• AMD says agentic AI will reshape enterprise computing budgets.
• AMD projects hybrid AI PCs could cut three-year costs by 40%–60%.

The use of agentic AI in enterprise operations is changing how organisations plan for computing resources and infrastructure spending. Unlike AI assistants that respond to individual prompts, AI agents can carry out multiple tasks continuously, increasing demand for processing power and memory.

Vinay Awasthi, Senior Vice President of AMD’s Asia Pacific, Japan and India Region, said in an exclusive interview with Tech Wire Asia that businesses will need to account for AI activity rather than relying solely on traditional measures such as employee headcount or software subscriptions.

A single employee can operate several agents and sub-agents simultaneously, with each performing tasks such as retrieving information, calling tools, and coordinating workflows. This increases compute consumption and makes the cost of running AI an ongoing operational consideration.

“The productivity multiplier from AI agents comes with a compute multiplier: more agents mean more reasoning, more context, more models and more actions,” Awasthi said.

Awasthi said enterprises should measure AI spending against productivity gains and business outcomes, using metrics such as cost per completed task, task completion rates, outcome quality, and recurring inference costs.

Managing AI costs across different environments

Awasthi said enterprises should evaluate workloads based on processing requirements, latency, data sensitivity, regulatory obligations, and cost. Applications requiring low latency or involving sensitive information can benefit from local AI PCs, edge systems, or nearby data centres.

Large-scale model training and other compute-intensive operations remain suited to cloud and data-centre infrastructure, where organisations can access greater computing capacity.

“Ultimately, there is no single environment that is optimal for every AI workload. Enterprises should place each workload where it delivers the best balance of performance, security, governance and cost,” Awasthi said.

Running AI across these environments also requires visibility into resource consumption and consistent infrastructure capabilities. Local deployments need sufficient memory and processing capacity, alongside the necessary security and software support.

The economics of enterprise AI PCs

Awasthi said organisations should determine which employees need local AI capabilities before investing in new devices. Knowledge workers can use on-device AI for everyday productivity tasks, while developers and other power users may require more memory and processing capacity to run larger models.

“AI PC investments should also be role- and workload-based. Different employees require different levels of local AI capability, so organisations do not need to take a one-size-fits-all approach across their PC fleet,” Awasthi said.

AMD’s analysis projects that a fleet of 500 AMD AI PCs, with AI workloads split equally between local devices and the cloud, can deliver savings of 40% to 60% over three years compared with a cloud-only deployment.

The estimate is based on a medium workload tier, with savings varying according to the cloud model used. AMD’s analysis also projects that fully local deployments can deliver higher savings, with break-even typically achieved in under 24 months.

Beyond cost considerations, Awasthi identified several applications for local AI in enterprise workflows. Developers can build and test agents locally, while other employees can use them for research, reporting, scheduling, and repetitive tasks.

In semiconductor design, agentic AI can assist with the testing and verification process. Chip-design workflows can involve millions of simulations and thousands of issues requiring classification and resolution.

AI agents can group similar issues and prioritise problems for engineering teams, helping them identify and resolve defects earlier in the design cycle.

Why CPUs remain important for agentic AI

Running AI agents involves more than model inference, placing additional demands on CPUs alongside graphics processing units (GPUs) and neural processing units (NPUs).

Awasthi explained that agents perform activities beyond model inference, including executing commands, processing files, retrieving information, and coordinating multiple tasks.

“A fast AI model delivers little value if the surrounding workflow cannot keep pace,” Awasthi said.

CPUs manage much of this orchestration work, while GPUs handle parallel computing operations such as model inference and content generation. NPUs support dedicated AI tasks that require low-power, continuous processing.

System performance also depends on how these processors work together with memory, networking, and software. AMD’s Ryzen AI platforms use this approach to support local AI workloads, while larger computing tasks can continue running in cloud or data-centre environments.

Security and governance challenges

Deploying AI agents in daily operations also raises security and governance requirements, particularly when agents access enterprise data, applications, and services.

Awasthi identified operational complexity, security, and governance as major challenges, particularly as organisations deploy increasing numbers of agents across different computing environments.

“Enterprises need visibility and control as agents access data, applications and tools, alongside a software foundation that allows AI workloads to scale across different environments without requiring teams to manage entirely separate technology stacks,” Awasthi said.

AMD’s ROCm software platform supports AI development across its computing products. Its software tools include ROCm CLI for installing and managing workloads, Hyperloom for inference performance optimisation, and Semantic Router for model routing and token management.

AMD is also working with partners including Cisco on observability, policy control, and agent security capabilities to support the management of AI deployments across different environments.

Agentic AI adoption in Malaysia

In Malaysia and Southeast Asia, Awasthi expects enterprise adoption of agentic AI PCs to depend on specific business requirements rather than widespread hardware replacement.

“Agentic PC adoption in Malaysia and Southeast Asia will be driven by clear business use cases rather than a uniform hardware refresh,” Awasthi said.

He identified telecommunications, healthcare, manufacturing, financial services, and retail as sectors where AI is already being applied. Existing applications include network optimisation, predictive maintenance, clinical decision support, fraud detection, and demand forecasting.

While these applications provide a starting point for adopting agentic AI, Awasthi did not provide figures on current enterprise adoption or deployment readiness in Malaysia.

Over the next three to five years, AMD expects enterprise computing infrastructure to become more distributed as organisations run AI workloads across devices, edge systems, and data centres.

Awasthi said infrastructure planning will need to account for employees running multiple agents simultaneously, increasing concurrent demand for processing capacity, memory, and data access.

AMD’s approach includes open software and standards such as ROCm, UALink, and Ultra Ethernet, which the company sees as supporting interoperability and software portability across distributed AI infrastructure.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.