And then the AI bubble burst. Or, as one instant meme put it, “ChatGPT just lost its job to AI”.
Instead of the previous few weeks’ obsession with TikTok’s 24-hour banning in the US, the world is in thrall to another Chinese app, DeepSeek. This AI start-up released its latest chatbot model last week, with many capabilities that match OpenAI’s ChatGPT — but developed at a fraction of the cost.
Instead of the hundreds of millions of dollars the US firms have invested in training advanced AI models, DeepSeek said it only spent $5.6m developing its R1 model. This is estimated at 3%-5% of what ChatGPT cost to develop.
The bubble didn’t just burst, it blew up. Since ChatGPT burst into the world on the last day of November 2022 there has been an arms race by US firms to build bigger data centres to run ever bigger large-language models (LLMs) that need more and more computing power. Or, as the industry itself calls it in flagrant disregard of grammar, “compute”.
To say the Nasdaq fell off the cliff is an understatement, as was the speed of the fall from grace of the tech stocks primed to take advantage of this AI hype cycle. On Monday $1-trillion, or 3.1% of its total market capitalisation, was wiped off the Nasdaq. Chipmaker Nvidia, which briefly eclipsed Apple and Microsoft last year as the most valuable listed company, lost $600bn alone. This is a dubious new record in its own right, the biggest one-day loss in US history. Ouch.
Google’s owner Alphabet lost $100bn, while Microsoft, which owns a large stake in OpenAI, lost $7bn. SoftBank — part of US President Donald Trump’s big announcement about building AI data centre capacity with OpenAI and Oracle, called Stargate — lost 8% on Monday and 5% on Tuesday. Eina on top of ouch.
Silicon Valley venture capitalist Marc Andreessen, who founded the Netscape browser, called this AI’s “Sputnik moment” — when the Soviet Union beat the better-resourced US into space with its Sputnik satellite.
Indeed, it certainly seems as if the AI bubble has burst. The sky-high projections for tech firms has been undermined by a little Chinese quantitative analysis firm called High-Flyer, that decided to build its own chatbot. Instead of creating its own large-language model, an enormously expensive undertaking, DeepSeek used the output of another LLM to build its R1 model, a process known as distillation.
Last week’s release was akin to that of OpenAI’s ChatGPT two years ago, when like DeepSeek this week it became the most downloaded app on Apple’s app store. DeepSeek’s release proved that it’s possible to create advanced so-called chatbot models using a fraction of the resources the big US AI firms have spent.
Companies such as OpenAI, Google, Microsoft and Facebook have invested tens of billions of dollars in high-end servers and powerful chips such as Nvidia’s H100, which is specifically designed for AI use. But DeepSeek, which published a paper on how it created R1, says it used Nvidia’s far less powerful H800 processor, which the chipmaker created because of restrictions on what US technology could be exported to China. These H800 chips were also banned last October.
“I don’t believe it’s $6m, but even if it’s $60m it’s a game-changer,” Umesh Padval, MD of Thomvest Ventures, which invests in AI firms, told Wired. “It will put pressure on the profitability of companies which are focused on consumer AI.”
As if to confirm this, this week DeepSeek released its own image-generating service called Janus-Pro, which already matches the output of OpenAI’s market-leading DALL-E 3.
Last year Goldman Sachs cautioned that the estimated $1-trillion in capex for “data centres, chips, other AI infrastructure, and the power grid,” which has “little to show for it so far,” might not “ever pay off”.
This week OpenAI CEO Sam Altman was full of praise. “DeepSeek’s R1 is an impressive model, particularly around what they’re able to deliver for the price,” he tweeted. “We will obviously deliver much better models and also it’s legit invigorating to have a new competitor. We will pull up some releases.”
Just last week Altman stood next to Trump as he announced the launch of Stargate, a project to build AI infrastructure (which includes OpenAI, Oracle and SoftBank) that will invest $500bn. This week Trump called DeepSeek a “wake-up call” and pointed out that it is “AI from a Chinese company”. Talk about focusing on the wrong Chinese app. TikTok owner ByteDance must be silently relieved there is a new baddy in town.
“The race is on,” said Bank Zero chair Michael Jordaan, himself a tech-savvy start-up investor since he resigned as FNB CEO a decade ago. “What DeepSeek demonstrated is that AI models can be built with far lower training cost on much less advanced and expensive chips. This also requires far less energy,” he told Business Day.
As he quite rightly points out, “the DeepSeek method is now open sourced, so it is no longer just Chinese. Everyone can use their insights for free if they want to. So, the US AI industry will catch up fast”.
However, the most cogent thing said about AI was by Scott Hanselman, Microsoft vice-president of developer community, at an AI event in Sandton last week: “We have to push back on the idea that AI is smarter than us. It’s artificial, but it’s not real intelligence. We are the real intelligence. Is the spanner the thing we hired, or is it the plumber that we hired?”
• Shapshak is editor-in-chief of Stuff.co.za
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.