AMD has introduced a new development strategy in the field of AI computing, placing primary emphasis on the transition from model training to exploitation (inference). According to the company's forecasts, by 2026, inference spending will exceed training costs by 1.6x, and by 2030, this ratio will reach 3.2:1.
What Happened
AMD announced a comprehensive roadmap including the new MI400 architecture codenamed "Altair," "Helios" rackscale solutions, and Epyc 9006 series "Venice" processors. These technologies are aimed at supporting agentic AI, which requires high-speed real-time data processing.
Context
The artificial intelligence industry is undergoing a fundamental shift: the phase of massive training of giant models is concluding, giving way to the phase of large-scale use and exploitation (inference). This transition creates colossal demand for optimized hardware capable of providing low latency during task execution.
Why It Matters for the Industry
The shift in focus toward inference opens opportunities for creating scalable agentic systems. AMD aims to become a key competitor to NVIDIA by moving from selling individual chips to providing ready-to-use rackscale solutions (complete server racks), which could significantly reduce the total cost of ownership (TCO) of infrastructure for major players.
Why It Matters for Users
For developers and users, this signifies the beginning of the era of mass adoption of AI agents. The development of specialized hardware will enable the implementation of agentic workflows, where computations are distributed across systems that ensure instantaneous response and real-time code execution.
Sources
Author
Look at AI, Editorial Team