A new contender has entered the AI ring. For years, since the time AI broke loose and started to get out of control, the perception was that only the US had the means and talent to make and lead the way. Remember how OpenAI Chief, Sam Altman, had written off every Indian startup who was trying to compete with it, calling it a “hopeless” endeavor, on a visit to India, no less. His comment reeked of arrogance, coated with confidence, something that might have gone unnoticed back then—this was in 2023—due to lack of any formidable competition, but someone, somewhere in China was listening closely and took it very seriously.
Spun from a hedge fund based out of Hangzhou, DeepSeek has single-handedly done for AI, what the— Sputnik—Soviet satellite launch did for space and geopolitics. There are different angles to this proverbial “underdog” story. Not only has DeepSeek managed to outpace and outsmart Big Tech, but it has also been able to do it at considerably less resources in the face of US sanctions that bar tech companies from the US to do business with the Chinese, in this case, powerful Nividia GPUs.
The tech world is abuzz with how DeepSeek was able to build a powerful AI model—LLM—with significantly less computing power than industry giants like OpenAI and Anthropic. DeepSeek’s DeepSeek-V3 model was trained on a budget of less than $6 million using Nvidia H800 chips, as per the Chinese company, but while it may have rendered it ineffective in edge cases, data suggests, it was able to run at par—and even excel in some areas—on benchmarks and capability with these giants. It has started a whole new conversation around if they are spending more when in fact, they could do more for less.
Mark Zuckerberg impressed, Elon Musk skeptical
Meta CEO Mark Zuckerberg has expressed admiration for DeepSeek’s achievement, during a recent earnings call, stating that the company’s ability to achieve such results with limited resources has “strengthened our conviction that this is the right thing to be focused on.” Zuckerberg also revealed that Meta plans to incorporate DeepSeek’s advancements into its own Llama model.
However, Tesla and SpaceX CEO Elon Musk remains unconvinced. He has expressed skepticism about the company’s claims regarding GPU usage hinting that DeepSeek isn’t being transparent enough with the numbers or even how it managed to train its model. Musk and Scale AI CEO Alexandr Wang believe that DeepSeek is likely using more GPUs than it publicly admits, possibly due to US export restrictions on advanced chips. OpenAI and Microsoft are investigating reports that DeepSeek might have used ChatGPT to train its model.
DeepSeek’s rise has raised questions about the future of AI development and the role of computing power. While OpenAI CEO Sam Altman has acknowledged the “impressiveness” of DeepSeek’s model, he maintains that greater computing power remains crucial for achieving breakthroughs in AI for the kind of things he is building on way to AGI or Artificial General Intelligence.
Zuckerberg, on the other hand, believes that DeepSeek’s approach validates Meta’s strategy of investing heavily in computing infrastructure. He argues that while DeepSeek’s model may be efficient, Meta’s focus on “reasoning” and multimodal capabilities will ultimately lead to more advanced AI assistants. Meta plans to spend roughly $60 billion on AI this year.
Future of AI
The debate surrounding DeepSeek highlights a fundamental question in AI development: is efficiency more important than raw computing power? DeepSeek’s success suggests that it may be possible to achieve impressive results with less expensive hardware, potentially democratising access to AI technology.
However, skeptics like Musk argue that DeepSeek’s claims should be taken with a grain of salt and that true AI breakthroughs still require significant computing resources. The coming months and years will likely reveal whether DeepSeek’s approach represents a paradigm shift in AI development or merely a temporary disruption.
There is no doubt that DeepSeek’s entry has injected new energy into the AI landscape, sparking a debate about the future of AI development and the role of computing power. India, too, has joined the bandwagon announcing an ambitious roadmap to build its own bias-free, homegrown ChatGPT rival with a massive 18,000 GPU compute facility within the next ten months.
Follow FE Tech Bytes on Twitter, Instagram, LinkedIn, Facebook