DeepSeek revelation likely adds, not detracts to AI compute boom
Jan. 27, 2025 9:16 AM ET
By: Brandon Evans, SA News Editor
 blackdovfx
The advent of the DeepSeek V3 large language model has prompted a mild panic among investors, causing stocks related to powering artificial intelligence to plunge as the trading week prepares to kick off.
Nvidia (NASDAQ: NVDA) and Broadcom (NASDAQ: AVGO) dove 11%, Advanced Micro Devices (NASDAQ: AMD) dipped 4% and ASML Holding (NASDAQ: ASML) declined 8%, just to name a few.
However, Cantor analysts surmised that the advent of a China-based LLM that requires less computing power, will likely create a boom for the developers of high-end graphic processing units and data center builders.
"Following release of DeepSeek's V3 LLM, there has been great angst as to the impact for compute demand, and therefore, fears of peak spending on GPUs," said Cantor analysts, led by C.J. Muse, in a Monday investor note. "We think this view is farthest from the truth and that the announcement is actually very bullish, with AGI seemingly closer to reality and Jevons Paradox almost certainly leading to the AI industry wanting more compute, not less."
Jevons Paradox is a 19th century economical idea that finds increased efficiency actually leads to increased use.
DeepSeek was founded in 2023 by the Chinese billionaire and hedge fund creator Liang Wenfeng. He reportedly stockpiled a supply of as many as 10,000 to 50,000 Nvidia GPUs, before the U.S. AI export chip ban went into effect, along with lower-end chips to build his LLM.
"DeepSeek's success highlights the great efficiency of open-source models," Muse said. "Second, we do question the veracity of this model only costing $5.57M vs. LLAMA (Meta) 3.1 $500M. We have seen some reports that suggest they are using 50k Hopper GPUs vs. stated 10k A100s."
"At the end of the day, DeepSeek's advances mean we are closer to AGI," Muse added. "Work will continue on pre-training, post-training, and time-based inference/reasoning, and future investments in large-scale clusters will only accelerate. All of which is bullish for AI ... We see this progress as positive in the need for more and more compute over time (not less)."
The analysts as UBS drew a similar conclusion, finding compute demand for AI will not fade on the reports of DeepSeek's efficiency gains in building an LLM.
"There has been some scrutiny surrounding the claim that R1 was trained using H800 GPUs (some have suggested that DeepSeek has access to 50,000 H100 GPUs) though this still does not detract from R1's efficiency in inference, where cost per token is >95% less than o1," said UBS analysts, led by Timothy Arcuri, in a Monday investor note. "If anything, we think model developers will look to incorporate some of R1's novel techniques into their own models which would help to improve efficiency. Even though this may read negatively for compute demand, the fact remains that compute continues to drive model performance even as models become more efficient."
(Update: Following an influx of new users, DeepSeek said it was restricting registration for its AI assistant to Chinese phone numbers only.) |
