To The Moon

What is DeepSeek?

• DeepSeek is an open-source AI model developed by a Chinese AI startup.

• It reportedly outperforms industry-leading models, such as OpenAI’s GPT-4, on several key benchmarks at a significantly lower cost. DeepSeek claims to have spent only $5.6 million to develop its model.

• However, Scale AI CEO Alexandr Wang mentioned that DeepSeek has access to around 50,000 NVIDIA H100 AI GPUs (each costing $27,000 to $40,000 per unit).

• Due to export restrictions, DeepSeek has avoided discussing its use of these GPUs, suggesting that their true costs may be understated.

The Impact of DeepSeek’s Success

• Cost Efficiency Concerns:

If a Chinese startup can develop an LLM (large language model) as good as OpenAI or Meta’s models with significantly fewer AI chips (lesser cost), why should companies buy so many NVIDIA GPUs?

• Investor Skepticism:

DeepSeek’s success has raised concerns about the massive AI spending by U.S. tech giants. Investors are questioning the ROI of these AI projects and speculating whether hyperscalers like Meta, Microsoft, and others might reduce their AI CapEx (capital expenditures), which could impact NVIDIA’s revenue growth.

Why You Shouldn’t Be Overly Worried?

1) Existing AI Players May Spend More to Compete

• $Meta Platforms, Inc.(META)$ ’s CapEx Plans: CEO Mark Zuckerberg recently revealed that Meta plans to invest $60 billion to $65 billion in AI infrastructure this year.

• If companies cannot quickly tweak their algorithms to compete with DeepSeek, they may need to increase spending on GPUs to boost computing power.

• Earnings to Watch:

  Keep an eye on AI-related CapEx announcements in upcoming earnings calls:

Jan 29: $Meta Platforms, Inc.(META)$ , $Microsoft(MSFT)$ , $Tesla Motors(TSLA)$

Jan 30: $Apple(AAPL)$

Feb 4: $Alphabet(GOOG)$

Feb 6: $Amazon.com(AMZN)$

2) DeepSeek May Be Overhyped

• Unlike major players, DeepSeek may have achieved rapid results by focusing exclusively on language-based AI, instead of diversifying efforts into multi-modal models (e.g., integrating text, image, and video).

• DeepSeek’s reliance on FP8 precision emphasizes efficiency but may sacrifice some accuracy compared to models using FP16 precision, which prioritize accuracy at a higher computational cost.

3)DeepSeek’s Success Could Benefit AI Chipmakers

Lower costs for training and inferencing could drive exponential AI adoption, ultimately benefiting companies like NVIDIA as demand for GPUs rises to power broader AI deployments.

Conclusion

• The hype surrounding DeepSeek is untimely, as it has momentarily overshadowed excitement around Project Stargate, a key AI initiative. Beneficiaries of Project Stargate, such as $ARM Holdings(ARM)$ , $Microsoft(MSFT)$ , $Oracle(ORCL)$ , $NVIDIA(NVDA)$ , and $Taiwan Semiconductor Manufacturing(TSM)$ , saw their stock prices drop in pre-market trading.

• Generative AI depends on two critical components:

1. Advanced algorithms

2. Massive computational power to process data, train models, and handle complex tasks.

• DeepSeek proves that scaling laws for AI have not yet plateaued, and better algorithms can achieve significant improvements. However, achieving Artificial General Intelligence (AGI) still requires vast amounts of computing power and data. NVIDIA’s GPUs remain central to this evolution.

• I reckon the sell-off triggered by DeepSeek’s news is likely to be short-lived.

• I remain constructive on AI-related stocks and ETFs, such as the Magnificent Seven (Meta, NVIDIA $NVIDIA(NVDA)$ , Apple, Tesla, Alphabet, Amazon, and Microsoft), SMH ETF $VanEck Semiconductor ETF(SMH)$ , TSMC $Taiwan Semiconductor Manufacturing(TSM)$ , Broadcom, ARM $ARM Holdings(ARM)$ , and Oracle $Oracle(ORCL)$ .