hankyoreh

Logos for DeepSeek, Open AI and ChatGPT. (AFP/Yonhap)

The news that an affordable artificial intelligence by DeepSeek, a Chinese AI startup, has outperformed ChatGPT developer OpenAI on several benchmarks has upended the global AI industry.

Tech analysts in the US are closely watching the emergence of a new contender in the field while noting that DeepSeek’s achievements epitomize the normal tendency for costs to fall with technological advancement.

It has generally been assumed that AI development was a cash-intensive endeavor that required investments numbering in the tens of billions of dollars. But now it seems the industry has reached an inflection point.

Dario Amodei, the co-founder and CEO of Anthropic, argued that DeepSeek, which exhibited an AI model on a level similar to those of Big Tech firms a step behind its competitors, can’t be viewed as true innovation. In a blog post published on Wednesday, Amodel wrote: “Sonnet’s training was conducted 9-12 months ago, and DeepSeek’s model was trained in November/December, while Sonnet remains notably ahead in many internal and external evals.” 

Amodei is referring to Claude 3.5 Sonnet, a mid-sized model from Anthropic. 

Amodei argued that according to historical trends, DeepSeek is actually in alignment with a normal cost-curve reduction, which is normally a reduction by a factor of four in one year. Amodei argues that DeepSeek is “worse than US frontier models” by a factor of two but that its training cost was reduced by a factor of eight compared to US models developed a year ago. This would equate to cost-curve reduction by a factor of four, which, as Amodei points out, is “on-trend at best and probably not even that.” 

The training cost for the V3 iteration of DeepSeek, a large-language model (LLM) publicly revealed last month, was reportedly only US$5.57 million, a figure that made waves. Amodei explains, however, that Anthropic’s model, which came out six months prior, cost US$10 million to train. Considering the six-month difference, the reduction in initial cost is a completely expected outcome. 

Big Tech firms create war rooms to respond

Big tech firms, while struggling to maintain their poker faces amid the emergence of such unexpected competition, have been spurred into action after being shaken out of complacency by DeepSeek. Sam Altman, the CEO of OpenAI, posted on X (formerly Twitter) on Tuesday: “deepseek’s r1 is an impressive model, particularly around what they’re able to deliver for the price.”

Satya Nadella, the CEO of Microsoft, posted: “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

Despite their praise, both OpenAI and Microsoft are raising suspicions about DeepSeek developers collecting OpenAI data without authorization, and have set out to investigate. Reuters reported that American AI experts view DeepSeek as a “distillation,” which is essentially when “one model learns from another model.” Through distillation, these American experts argue, DeepSeek copied OpenAI’s model while extracting its data without authorization to develop its own model.  

Meta, the parent company of Facebook, which has offered an open-source AI model similar to DeepSeek, reportedly set up a “war room” to draft a strategy for responding to the competitive threat posed by DeepSeek. Yet during an earnings call on Wednesday, Meta CEO Mark Zuckerberg was reportedly untroubled by the emergence of DeepSeek, despite others in the industry voicing concerns.

Impact on AI semiconductor stocks

Some people expect the Trump administration to implement export controls on semiconductors to prevent Chinese advances in AI technology. In addition to export controls on high-spec semiconductors, there are even considerations of restricting exports of low-spec semiconductors, such as Nvidia’s H20 chip, to China. DeepSeek has revealed that it utilized around 2,000 H800 Nvidia chips in its development of DeepSeek V3, which was released last month. 

Amid reports that DeepSeek utilized low-spec semiconductors to develop AI technology on the same level as Big Tech firms, the stocks of firms that manufacture AI semiconductors, such as Nvidia and Broadcom, plummeted on Monday and have been volatile ever since. Shares of Nvidia, which controls around 70% of the AI chip market, made a recovery of 8.93% on Jan. 28 but fell 4.1% on the following day. 

By Sun Dam-eun, staff reporter; Jeong Nam-ku, staff reporter

Please direct questions or comments to [english@hani.co.kr]