Story Stocks^®

Artificial intelligence stocks are taking on water today following the release of a groundbreaking AI model from a Chinese startup, DeepSeek. Semiconductors are among some of today's biggest losers, with leaders in the sector, including NVIDIA (NVDA -14%), Broadcom (AVGO -15%), and Taiwan Semi (TSM -13%), down by a double-digit percentage. The news has rocked big tech as well, placing selling pressure on Microsoft (MSFT -3%), Amazon (AMZN -1%), and Alphabet (GOOG -2%), all of which are deeply invested in AI, continuing to spend on the technology to pounce on unwavering demand. Other industries with significant ties to AI are also selling off, including nuclear stocks like Constellation Energy (CEG -19%), Nuscale Power (SMR -24%), and Vistra (VST -26%), which have seen substantial investments from tech firms due to the power demands of AI workloads.

• What is DeepSeek? The China-based AI platform is similar to the AI platforms offered by OpenAI (MSFT), such as ChatGPT, allowing it to generate content, debug code, and execute multi-step workflows. The latest release of DeepSeek, dubbed DeepSeek-R1, touts performance on par with OpenAI-o1, the latest series of reasoning models from OpenAI. This in itself is not so much the part that is causing some panic today. Instead, it is the cost at which DeepSeek achieved this milestone that has Wall Street concerned.
• According to DeepSeek's research paper, it cost the company just over $5.5 mln to train its latest model, far less than the estimated $80 mln and $190 mln OpenAI and Google spent to train some of their more recent AI models, respectively. This is a serious issue for big tech, which is spending billions on its AI infrastructure. If DeepSeek can do the same or better for considerably less money, it can charge far less than U.S.-based competitors. Furthermore, DeepSeek-R1 is open-source, making it even cheaper to train and leverage, threatening big tech's leadership in AI.
• While DeepSeek has purchased NVDA chips, noting that the H800 was used for DeepSeek-V3, released in December, these chips were sold at a fraction of the cost of NVDA's flagship H100 chips. By demonstrating that it can extract similar gains using throttled chips, which were re-tooled to comply with U.S. trade restrictions, DeepSeek is injecting further worry that big tech could reduce their spending on NVDA hardware, hurting the future financials of the company and its supplier, TSM.

DeepSeek's R1 performance metrics are sparking panic across tech companies today, particularly those with heavy exposure to AI. However, there is the possibility that DeepSeek's numbers may not be accurate. Perhaps the company spent significantly more than it claimed. Still, in a market where tech companies like NVDA have stretched valuations, trading at a rich multiple of around 32x as of Friday's close, investors are taking profits now instead of waiting until the dust settles.