Nvidia’s GTC Highlights AI Chip Shift as CPUs Take Center Stage

Nvidia’s Grace CPU, unveiled at GTC, marks the company’s first major foray into data‑center‑class general‑purpose processors, a shift that underscores the growing importance of CPUs in AI workloads. Grace combines an Arm‑based core with Nvidia’s NVLink interconnect, allowing tight coupling with the firm’s Hopper GPUs. According to ZDNet, the chip is positioned to handle “agentic AI” workloads that require large memory footprints and low‑latency data movement, a capability that pure‑GPU solutions struggle to provide. The architecture also integrates high‑bandwidth memory (HBM) directly onto the CPU package, a design choice intended to reduce the memory bottleneck that has hampered transformer‑based models at scale.

The strategic pivot is reflected in demand metrics reported by both Nvidia and AMD. CNBC notes that the two rivals are seeing “huge demand for CPUs,” a trend the analysts attribute to enterprises deploying increasingly autonomous AI agents that need both inference speed and complex decision‑making logic. This demand is driving Nvidia to expand its silicon portfolio beyond GPUs, a move that could reshape the competitive landscape traditionally dominated by Intel and AMD in the server CPU market. By leveraging its GPU expertise and the NVLink fabric, Nvidia aims to deliver a unified compute stack that can serve both training and inference phases without the typical PCIe latency penalties.

Jensen Huang’s keynote hinted at a roadmap that extends the CPU‑centric approach beyond Grace. While the specifics were not disclosed, the presentation alluded to upcoming processors tailored explicitly for “agentic AI,” a term Huang uses to describe AI systems capable of autonomous planning and execution across multiple domains. Ars Technica reports that Nvidia plans to introduce “Rubin Ultra” and “Feynman” chips for 2027 and 2028, respectively, which will further integrate CPU and GPU functions on a single die. These future silicon offerings are expected to power “robots and billions of AI agents,” suggesting a long‑term vision where heterogeneous compute becomes the default architecture for AI‑driven services.

The shift also has implications for Nvidia’s software ecosystem. Grace is paired with a new software development kit that includes libraries for quantized inference, memory management, and cross‑processor scheduling, according to the ZDNet coverage of the GTC announcements. By providing a unified SDK, Nvidia hopes to lower the barrier for developers transitioning from GPU‑only pipelines to hybrid CPU‑GPU workflows, a move that could accelerate adoption of its new hardware in data‑center environments where mixed‑precision workloads dominate.

Analysts observing the market note that the CPU’s resurgence is not merely a reaction to Nvidia’s product rollout but also a response to the architectural demands of next‑generation AI models. As models grow in parameter count and require more sophisticated control logic, the need for general‑purpose processing power becomes acute. The convergence of high‑bandwidth memory, NVLink interconnect, and Arm‑based cores in Grace represents a concrete step toward meeting that demand, positioning Nvidia to compete directly with traditional CPU vendors while leveraging its dominant GPU ecosystem.