This is not investment advice. The author has no position in any of the stocks mentioned. Wccftech.com has a disclosure and ethics policy.
Despite mega cap US firms allocating billions of dollars in capital expenditure spending for AI GPUs, Chinese multinational technology firm Tencent's chief strategy officer James Mitchell believes that DeepSeek's AI breakthroughs demonstrate that such spending might not be needed. DeepSeek's claims of having developed its AI models similar in capabilities to some American counterparts but at a fraction of a cost appear to have permanently altered the stock market.
Their biggest victim was NVIDIA's stock, which has yet to recover nearly $600 billion in losses during January's selloff. NVIDIA's GTC conference earlier this week failed to impress investors either, and the stock has remained flat despite CEO Jensen Huang pointing out multiple trillion dollar markets for his firm's products.
Chinese Technology Companies Reduce GPU AI Capital Expenditure Purportedly Due To DeepSeek's Breakthroughs
In a recent talk, Tencent CSO James Mitchell shared that one use case his firm had for spending money on NVIDIA's AI GPUs was to train large language models or LLMs. Tencent released its Hunyuan Turbo S AI model a little over a month after DeepSeek's products gained popularity in January. The firm aimed at the Chinese startup with its products as it claimed that its model could respond to queries in less than a second.
However, while competing with DeepSeek to develop the fastest AI models, Tencent also believes that the startup's training breakthroughs significantly reduce AI development costs. DeepSeek claims to have reduced AI development costs by using advanced software engineering, which allowed it to access a GPU's core functions. Typically, engineers rely on NVIDIA's CUDA software to use the chips, but the convenience comes at the expense of tighter control over the products.
Commenting on capital spending for AI GPUs, Mitchell shared that "A second use of CapEx was GPUs for large language model training," This use case was important before DeepSeek shared its model development process. The Tencent CSO added that "there was a period of time last year when there was a belief that every new generation of large language model required an order of magnitude more GPUs." However, according to him, DeepSeek changed this perception, at least among Chinese firms. "That period of time ended with the breakthroughs that DeepSeek demonstrated," Mitchell said.
While he didn't share how, Mitchell outlined that after the breakthroughs, "the industry and we within the industry are getting much higher productivity on a large language model training from existing GPUs without needing to add additional GPUs at the pace previously expected." Chinese firms are restricted from buying NVIDIA's latest AI GPUs, such as its Blackwell and Hopper products. The restrictions have forced them to rely on older chips or large GPU clusters to overcome the limitations of limited computing power.
Tencent has claimed that its Turbo S model has beaten DeepSeek's products in math, reasoning and other AI capabilities. Industry sources believe Chinese firms are interested in working with Huawei and its Ascend AI chips as they navigate the chip blockade.
Like NVIDIA, Huawei also offers its chip customers with software to control the chips. However, purportedly, DeepSeek has tested the software and found it lagging to NVIDIA's products. NVIDIA's shares, on the other hand, are yet to recover their pre-January selloff highs. The stock is down by 14% year-to-date as investors wait for more data for its demand to materialize. Mitchell's firm trades on the OTC markets and has a market cap of $601 billion.
James Mitchell of Tencent quite clearly laying out the bear case for $NVDA - DeepSeek's innovations mean fewer GPUs needed vs. what Tencent previously thought, despite the explosion in AI usage. @doodlestein
"There was a period of time last year when there was a belief that… pic.twitter.com/f80UXyQ8Ny
— Timothy Liu (@timothyhliu5) March 20, 2025
