The Shifting Sands of Data Center Silicon: CPUs Develop a Comeback
For the last three years, the data center world has been almost entirely focused on Graphics Processing Units (GPUs). From training massive AI models to powering inference workloads, GPUs have dominated the conversation. But a significant shift is underway. The industry is beginning to recognize that CPUs aren’t going anywhere, and in fact, are poised for a resurgence, particularly as the landscape of artificial intelligence evolves.
Beyond GPUs: The Rise of Specialized Processors
The initial rush to GPUs was driven by their parallel processing capabilities, ideal for the matrix multiplications at the heart of deep learning. However, as AI models become more complex and diverse, the need for specialized silicon is growing. This isn’t just about GPUs anymore. Companies are realizing that a heterogeneous computing approach – utilizing a mix of CPUs, GPUs, and other accelerators – is the most efficient path forward.
Meta’s recent unveiling of its MTIA chip family is a prime example. The company is developing four custom, in-house chips – MTIA 300, MTIA 400, MTIA 450, and MTIA 500 – to diversify its silicon supply and reduce reliance on external vendors like Nvidia and AMD. The MTIA 300 is specifically designed for training smaller AI models used in ranking and recommendation tasks, whereas the upcoming chips will tackle more advanced generative AI inference. This demonstrates a move towards tailoring silicon to specific workloads, and CPUs are a crucial part of that equation.
Nvidia’s Expanding Ecosystem and the Vera Rubin Platform
Even Nvidia, the undisputed leader in GPU technology, acknowledges the need for a broader silicon stack. Their Vera Rubin platform, announced at GTC 2026, isn’t just about GPUs. It encompasses CPUs, dedicated inference accelerators, networking ASICs, data processing units, and Ethernet switches – all designed to work together as a cohesive system. This “AI factory” approach signifies a fundamental shift in how Nvidia packages and sells its hardware, moving beyond individual components to integrated solutions.
The Vera Rubin platform delivers 60 exaflops of compute, showcasing the power of co-designed systems. This highlights that while GPUs remain vital, they are most effective when integrated with other specialized processors, including CPUs.
The Role of Arm and Agentic AI
Arm’s recent decision to build its first-ever CPU signals a significant change in the industry. This move is directly tied to the growing demand for agentic AI – AI systems capable of autonomous action and decision-making. Agentic AI requires a different kind of processing power than traditional deep learning, one that emphasizes control, flexibility, and real-time responsiveness. CPUs, with their ability to handle a wider range of tasks and manage complex workflows, are well-suited for this role.
Power Considerations and Data Center Infrastructure
The increasing demand for processing power is also driving innovation in data center infrastructure. Nvidia is partnering with companies like Eaton, Schneider Electric, and Vertiv to address power delivery challenges, from facility-level grid connections to rack-level power management. These partnerships are focused on innovations like solid-state transformers and 800VDC power distribution, essential for supporting the energy demands of next-generation AI workloads.
FAQ
Q: Are GPUs becoming obsolete?
A: No, GPUs remain crucial for many AI workloads, particularly training. However, the industry is moving towards a more heterogeneous computing approach that incorporates CPUs and other specialized processors.
Q: What is agentic AI?
A: Agentic AI refers to AI systems that can autonomously plan, execute, and adapt to achieve specific goals. It requires a different type of processing power than traditional AI.
Q: What is the Vera Rubin platform?
A: The Vera Rubin platform is Nvidia’s new data center architecture, encompassing GPUs, CPUs, and other specialized chips designed to work together as an “AI factory.”
Q: Why are companies like Meta building their own chips?
A: Building custom chips allows companies to optimize performance for specific workloads, reduce reliance on external vendors, and gain more control over their silicon supply chain.
Did you know? Meta plans to release a new chip approximately every six months as part of its MTIA family.
Pro Tip: When evaluating data center solutions, consider the entire silicon stack, not just the GPUs. A balanced approach with CPUs and specialized accelerators will likely deliver the best performance and efficiency.
Want to learn more about the latest trends in data center technology? Explore our other articles or subscribe to our newsletter for regular updates.
Related reading