DeepSeek’s Disruption: The Impact on Nvidia and the Semiconductor Industry

Please note that VanEck may have positions in the firms referenced herein.

DeepSeek’s recent claim of high-end results at a fraction of typical costs has rattled investors and raised questions about Nvidia’s dominance. However, it’s important to consider the broader context: hyperscalers and enterprises could now shift focus to leveraging existing AI infrastructure for inference, where Nvidia may still hold significant advantages with its software ecosystem and next-gen products. Meanwhile, the broader semiconductor market presents opportunities for diversification, with specialized ASICs and other chipmakers potentially capturing demand. However, investors should remain aware of key risks, including market volatility, disruptive technological advancements, and regulatory uncertainties that could affect industry dynamics.

The Breakthrough Claim: DeepSeek’s R1 Model

• DeepSeek asserts that its R1 model delivers competitive AI performance while reducing training costs significantly (Source: Reuters, January 2024).
• Unlike major AI labs that rely on cutting-edge GPUs, DeepSeek claims to have used lower-cost hardware while achieving strong results (Source: Bloomberg, January 2024).
• If such efficiency gains are scalable, they could alter hyperscalers’ (large cloud providers) future capital expenditures on AI infrastructure.
• A key industry question arises: If new models can be trained more cheaply, do hyperscalers still need massive GPU investments to achieve top performance?

The Hyperscalers’ Massive CapEx on AI Training

• Over the past two years, major cloud providers (AWS, Azure, Google Cloud, etc.) have invested heavily into data centers outfitted with Nvidia GPUs to handle AI training at scale.
• The open question has been how these hyperscalers plan to monetize that infrastructure.
• AI researchers’ commentary now suggests that performance gains from ever-bigger models may be reaching diminishing returns, nudging the market from investing in bigger clusters toward monetizing existing models.
• The GPU power dynamic had a potential shift from Nvidia’s ownership of the market to the hyperscalers’ increased purchasing power.
• However, the AI hardware market is subject to volatility, and changing industry dynamics could introduce regulatory and competitive risks.

Note: Each hyperscaler has different paths to monetization—through cloud subscriptions, consumer apps, enterprise services, etc.—which reduces the likelihood that open-source alone will displace them.

Training vs. Inference in the Semiconductor Industry

• The industry is moving from a training-dominated focus to the inference phase, where real-world applications and monetization happen.
• Tesla offers a perfect illustration:
  • They train their Full Self-Driving (FSD) models in big data centers.
  • Inference happens in each car (edge computing)—with the car running the model locally without constant round trips to the cloud.
  • This edge-based inference is monetized through subscription fees (e.g., $99/month for FSD)(Source: Tesla Q4 2023 Earnings Report).

DeepSeek’s Disruption: Concerns for Nvidia and Other Chipmakers

• DeepSeek’s announcement—training a high-performance model on cheaper hardware—caused market jitters about future GPU demand.
• Previous assumptions around scaling laws could be disproven with this model. R1 did more with less computing.
• The concern: if hyperscalers can achieve state-of-the-art results with less costly or alternative hardware, Nvidia’s growth in data center GPUs specifically could slow. This sentiment contributed to recent stock price drops across AI chipmakers.
• However, technological shifts can be unpredictable, and regulatory policies or competitive innovations may alter these dynamics.

If DeepSeek’s Claims Are Fully True…

• DeepSeek’s R1 shows that you can train top-tier models with cheaper hardware.
• This doesn’t eliminate Nvidia, but it potentially accelerates a timeline where Nvidia’s near-total dominance in AI training normalizes.
• As training hardware demand broadens, other chip designers (e.g., ASIC makers, Broadcom, Intel, AMD, specialized startups) can gain footholds—especially in inference, where purpose-built chips are cost- and power-efficient.
• Takeaway: Even if GPU training demand levels out, there is still a notable opportunity across the semiconductor ecosystem.

If DeepSeek’s Claims Are Overstated…

• Perhaps DeepSeek’s achievement isn’t as groundbreaking as it seems, or there are undisclosed constraints.
• Still, the long-term AI cycle naturally shifts to monetizing models via inference.
• That inference stage could favor a variety of hardware solutions, including specialized ASICs, smaller GPU instances, and enhanced CPUs. These alternatives could increase competition with Nvidia’s high-end GPUs by offering more efficient and cost-effective solutions.
• Takeaway: Nvidia may continue leading in training, but, as AI matures, more players will compete for different parts of the AI hardware stack.

Nvidia’s Continued Strengths and the Move to Inference

• Nvidia remains a highly innovative leader in AI hardware and software.
• They have announced new products designed for inference workloads—like next-gen GPU architectures (e.g., Hopper) and specialized platforms that bridge training and inference.
• As the market transitions, Nvidia’s data center GPU business could likely see more normalized growth—but their comprehensive ecosystem (hardware, CUDA software, enterprise partnerships) still positions them as a key player.
• In parallel, other chipmakers are ramping up, and a broader selection of ASICs and CPUs is emerging, enabling a variety of cost-effective inference solutions.

Why We Favor a Diversified Approach

1. Training vs. Inference: Industry focus is shifting from massive training (where Nvidia dominated) to inference (where more players will have competitive offerings).
2. Monetization: AI is moving into real-world deployments and subscriptions (e.g., Tesla’s FSD), highlighting the importance of efficient inference hardware and networks.
3. Nvidia’s Role: Nvidia is still well-positioned with leading GPU and software solutions, plus new products targeting inference. However, the days of unbounded data center GPU demand may be giving way to a more balanced, multi-vendor market.
4. Investment Strategy: In this environment, diversifying across the semiconductor sector can hedge against potential shifts in market leadership—from Nvidia’s GPUs to specialized ASICs and other hardware innovators.

Ultimately, the AI hardware landscape remains dynamic. While Nvidia is poised to remain a major force, Deepseek’s story—and the broader shift to inference—underscore the value of broad exposure to the entire semiconductor value chain.