AMD's Lisa Su forecasts 1:1 CPU-GPU ratio, partners 01.AI for agents
AMD targets 50% GPU share in data centers by 2026 alongside new enterprise AI appliance.
At an AI Developer Day in Shanghai, AMD Chair and CEO Lisa Su painted a bold picture of the data center's future, emphasizing the meteoric rise of AI agents as a primary workload driver. She projected that the CPU-to-GPU ratio in data centers would shift dramatically from today's 1:4 to a balanced 1:1 by 2026, signaling a massive uptick in GPU demand. This shift underscores AMD's bet that agentic AI—autonomous systems that plan and execute tasks—will require heterogeneous compute, blending traditional CPUs for orchestration with GPUs for heavy inference and training.
To capitalize on this trend, AMD announced a joint enterprise agent appliance developed with 01.AI, the Chinese AI startup founded by Kai-Fu Lee. The appliance bundles AMD's Instinct GPUs with 01.AI's agent orchestration software, targeting enterprises that want to deploy AI agents on-premises for lower latency and data sovereignty. Additionally, AMD revealed it has optimized its ROCm software stack for leading Chinese foundation models, including Xiaomi's MIMO, StepFun, and MiniMax, with the goal of reducing API call costs for developers by up to 40% through better hardware utilization and memory management. These moves position AMD to compete aggressively against NVIDIA in China's AI infrastructure market, where domestic model providers are seeking cost-effective alternatives for inference workloads.
- AMD CEO Lisa Su projects data center CPU-to-GPU ratio will shift from 1:4 to 1:1 by 2026, driven by AI agent workloads.
- AMD announces joint enterprise AI agent appliance with 01.AI, combining Instinct GPUs with agent orchestration software.
- AMD optimizes ROCm platform for Chinese models (Xiaomi MIMO, StepFun, MiniMax), aiming to cut API call costs by up to 40%.
Why It Matters
AMD's agent-focused strategy could reshape enterprise AI deployment in China, offering lower-cost inference alternatives.