Deep in the heart of China’s bustling tech hub, a fledgling company named DeepSeek has embarked on a mission that could potentially reshape the technological balance of power. This startup audaciously claims its new AI models rival, or in some instances, even surpass the giants of American tech, all while dramatically slashing costs.
In the labyrinthine corridors of artificial intelligence, the buzz is palpable. Just last month, DeepSeek unlocked a wave of curiosity with a paper claiming the training of their latest marvel, the DeepSeek-V3, cost a modest sub-$6 million using Nvidia’s H800 chips. It’s an assertion that has sent AI aficionados worldwide scrambling for their calculators.
With the unveiling of this model, DeepSeek’s AI Assistant has swept past its American competitor, ChatGPT, firmly planting itself on the zenith of Apple’s App Store chart stateside. This insurgence isn’t merely rattling the tech landscape; it’s stirring questions. Why are U.S. technocrats channeling vast fortunes into similar ventures?
Shares of titanic tech outfits, Nvidia included, are feeling this tremor. Across the digital sphere, whispered conversations turn to speculation about the innovation surge heralded by DeepSeek. Let’s unpack this seismic shift initiated by the trailblazing Chinese enterprise.
WHAT MAKES DEEPSEEK A GAME-CHANGER?
The debut of ChatGPT by OpenAI in late 2022 triggered a frenetic surge among Chinese tech conglomerates. They hustled to crank out their interpretations of these brainy, AI-driven interlocutors. Yet, reality delivered a harsh verdict. Despite Baidu’s notable attempt with its ChatGPT clone, the discrepancy in capabilities between U.S. and Chinese models was glaring.
Enter DeepSeek, stage left. This young company managed to defy expectations, turning the tables on this narrative. According to its proclamations, models such as DeepSeek-V3 and its sibling, the DeepSeek-R1, are not merely comparable but notoriously efficient in cost. Silicon Valley’s tech elite and U.S. engineers have noted their prowess, aligning them closely with the flagship models of OpenAI and Meta.
Let’s talk numbers. The economic savvy of DeepSeek-R1 is downright impressive. Newly unveiled last week, this model promises a cost-efficiency of 20 to 50 times that of OpenAI’s offerings, varying by application, as disclosed on DeepSeek’s WeChat portal.
Yet, some skeptics voice their incredulity. One such naysayer, Scale AI CEO Alexandr Wang, disputed the legitimacy of DeepSeek’s narrative in a chat with CNBC. Without presenting evidence, Wang claimed the Chinese firm possesses a staggering 50,000 Nvidia H100 chips. He conjectured that DeepSeek’s withholding of this information flouted U.S. export restrictions, disallowing such advanced components to Chinese firms. DeepSeek has yet to comment on these allegations.
Adding fuel to the skepticism, Bernstein analysts opined in a recent report that actual training expenses for DeepSeek’s V3 are cloaked in enigma, exceeding the meager $5.58 million suggested for computing costs. The training expenditure for the equally lauded R1 model remains undisclosed too.
WHOSE DREAM IS DEEPSEEK?
Situated amidst the innovative energy of Hangzhou, DeepSeek’s vision is helmed by Liang Wenfeng. A notable figure, Liang co-initiated the quantitative hedge fund High-Flyer, as documented in Chinese company reports. Earlier this year, Liang’s fund revealed its strategic pivot: transcending trading to invest in forging an independent research collective. This initiative, aimed at unraveling the quintessence of Artificial General Intelligence (AGI), charted a new era, birthing DeepSeek.
OpenAI, the creators of ChatGPT, frames AGI as autonomous systems surpassing human capabilities across financially critical tasks. The financials, however, linking High-Flyer’s involvement with DeepSeek remain an enigma. The two cohabit a building, and High-Flyer holds patents pertinent to chip clusters conducive to AI model training, as corporate documentation illustrates.
A revelation on High-Flyer’s WeChat in mid-2022 announced its operation of an impressive assembly of 10,000 A100 chips.
DEEPSEEK’S PLACE IN CHINA’S POLITICAL ARENA
DeepSeek’s meteoric rise hasn’t escaped the gaze of Beijing’s politico-economic stalwarts. On the very day DeepSeek-R1 graced the public arena, January 20, its founder Liang Wenfeng mingled at a secluded symposium convened by Premier Li Qiang. The state-controlled news outlet, Xinhua, chronicled Liang’s presence at this exclusive conclave.
Liang’s attendance hints at greater strategic significance. As China wrestles for independence from U.S. export policies, DeepSeek’s ascendancy could be pivotal, serving Beijing’s ambitions for self-reliance in crucial domains like AI.
This symposium bore resemblance to a previous gathering, which too witnessed the participation of high-profile entrepreneurs such as Baidu’s CEO, Robin Li.
Report By Axadle Desktop.
