What’s Hot: Halftime Heroes or Final Whistle Winners? Lessons from DeepSeek’s AI Breakthrough for Thematic Investors

Originally Posted, 29 January 2025 – What’s Hot: Halftime Heroes or Final Whistle Winners? Lessons from DeepSeek’s AI Breakthrough for Thematic Investors

Key Takeaways

• The emergence of DeepSeek’s cost-effective AI models challenges established industry leaders, highlighting how sensitive to disruption thematic investment can be
• Relying heavily on current market leaders can be very risky in thematics. Historically, early leaders like AltaVista or Netscape did not end up creating value for investors

Over the weekend, DeepSeek, a Chinese artificial intelligence (AI) lab, made headlines by unveiling large language models (LLMs) that rival those of OpenAI and Meta, but at a fraction of the training cost. This breakthrough has sent shockwaves through the AI World, leading to a significant decline in the Nasdaq index. DeepSeek’s advancements challenge existing assumptions about the AI industry, reshaping both the competitive and investment landscapes.

This development serves as a wake-up call for AI and thematic investors alike. Regardless of their size, companies leading in a particular theme can be challenged over time. In thematic investing, disruption isn’t the exception—it’s the rule.

Is DeepSeek’s arrival, the turnaround that challenges Nvidia’s early Lead?

DeepSeek’s milestone involves developing AI models comparable to leading U.S. counterparts but at substantially lower costs. Their R1 model was trained using 2,048 GPUs over 57 days, consuming 2.78 million GPU hours on Nvidia H800 chips—a modest figure for a 671-billion-parameter model, costing approximately $5 million in hardware¹. In contrast, an earlier Open AI model like GPT-4, released in March 2023, required about 25,000 NVIDIA A100 GPUs over 90 to 100 days, costing around $63 million².

The market responded swiftly to DeepSeek’s advancements. Nvidia, a dominant player in AI hardware, experienced a notable stock decline, reflecting investor concerns about potential shifts in AI infrastructure demand. Prior to this development, it was believed that AI-driven gains were primarily captured by companies with substantial tech investments and high-end chip producers. DeepSeek’s breakthrough suggests that:

• Companies with significantly smaller investment resources than U.S. mega-caps like Microsoft or Meta can still question the leadership and disrupt the space
• Training of some AI models can be accomplished with more cost-efficient alternatives than NVIDIA chips, highlighting opportunities in the semiconductor market.
• Lower cost of inference, which means more companies can integrate AI capabilities into their products with less costs than previously assumed, e.g. presenting a tailwind for software companies leveraging AI agents.

From an investment standpoint, strategies that prematurely concentrated investments in Nvidia and hyperscalers may find that dominating the first half does not always imply winning the game.

Beyond the Early Lead

Thematic investments seek to capitalize on the exceptional growth driven by structural changes in the world. Given that equity markets are often governed by the “winner takes it all” spirit, only a handful of companies typically capture the majority of these gains.

In the early days of the internet, companies like Yahoo, AltaVista, and Netscape seemed to have insurmountable leads. However, investors who bet solely on these firms were ultimately disappointed as they were challenged and overtaken by competitors such as Amazon, Google, and Microsoft. This underscores the importance of not drawing premature conclusions about market leaders.

WisdomTree’s Thematic Approach: Expertise and Differentiation to Benefit from and build resilience against Disruption

At WisdomTree, we have developed a unique Thematic Investment Philosophy anchored around four key pillars to address these challenges:

1. Expertise: Our investment process leverages domain expertise, often through partnerships with external specialists, to identify companies that offer focused exposure to a theme and stand to benefit most from its evolution.

2. Purity of Exposure: We concentrate on pure-play companies within a theme’s ecosystem, reducing exposure to companies that only marginally engage with the theme.

3. Differentiation: Our portfolios are constructed to add distinct exposure, minimizing overlap with traditional equity benchmarks and offering distinct geographical or sectorial insights.

4. Transparency, Investability & Discipline: We build scalable and liquid portfolios with a transparent investment process, allowing investors to monitor and manage their investments effectively. This approach combines our expert partners’ proprietary knowledge with disciplined index rules to avoid the pitfalls of “hype investing.” The rigidity of the index rules ensures that the portfolio remains focused on the theme and that the manager does not end up investing in the fads of the moment.

By adhering to these principles, we create diversified portfolios that invest across the full breadth of the thematic opportunity, from established leaders to emerging disruptors. This approach allows our thematic strategies to benefit from the growth of current leaders while remaining resilient to potential disruptions.

The weighting methodology is also crucial. For themes early in the adoption curve, an equal-weighted approach acknowledges that no leaders have yet emerged. As themes mature, a modified market cap weighting—where weights depend on company size and relevance to the theme—ensures that early leaders receive significant weightings while accommodating potential disruptors. Only themes late in the adoption curve benefit from a market cap-weighted approach, as leaders are firmly established, and disruption is less likely.

Conclusion

DeepSeek’s recent advancements serve as a reminder of the importance of a disciplined, diversified approach.

¹ https://www.unite.ai/how-deepseek-cracked-the-cost-barrier-with-5-6m/?utm_source=chatgpt.com
² https://plainswipe.com/gpt-4-details-leaked/index.html