Nvidia Keeps AI Chip Crown With CPU Push, Huang's Vision

Nvidia was founded in 1993 and invented the graphics processing unit (GPU) in 1999, a breakthrough that accelerated graphics rendering and transformed the video game market. However, CEO and co-founder Jensen Huang's most strategically important move was the creation of CUDA, a software platform that enabled Nvidia's chips to be programmed for tasks far beyond graphics. Huang seeded this technology into universities and research labs conducting early AI work, a decision that proved prescient: most foundational AI code was written on CUDA for Nvidia GPUs, creating a durable competitive moat in AI model training.

Huang did not pause there. In 2020, Nvidia acquired Mellanox, a networking company whose technology was ahead of its time. Huang recognized that AI infrastructure would eventually demand sophisticated networking capabilities; today, Mellanox's networking portfolio ranks as Nvidia's fastest-growing business segment and has transformed the company from a GPU specialist into a complete AI infrastructure provider. Similarly, Huang anticipated the industry's shift toward inference (the phase where AI models respond to user queries) and agentic AI (AI systems that manage tasks autonomously). In response, Nvidia developed its own ARM-based central processing units (CPUs) to manage AI agents in data centers.

This expansion into CPUs addresses a fundamental change in data center architecture. When AI workloads were dominated by training, the ratio of GPUs to CPUs in data centers was 8 to 1. As cloud companies build infrastructure for agentic AI, Nvidia projects this ratio could shift to 1 to 1, opening a massive new market. The company has estimated that the data center CPU market could reach a value of $200 billion（約32兆円） in the next few years. Additionally, Nvidia acquired assets and key personnel from Groq, including language processing units (LPUs), which it has integrated into CUDA. These LPUs are designed for servers optimized for inference, a market segment expected to eventually exceed AI model training in size. Nvidia's new server design will pair GPUs—handling the prefill phase of understanding user prompts—with LPUs, which manage the decode phase of delivering quicker responses, positioning this as a potential next major growth driver. Currently, Nvidia trades at 16 times analysts' earnings estimates for fiscal 2028 (ending January 2028), and the company's top and bottom lines are growing rapidly. Huang's track record of visionary positioning suggests the stock remains attractive for long-term AI exposure.

Nvidia's dominance in AI rests on a deliberate, multi-decade pattern of anticipating market shifts before they happen. The 1999 invention of the GPU accelerated computer graphics, but Huang's more decisive move was creating CUDA, a software platform that made Nvidia's chips programmable for tasks beyond graphics. By seeding CUDA into universities and early AI research labs, Nvidia ensured that most foundational AI code was written on its hardware—a structural advantage that persists today in AI model training.

The company has extended this playbook into newer AI segments. The 2020 acquisition of Mellanox was not justified by immediate returns but by Huang's foresight into the networking demands of AI infrastructure; today, Mellanox's networking portfolio is Nvidia's fastest-growing business unit. More recently, the incorporation of Groq's language processing units signals Nvidia's pivot toward inference—the phase where AI models answer user queries. This shift matters because cloud providers must now build dual-GPU and CPU systems to manage both training and agentic workloads, expanding Nvidia's addressable market and changing the hardware mix in data centers from 8-to-1 (GPU-to-CPU) toward parity.