Summary
AMD says AI infrastructure should be designed around adaptability as training, inference and agentic workloads evolve. Its framework emphasises hardware breadth, common software, cost per task and open standards.
AMD is arguing that AI infrastructure should be designed for changing workloads rather than optimised around a single chip, networking fabric or model architecture.
In a September 17, 2026 Newsroom post, Forrest Norrod, AMD’s executive vice president and general manager of its Data Center Solutions Business Group, said infrastructure leaders are making investments intended to support businesses for five to seven years even though the future mix of AI workloads is difficult to predict. The post presents adaptability as the central design principle for those systems.
Why AI workloads are changing shape
Large training runs remain an important use of computing infrastructure, but AMD says they now coexist with inference, agentic AI and more distributed workloads.
Training is generally organised as a batch workload. Inference, by contrast, runs continuously and places greater emphasis on latency—the time taken to respond—and cost per request. Agentic AI can make a single user request into a chain of operations: retrieving and preparing information, calling tools, executing code, coordinating other agents and retaining state across steps.
That changes the balance required between compute, memory and networking. AMD’s position is that an infrastructure system built primarily for large training clusters may have the wrong configuration for these newer patterns; simply adding more of the same hardware may not address that mismatch.
The company says the system therefore has to be designed as a whole, with compute, memory, networking, storage, software, power and reliability considered together.
AMD’s three principles for AI infrastructure
AMD describes three properties that it says can help organisations preserve flexibility as requirements change:
- Adaptable: A broad hardware portfolio should allow organisations to select different configurations for different workloads while retaining a common software foundation.
- Economical: The goal is not necessarily the lowest purchase price, but matching the tool to the job. AMD distinguishes the capital decision involved in training infrastructure from the repeated operating cost of inference, and says agentic systems make cost per completed task more important than cost per token alone.
- Open: Broadly adopted standards, open interconnects and open networking can give customers more choice across suppliers and reduce dependence on a single vendor’s product roadmap.
In AMD’s framing, these principles are intended to preserve the ability to change hardware or suppliers when workload requirements evolve, rather than requiring a complete infrastructure redesign.
How AMD positions its portfolio
AMD says its products span several layers of AI computing. EPYC server CPUs are positioned for inference, tokenisation, orchestration and other general-purpose computing tasks. Radeon and Ryzen AI platforms extend the company’s offering to smaller models and local AI workloads.
For larger deployments, AMD identifies Instinct PCIe cards for large-language-model inference and fine-tuning, eight-way Instinct GPU systems for distributed inference and training, and the AMD Helios rack-scale solution for what it calls mega-scale deployments.
The company says this range is useful only when it is connected by a common software environment. AMD identifies ROCm software as that foundation, arguing that workloads can move between hardware configurations without requiring the software to be started from scratch. The post also presents open standards as a way to use selected AMD components without making the entire infrastructure dependent on AMD.
Why the design choice matters
The practical issue raised by AMD is that AI infrastructure decisions are made before the industry knows which model architectures, deployment locations or workload mixes will dominate. AI may run in central data centres, in third-party infrastructure, on laptops or on smaller local devices, while the balance between training and inference may continue to change.
That makes flexibility an operational concern as well as a hardware choice. Systems must be able to respond to differences in latency, memory, networking and repeated inference cost, particularly when agentic workflows turn one request into multiple coordinated steps.
AMD’s post is a strategic framework and portfolio position rather than a new benchmark or product announcement. Its central recommendation is to treat adaptability, operating economics and openness as system-level design requirements from the beginning.