AI Computing Demand Hits a New Inflection Point: Inference Workloads Surpass Training
In August 2026, the global semiconductor industry is undergoing a quiet yet far-reaching architectural transformation. Over the past two years, the market focused on NVIDIA's GPUs for AI model training, but as large language models like ChatGPT and Gemini enter the commercial deployment phase, AI inference workloads have officially surpassed training, becoming the new main driver of chip demand. According to the latest supply chain reports, custom Application-Specific Integrated Circuits (ASICs) are rapidly eroding the market share of general-purpose GPUs. This 'inference chip war' is reshaping the global semiconductor supply chain landscape, and Thailand, as a key packaging and testing hub in Southeast Asia, is facing a new wave of technological upgrades and incoming orders.
GPU Is No Longer the Only Solution: ASIC Inference Performance and Cost Dual Suppression
In AI inference scenarios, chips do not need to handle the massive bidirectional data transmission required for training; instead, the focus is on low latency, low power consumption, and high throughput per unit cost. This allows the advantages of Application-Specific Integrated Circuits (ASICs) to fully emerge. Compared to general-purpose GPUs, ASICs optimized for specific algorithms can deliver 5 to 10 times the energy efficiency ratio in inference tasks, at only one-third to one-half the cost of equivalent GPU computing power. Recent financial reports from Broadcom and Marvell show their custom AI chip business revenue grew over 120% year-over-year in Q2 2026, with major clients including cloud giants like Google, Meta, and Amazon, who are accelerating the replacement of some GPU purchases with self-developed ASICs. This trend not only weakens NVIDIA's monopoly in the inference market but also significantly increases the demand for "diversification" in wafer foundry and packaging and testing segments.
Thailand's Packaging and Testing Chain Hits a Customization Sweet Spot
The explosive growth of ASICs brings new challenges and opportunities to the packaging and testing industry. Unlike standardized GPUs, ASIC chips are highly customized, with varying chip designs, power management, and thermal structures, making advanced packaging and heterogeneous integration rigid demands. Thailand has long been known in the global semiconductor supply chain for its mature back-end packaging and testing capacity, especially in automotive electronics and power management ICs. As the ASIC wave arrives, several major Thai packaging and testing companies are accelerating the adoption of Fan-Out and 2.5D packaging technologies to meet the rapid prototyping and mass production needs of cloud giants for AI inference chips. Industry insiders point out that construction plans for related advanced packaging lines in industrial parks in Samut Prakan and Ayutthaya provinces are scheduled through 2027, with capacity utilization expected to exceed 95% in the third quarter.
Investment Perspective: Shifting from General-Purpose Computing Power to Specialized Ecosystems
For semiconductor investors, the core indicator for the second half of 2026 has shifted from "who can provide the most powerful GPU" to "who can build the most comprehensive ASIC ecosystem." This includes not only the CoWoS capacity allocation of leading foundries like TSMC but also the back-end supporting capabilities in Southeast Asia. With its cost advantages and policy stability within the ASEAN supply chain, Thailand is becoming the preferred spillover base for customized ASIC packaging and testing. In terms of investment strategy, beyond focusing on design-side companies like Broadcom and MediaTek, Thai local packaging and testing firms with advanced packaging technologies and related equipment suppliers have become the new focus of capital pursuit. This wave of ASICs driven by AI inference is propelling Thailand from a mere "cost depression" to a core link in the global semiconductor "technology supply chain."
Future Outlook: The Long-Term Co-opetition Between ASICs and GPUs
While ASICs are unstoppable in the inference domain, this does not mean the exit of GPUs. In complex general AI training and High-Performance Computing (HPC), the flexibility of GPUs remains irreplaceable. The future chip market will exhibit a polarized pattern of "GPUs for training, ASICs for inference." For Thailand's semiconductor industry, this means it must simultaneously possess the dual capabilities to handle large-size GPU chips (via Chiplet technology) and highly complex ASIC packaging. The chip industry's direction for the second half of 2026 points clearly to: Whoever can solve the "last mile" packaging and testing challenges of customized chips will hold the hardware lifeline of the AI era.
