Non-Western founders say DeepSeek is proof that innovation need not cost billions of dollars

On January 20, DeepSeek, a 2-year-old company, unveiled an artificial intelligence model that it claimed to have trained with less than $6 million. Within a week, its chatbot app reached the top rank on Apple’s App Store and the Google Play store in the U.S. DeepSeek’s model fares at-par with other popular AI models in terms of performance, according to an independent AI quality index.

OpenAI has raised nearly $18 billion to date. In April 2023, Altman said the company had spent over $100 million just to train ChatGPT’s latest version, GPT-4.

The conversation around DeepSeek in the West has ranged from excitement and surprise to skepticism about the veracity of the low-cost claims, the lack of clarity around data, security flaws, and allegations of IP theft. However, the AI industries in Asia and Africa are cheering DeepSeek not for what it is, but for what it stands for: Proof that frugality and innovation can go hand in hand.

“This is the next inflection point in AI after ChatGPT … where a product or a technology has shown how it can move the needle of AI even further,” Aakrit Vaish, founder of Haptik, one of the world’s largest conversational AI companies, and an adviser to the India government’s AI mission, told Rest of World.

DeepSeek’s feat of building an AI model that is faster and cheaper than others, which it detailed in a research paper, opens the possibility that “more teams can now take a crack at these things without having to do the base research themselves,” Vaish said.

The launch of DeepSeek’s latest model rattled the U.S. stock markets. On January 27, American chipmaker Nvidia’s stock plunged 17% to become the biggest single-day wipeout in U.S. history because the Chinese company said it had built its model using Nvidia’s less-advanced H800 chips, and only 2,788 GPUs, in comparison to the 10,000 GPUs used by OpenAI. A GPU or graphics processing unit is an electronic circuit that performs mathematical calculations at high speed.

“DeepSeek has just blown the roof off of this thing because they’re showing us that there’s not just one way to get it done right by introducing a competitive approach … that has now shifted the power dynamics,” Oswald Osaretin Guobadia, founder and CEO of policy advisory firm DigitA and a senior special assistant on digital transformation to Nigeria’s former president, Muhammadu Buhari, told Rest of World.

“This should motivate and help us [Nigerians] restructure our government and program,” Guobadia said. “Africa as a whole should be looking at a consolidated collaborative, coordinated effort. That’s what China is: one big ball of consolidation, collaboration, and coordination.”

DeepSeek is “an example of clever engineering,” Karim Beguir, co-founder and CEO of one of Africa’s most successful AI startups, InstaDeep, told Rest of World.

However, Beguir said Africa needs more GPUs for its startups to compete on a global level. “Africa has less than 0.1% of the world’s GPUs and that needs to change,” he said.

On January 27, threat intelligence firm Kela said it had seen several security flaws in DeepSeek’s model. Later in the week, OpenAI’s parent company Microsoft started probing if DeepSeek had stolen intellectual property from its AI rival. Amid soaring global popularity, Ireland and Italy blocked the app citing data privacy concerns.

While DeepSeek shared some details about its model in its research paper, “there’s no training / data processing code, and hardly any information about the data,” Stanford professor Percy Liang wrote in a post on X. 

“It’s also true that DeepSeek isn’t the best model out there. But it’s the closest thing to the best new approach for training AI models that will make building easy for the rest of the world,” InstaDeep’s Beguir said. “But I’m not certain they have done it on the quoted budget.”

Vaish agreed that there was a need to know more about DeepSeek’s models than what the company has so far shared.

“There are a lot of hidden pieces behind it, which nobody can tell given the country that they come from” and details like who the people behind it are “continue to be a bit of a black box,” he said. Despite the debate around DeepSeek publicizing deflated costs, Vaish said “it’s definitely not 100 or 1,000 [times] of what DeepSeek claims. … Worst case is what — 10, 15, 20, 30 [times]? In which case, it’s still not $1 billion.”

In 2024, AI and machine learning startups received a third of the total venture capital funding globally, according to research firm Pitchbook. However, there have been concerns that too much money is being pumped into foundational models and infrastructure, while not enough is being done toward the real-world application of AI.

For the industry to move at a faster pace, it is important that every company does not start from scratch, Vaish said.

“We never built the internet browser, we never built the cloud infrastructure that all of the world’s data runs on. So why suddenly go on this bandwagon and say let’s build the AI infrastructure?” Vaish said about India. “But in all these cases, whenever technology becomes commoditized and cheaper and available to use for everybody — that’s where India shines and that’s our moment today.”