Hit enter to search or ESC to close
Trending Now
AI is evolving too quickly for infrastructure leaders to predict exactly what the technology will look like five years from now. The bigger challenge is building systems today that can adapt as workloads, architectures and requirements continue to change.
Attributed to Alexey Navolokin, General Manager, APAC
No one in the industry can say with certainty what artificial intelligence will look like five years from now. The next dominant model architecture remains unknown, as does the eventual balance between training and inference. It is also unclear how much AI will ultimately run in an organisation's own data centre, someone else's data centre, on a laptop or on something even smaller.
Despite that uncertainty, infrastructure leaders still need to make decisions and investments today that could form the IT foundations of their businesses for the next five to seven years.
This makes the ability to adapt increasingly important. The greatest risk is not necessarily choosing the wrong technology today, but building infrastructure that cannot adjust as workloads and requirements evolve tomorrow.
At the AI Infra Summit 2026, the discussion centred on the need to rethink how AI infrastructure is designed. Rather than designing around a particular chip or fabric, the more fundamental consideration is how easily a system can adapt when its requirements change.
Adaptability Becomes the Design Principle
Only a few years ago, AI infrastructure was largely associated with large training runs that generated predictable, batch-shaped demand on a major cluster. While training remains important, infrastructure must now accommodate inference, agentic AI and an increasingly distributed range of workloads, each with different requirements.
Inference is continuously active and places greater emphasis on latency and cost per request. Agentic AI can compound those demands because a single request may trigger a sequence of steps. It can retrieve and shape inputs, call tools, execute code, coordinate sub-agents and maintain a state that extends beyond an individual request.
The result is not simply a need for more computing power. The shape of the overall system also changes. Infrastructure optimised around yesterday's training clusters may have been designed around a balance of compute, memory and networking that no longer reflects where workloads are heading. Simply adding more of the same infrastructure does not necessarily solve that mismatch.
Instead, the focus needs to be on optimising the system as a whole. Compute, memory, networking, storage, software, power and reliability all need to work together. Scaling an existing system can largely become a budgeting exercise, while accommodating something entirely new presents architectural and operational challenges.
Three Properties That Preserve Flexibility
If AI requirements continue to change, the objective is not necessarily to predict which technology will ultimately prevail. Instead, infrastructure needs to preserve the freedom to move towards whatever comes next. In practice, this can be considered through three auditable properties.
Adaptable: Organisations need a portfolio broad enough that when workloads shift, the solution can be a different configuration rather than requiring a different vendor. This means matching the appropriate compute to each workload while maintaining a common software foundation that makes those choices practical.
Economical: Being economical does not simply mean choosing the cheapest option. It means paying for the right tool for the job. Training may have traditionally represented a capital decision made at a particular point in time, while inference becomes an operating decision potentially repeated billions of times each day. With an agentic request potentially involving a chain of inferences, cost per completed task becomes more relevant than simply looking at cost per token.
Open: Open ecosystems provide greater choice through broadly adopted standards, allowing customers to reduce their dependence on a single vendor's product roadmap. This provides choice across suppliers and highlights the importance of open interconnects and open networking.
Where the AMD Portfolio Comes In
The breadth of available technology becomes particularly important in this environment. AMD has invested across the technology stack rather than focusing solely on a single flagship product, reflecting the range of requirements emerging across AI workloads today.
These requirements span:
AMD's approach is to build across the stack while allowing each component to compete on its own merits, rather than requiring customers to purchase the entire stack for it to function. Open standards play an important role in making this possible, enabling organisations to choose the AMD technologies that best suit their workloads while retaining the freedom to make different choices elsewhere.
Over the next five years, organisations that navigate AI infrastructure successfully may not necessarily be those that predicted every change in advance. Instead, their advantage may come from maintaining enough freedom and flexibility to respond when those changes arrive.
The approach is to build for adaptability so infrastructure can accommodate whatever shape the future takes, build economically so organisations are paying for the appropriate technology rather than simply the largest option, and build around openness to preserve choice.
While the exact direction of AI cannot be guaranteed, continued change is one factor infrastructure leaders can expect. Designing infrastructure around that assumption from the beginning could therefore become increasingly important.
No one in the industry can say with certainty what artificial intelligence will look like five years from now. The next dominant model architecture remains unknown, as does the eventual balance between training and inference. It is also unclear how much AI will ultimately run in an organisation's own data centre, someone else's data centre, on a laptop or on something even smaller.
Despite that uncertainty, infrastructure leaders still need to make decisions and investments today that could form the IT foundations of their businesses for the next five to seven years.
This makes the ability to adapt increasingly important. The greatest risk is not necessarily choosing the wrong technology today, but building infrastructure that cannot adjust as workloads and requirements evolve tomorrow.
At the AI Infra Summit 2026, the discussion centred on the need to rethink how AI infrastructure is designed. Rather than designing around a particular chip or fabric, the more fundamental consideration is how easily a system can adapt when its requirements change.
Adaptability Becomes the Design Principle
Only a few years ago, AI infrastructure was largely associated with large training runs that generated predictable, batch-shaped demand on a major cluster. While training remains important, infrastructure must now accommodate inference, agentic AI and an increasingly distributed range of workloads, each with different requirements.
Inference is continuously active and places greater emphasis on latency and cost per request. Agentic AI can compound those demands because a single request may trigger a sequence of steps. It can retrieve and shape inputs, call tools, execute code, coordinate sub-agents and maintain a state that extends beyond an individual request.
The result is not simply a need for more computing power. The shape of the overall system also changes. Infrastructure optimised around yesterday's training clusters may have been designed around a balance of compute, memory and networking that no longer reflects where workloads are heading. Simply adding more of the same infrastructure does not necessarily solve that mismatch.
Instead, the focus needs to be on optimising the system as a whole. Compute, memory, networking, storage, software, power and reliability all need to work together. Scaling an existing system can largely become a budgeting exercise, while accommodating something entirely new presents architectural and operational challenges.
Three Properties That Preserve Flexibility
If AI requirements continue to change, the objective is not necessarily to predict which technology will ultimately prevail. Instead, infrastructure needs to preserve the freedom to move towards whatever comes next. In practice, this can be considered through three auditable properties.
Adaptable: Organisations need a portfolio broad enough that when workloads shift, the solution can be a different configuration rather than requiring a different vendor. This means matching the appropriate compute to each workload while maintaining a common software foundation that makes those choices practical.
Economical: Being economical does not simply mean choosing the cheapest option. It means paying for the right tool for the job. Training may have traditionally represented a capital decision made at a particular point in time, while inference becomes an operating decision potentially repeated billions of times each day. With an agentic request potentially involving a chain of inferences, cost per completed task becomes more relevant than simply looking at cost per token.
Open: Open ecosystems provide greater choice through broadly adopted standards, allowing customers to reduce their dependence on a single vendor's product roadmap. This provides choice across suppliers and highlights the importance of open interconnects and open networking.
Where the AMD Portfolio Comes In
The breadth of available technology becomes particularly important in this environment. AMD has invested across the technology stack rather than focusing solely on a single flagship product, reflecting the range of requirements emerging across AI workloads today.
These requirements span:
- Inference, tokenisation, orchestration and other general-purpose compute on AMD EPYC™ Server CPUs, extending to small models and local AI on AMD Radeon™ and AMD Ryzen™ AI platforms.
- Large language model inference and fine-tuning on AMD Instinct™ PCIe cards.
- Distributed inference and training on eight-way AMD Instinct™ GPU systems, scaling to the AMD Helios™ rackscale solution for mega-scale deployments.
AMD's approach is to build across the stack while allowing each component to compete on its own merits, rather than requiring customers to purchase the entire stack for it to function. Open standards play an important role in making this possible, enabling organisations to choose the AMD technologies that best suit their workloads while retaining the freedom to make different choices elsewhere.
Over the next five years, organisations that navigate AI infrastructure successfully may not necessarily be those that predicted every change in advance. Instead, their advantage may come from maintaining enough freedom and flexibility to respond when those changes arrive.
The approach is to build for adaptability so infrastructure can accommodate whatever shape the future takes, build economically so organisations are paying for the appropriate technology rather than simply the largest option, and build around openness to preserve choice.
While the exact direction of AI cannot be guaranteed, continued change is one factor infrastructure leaders can expect. Designing infrastructure around that assumption from the beginning could therefore become increasingly important.
.jpeg)
