enterprise chain, creating fresh demand for its technology.

Most companies will not be able to replicate the foundational work that giants like Meta and Google have invested in to kickstart their AI journeys. Instead, Agrawal noted that industries such as telecoms will benefit from AI through SaaS providers, who will enhance their services with more affordable AI solutions.

Can DeepSeek continue its challenge to ChatGPT?

DeepSeek has shown remarkable ingenuity - so much so that OpenAI’s chief executive, Sam Altman, has praised its ability to achieve so much with limited resources. However, Agrawal argued that DeepSeek won’t be able to keep pace with ChatGPT in the long term, as US restrictions on selling advanced technology to Chinese firms continue to tighten.

“DeepSeek claims they trained the model on a small pool of old Nvidia chipsets. To maintain that momentum, they will need access to higher-capability chips. It will be difficult for them to keep moving at the same pace without access to high-end chipsets,” said Agrawal.

Whether DeepSeek is here to stay for the long term - or whether geopolitical tensions will cut its trajectory short - remains to be seen. What it has achieved with limited resources is nothing short of phenomenal (if its claims hold true).

DeepSeek has been a trigger for making AI more affordable for telecoms and other industries. We can expect to see more innovative applications and services from telecom players as global AI innovation continues. But how much of that progress will be hamstrung - or even accelerated - by geopolitical wrangling between the US and China? The coming months will offer a clearer picture as the AI landscape evolves at breakneck speed.