en.Wedoany.com Reported - July 3 news - South Korea's Samsung Electronics is emerging as a key foundry choice for global tech giants developing their own AI chip ASICs. Industry sources indicate that Samsung's foundry long-term backlog orders are expected to approach 50 trillion Korean won, with Meta advancing collaboration with Samsung to design and produce next-generation AI ASICs, with related orders potentially exceeding 10 trillion Korean won.
Meta's self-developed AI accelerator, MTIA, is moving from internal inference workloads to larger-scale deployment. MTIA primarily serves Meta's AI recommendations, advertising, content understanding, and generative AI applications, aiming to reduce full reliance on general-purpose GPUs. While the first two generations of MTIA were mainly produced by TSMC, the latest generation has shifted to collaboration with Samsung, planning to mass-produce hundreds of thousands of units using advanced 2nm process technology. For hyperscale platforms like Meta, the value of self-developed ASICs lies in optimizing energy efficiency, cost, and inference throughput around their own models and business workloads, binding hardware design, software stack, data center scheduling, and AI service demands into a unified infrastructure.
Anthropic is also evaluating the use of Samsung's 2nm process to develop AI chips. This project remains in early planning stages, not yet entering detailed design and manufacturing, but it already signals rising interest in custom chips among large model companies.
Samsung's pursuit of AI ASIC orders relies not only on advanced process technology. AI chip projects typically involve wafer manufacturing, advanced packaging, high-bandwidth memory (HBM), power management, server boards, and data center deployment. The 2nm process enhances transistor density and energy efficiency, while advanced packaging determines data exchange efficiency between compute chips and HBM. Large model inference and training demand high memory bandwidth, chip interconnect, and thermal dissipation capabilities. If a single chip's performance cannot integrate with packaging, HBM, and system architecture, it is difficult to deliver stable computing output in data centers. Samsung's simultaneous ownership of foundry, memory, and advanced packaging resources provides a more comprehensive supply chain portfolio when competing for AI clients like Meta and Anthropic.
Global tech giants are accelerating their self-developed AI chip strategies. Google has TPU, Amazon has Trainium and Inferentia, Meta is advancing MTIA, and model companies like OpenAI and Anthropic are also evaluating custom chips better suited to their workloads. ASICs will not immediately replace GPUs, but they will take on more computing power in specific inference tasks, internal cloud platforms, and large-scale stable workloads. For Samsung, if projects with Meta and Anthropic continue to materialize, it will help its foundry business improve 2nm production line utilization and compete for more client resources in the AI chip foundry market against TSMC.
