AMD poised to outperform Nvidia as AI shifts to inference, agents

During the first wave of the artificial intelligence boom, Nvidia (NVDA) was the dominant chip supplier, its graphics processing units (GPUs) ideal for training frontier large language models. Its lead was reinforced by CUDA, its proprietary software platform that became the standard foundation for AI code development. Nvidia sold vast numbers of GPUs and continues to do so.

However, the article argues that the center of gravity in AI infrastructure is shifting, with two major trends about to reshape the competitive landscape. The first is a transition from training to inference — the process of applying trained AI models to real-world problems. Unlike training, which is a one-time cost, inference is an ongoing operational expense where cost-per-inference becomes the dominant concern. Inference workloads, the article notes, depend much more on memory access than raw compute power. AMD is addressing this advantage through two mechanisms: its chiplet design allows it to package more memory per GPU than Nvidia, making its hardware inherently better for inference jobs. Additionally, AMD acquired MEXT, a memory optimization company whose technology uses AI itself to intelligently offload rarely-used data to flash storage and retrieve it back to DRAM just before it is needed, effectively expanding memory capacity virtually without performance penalties and reducing overall costs. AMD already has large GPU deals in place, positioning it for significant revenue growth in this segment.

The second trend is the rise of agentic AI — AI systems that autonomously plan and execute tasks through sequential reasoning. Unlike prior AI workloads that could be parallelized across GPUs, agentic systems require central processing units (CPUs) to handle sequential operations. This fundamentally changes the hardware mix data centers must purchase. In inference-optimized data centers, the typical ratio recently was around 1 CPU for every 8 GPUs. For agentic AI, however, Intel predicts that ratio will shift to 1 CPU for every 1 GPU — a doubling of CPU demand. AMD has long been a leader in high-performance data center CPUs and has already released high-core CPUs specifically designed to handle AI agents, positioning it to gain market share from Intel, its traditional rival in the CPU market. The article concludes that with AMD riding both of these powerful infrastructure trends — inference cost optimization and agentic AI CPU demand — its stock is expected to outperform Nvidia's in the years ahead, even though Nvidia will retain its dominance in AI training chips.

Nvidia's dominance during the first phase of the AI boom — centered on training frontier large language models — was built on two advantages: powerful GPUs and the CUDA software platform that became the industry standard for foundational AI code. However, the article identifies a structural shift in AI infrastructure spending that could redistribute market leadership without overthrowing Nvidia's absolute scale. The move from training to inference as the primary workload represents not just a different computational task, but a different economic constraint: inference is a recurring operational expense where cost-per-inference and memory efficiency matter far more than the raw parallel compute that GPUs excel at.

AMD's positioning rests on two concrete technical advantages aligned with this shift. First, its chiplet architecture allows it to package more memory per GPU than Nvidia's designs, a direct answer to the memory-centric nature of inference workloads. The MEXT acquisition signals AMD's intent to control not just hardware but the full stack — allowing it to virtualize memory management through software, reducing the absolute memory a customer must buy while preserving performance. Second, the rise of agentic AI (systems that plan and execute sequential tasks) structurally changes the CPU-to-GPU ratio data centers need to buy. AMD already leads in high-performance data center CPUs and has released processors designed specifically for AI agents, positioning it to capture share from Intel — its traditional rival in that segment — rather than Nvidia directly. The article's claim that AMD's stock is "poised to outperform" rests on these two independent growth vectors hitting simultaneously, each grounded in a different hardware bottleneck and buying pattern.