In an interview with Tech Wire Asia, Vinay Awasthi, AMD Senior Vice President for Asia Pacific and Japan and India, explained that unlike AI Agents that provide one-off responses, AI Agents continuously execute tasks and invoke sub-agents, driving up both computing and memory requirements.
This observation highlights one of the most commonly underestimated considerations when manufacturers adopt AI Agents: the cost of computing. For Taiwan’s industrial supply chain, which relies heavily on semiconductors and precision manufacturing, this shift in thinking is particularly relevant.
The Cost of Multiplied Productivity Is Multiplied Computing Demand
Awasthi identified the core logic in the interview: the other side of multiplied productivity is multiplied computing demand. If one employee operates multiple Agents and sub-agents simultaneously, compute consumption is no longer an occasional spike but an ongoing operating expense. At the same time, knowledge workers, developers, and advanced users do not require the same computing and memory configurations.
He recommends that enterprises stop measuring AI spending by headcount or software licensing fees. Instead, they should use four metrics: cost per completed task, task completion rate, output quality, and the cost of continuous inference. Only then can they see the relationship between AI spending and actual outcomes. He also cautioned that over the next three to five years, enterprise computing infrastructure will become more distributed, with AI workloads running across devices, edge systems, and data centers.
This division of responsibilities also extends to employee roles. Developers can build and test Agents locally, while other employees are more likely to use them for repetitive work such as research, reporting, and scheduling. The two groups should not be expected to use identical computing and memory configurations. The article also cites another important point: no matter how fast a model is, its overall value remains limited if the surrounding workflows cannot keep pace. This is why Awasthi emphasizes task-completion efficiency, with speed being only one of several contributing factors.
Manufacturing Use Cases: Predictive Maintenance and Semiconductor Verification
The report identifies telecommunications, healthcare, manufacturing, financial services, and retail as five major industries already deploying AI Agents. Applications include network optimization, predictive maintenance, clinical decision support, fraud detection, and demand forecasting. The use case most directly relevant to Taiwan’s industrial supply chain is semiconductor design, testing, and verification. Agents can help classify and process large volumes of anomalous signals generated through simulation, enabling engineering teams to identify and eliminate defects earlier and directly shorten the overall verification cycle.
The article also discusses how workloads should be allocated. Latency-sensitive applications or those involving sensitive data—such as real-time production-line data and process parameters—are best kept on local devices or edge systems. Compute-intensive workloads such as large-model training remain better suited to the cloud. CPUs handle instruction orchestration, file access, and multitasking coordination; GPUs handle parallel computing and model inference; and NPUs handle low-power, continuous workloads. For production-line deployments in Taiwanese factories, this hybrid division of labor between the edge and the cloud is more practical than moving everything to the cloud. AMD also highlighted tools such as its ROCm software platform and Semantic Router, which are designed to address scheduling challenges such as model routing and token management, preventing distributed deployments from becoming unmanageable.
What This Means for Taiwanese Manufacturers
Taiwan’s electronics manufacturing services, semiconductor supply chain, and precision machinery industries already have the digital foundations required for systems such as predictive maintenance and automated optical inspection. Adopting Agentic AI is therefore an extension of existing capabilities. The real differentiator is whether governance and access controls have been established in advance. AMD’s collaboration with Cisco on observability and policy controls reflects this reality: once Agents can access internal enterprise systems and tools, security and governance are not matters to address after launch. They are prerequisites that must be clearly defined before deployment.
The article states that “a hybrid deployment of 500 AI PCs can reduce computing costs by 40% to 60% over three years,” with even greater savings and a payback period typically under 24 months for fully local deployments. However, these figures come from AMD’s own analysis, and neither the methodology nor real-world customer data has been disclosed. They should therefore be viewed as an optimistic vendor estimate rather than a directly replicable outcome. Awasthi also noted that the adoption of agentic AI PCs in Southeast Asia and Malaysia will depend on specific business needs rather than the timing of a one-off, fleet-wide hardware refresh. The same criterion applies to Taiwanese manufacturers: first determine which processes genuinely require the technology, then decide whether the device fleet needs to be upgraded. A more pragmatic path for Taiwanese manufacturers is to begin with small-scale, task-oriented pilots, calculate their own cost per task, and then determine which production lines should be expanded and to what scale. EgentWrX approaches enterprise adoption through a three-layer architecture of Agent, Task, and Skill, supported by governance mechanisms—reflecting the same sequence of defining and quantifying tasks before pursuing scale.
FAQ
How credible is the report’s claimed savings of 40% to 60%? The article itself cautions that these figures are AMD’s own estimates and are not supported by a publicly disclosed methodology or actual customer deployment data. Given the clear vendor perspective, enterprises should validate the economics using data from their own pilots rather than applying the figures directly.
Can Taiwan’s traditional manufacturing sector keep pace with this wave of Agentic AI? Based on the manufacturing use cases highlighted in the report, predictive maintenance and semiconductor design, testing, and verification are not far removed from the existing digital capabilities of Taiwan’s electronics manufacturing services and semiconductor supply chain. The primary gap is usually whether governance and access controls are in place; the technology itself is not the real bottleneck.
Should Agents run at the factory edge or in the cloud? The report indicates that latency-sensitive applications or those involving sensitive data—such as real-time production-line data and process parameters—are best handled locally or at the edge. Compute-intensive tasks such as large-model training remain better suited to the cloud. Enterprises should adopt a hybrid deployment based on workload characteristics rather than treating edge and cloud as an either-or choice.
How can enterprises determine whether adopting AI Agents is worthwhile? Awasthi recommends replacing simple headcount or subscription-cost calculations with four metrics: cost per completed task, task completion rate, output quality, and the cost of continuous inference. The same measurement framework can be applied when evaluating manufacturing pilots.
References
- Tech Wire Asia, “How agentic AI is changing enterprise computing costs, according to AMD”, industry media interview (interviewee: Vinay Awasthi, AMD Senior Vice President for Asia Pacific and Japan and India; published October 19, 2026)
