Nvidia’s Moat is Getting Attacked from Multiple Sides
Author: Ruben Dalfovo, Investment Strategist
Date: August 20, 2026
Key Takeaways
- Cerebras is challenging Nvidia on inference speed.
- Google is deepening its custom chip initiatives with Marvell.
- Hyperscalers prefer chips tailored for specific workloads over all-purpose processors.
- Nvidia retains advantages in software and networking, but the AI chip market is becoming more specialized.
Market Dynamics
Nvidia has dominated the AI chip market by providing versatile GPUs that serve as the industry's go-to solution. However, as the demand for AI applications grows, major customers are increasingly seeking chips designed specifically for their unique workloads. This shift is evidenced by two significant developments:
- Cerebras' New Hardware: On August 18, Cerebras introduced hardware aimed at significantly enhancing AI inference speed.
- Google's Partnership with Marvell: Following this, Google announced an expanded collaboration with Marvell Technology, which could generate up to $120 billion in revenue for Marvell by fiscal 2033, contingent on meeting performance benchmarks.
The Evolving Landscape of AI Chips
The AI chip market is diversifying. While Nvidia's GPUs are versatile and powerful, they are now one option among many. Companies like Google and Amazon are developing custom processors, such as Google's Tensor Processing Unit (TPU) and Amazon's Trainium, which are optimized for specific tasks, trading off some flexibility for enhanced performance and efficiency.
Cerebras' Innovative Approach
Cerebras is pushing the boundaries of chip architecture by creating processors that are significantly larger than traditional chips, designed to minimize data movement and improve efficiency. Their new CS-4 system reportedly uses 50% fewer components than previous models, specifically targeting inference tasks where speed is crucial.
Google's Strategic Moves
Google's strategy reflects a broader trend among hyperscalers who are leveraging their scale to design chips tailored to their specific needs. The expanded agreement with Marvell encompasses not just processors but also storage and networking solutions, indicating a shift towards more customized hardware solutions.
Challenges for Specialization
Despite the advantages of specialized chips, there are significant challenges. The design and manufacturing of custom chips can be costly and complex, and software compatibility is crucial for successful deployment. Nvidia's ability to enhance inference performance while keeping customers within its ecosystem remains a key defensive strategy.
Investor Considerations
Investors should focus on the actual deployment of custom chips rather than just their announcements. It's essential to differentiate between training and inference capabilities, as Nvidia may still excel in training even as alternatives gain traction in inference. Additionally, the surrounding infrastructure, including networking and memory, will play a critical role as data centers adopt a mix of processors.
Conclusion
The competitive landscape for AI chips is evolving, with new players and specialized solutions emerging. While Nvidia's moat is under pressure, it remains formidable due to its established ecosystem and scale. The key question is not whether faster chips can be built, but whether these alternatives can match Nvidia's comprehensive advantages in performance, software, and deployment ease.