[Edaily Reporter Choi Oh-hyun ] As global Big Tech companies accelerate their shift away from NVIDIA, attention is focused on whether this could become another growth driver for Samsung Electronics and SK Hynix. This is driven by expectations that the pool of suppliers for High-Bandwidth Memory (HBM) could expand as companies like Microsoft (MS), Google, and Amazon successively accelerate their development of proprietary artificial intelligence (AI) chips. In addition to memory, Samsung Electronics could aim to expand its foundry (semiconductor contract manufacturing) orders. However, since memory specifications and semiconductor packaging technologies vary by customer, the ability to provide customized development is expected to emerge as a new competitive factor.

According to industry sources on the 15th, Microsoft recently announced that starting next year, it will adopt its proprietary AI accelerator, the “Azure Maia” chip, as the core AI inference system for its products and services. According to Reuters, Microsoft plans to deploy more than 1 million “Maia 300” chips next year. This follows the company’s unveiling of the Maia 200 earlier this year and signals that it is accelerating the release of its next-generation products. It is reported that the Maia 300 will be equipped with eight 6th-generation high-bandwidth memory (HBM4) modules, each with a capacity of 288 gigabytes (GB). Even by a simple calculation, this translates to an increase in demand for 8 million HBM modules. Compared to the initial model, the Maia 100 (64GB), the memory capacity has increased 4.5-fold.

Other big tech companies are also accelerating the development of their own AI chips. Google is continuously enhancing its proprietary AI semiconductor, the Tensor Processing Unit (TPU). The HBM capacity of the next-generation TPU 8i is expected to be 288 GB, a 50% increase over the existing 7th-generation Ironwood (192 GB). Amazon’s Trainium 3 also features 144GB of HBM, a 50% increase over its predecessor, Trainium 2 (96GB). Trainium 4 is also under development, with a target launch date of 2027. Meta is likewise pushing forward with the development of its own AI accelerator, the MTIA.

Big Tech companies are staking their futures on developing their own AI chips for reasons of cost and efficiency. As the AI industry’s focus shifts from training to inference—the process of delivering actual services—power efficiency, response latency, and return on investment are becoming increasingly critical. Utilizing ASICs optimized for their own AI services allows them to reduce unnecessary computations and improve power efficiency. There is also an underlying intention to reduce dependence on NVIDIA and diversify the semiconductor supply chain. Even if initial development costs are incurred, once mass production begins, the unit cost per chip can be significantly reduced compared to GPUs in the long term.

As demand for HBM—which had previously been concentrated on NVIDIA GPUs—expands to include Big Tech companies’ proprietary ASICs, a new market has opened up for domestic memory semiconductor companies. Samsung Electronics and SK Hynix hold overwhelming market shares, accounting for 39% and 48%, respectively, of the global HBM market last year. SK Hynix is reportedly the sole supplier of the HBM3E used in Microsoft’s previous-generation Maia 200. All eyes are on whether SK Hynix will be able to secure orders from Microsoft for the Maia 300 as well. HBM from Samsung Electronics and SK Hynix is also used in Google’s TPUs.

Market research firm Counterpoint Research forecasts that the total bit demand for HBM used in ASICs for AI servers will increase 35-fold from 2024 to 2028. It also forecast that the average HBM capacity per ASIC will increase approximately fivefold during the same period. Analysts suggest that HBM demand will be driven by a combination of expanding shipments of proprietary AI chips and increased memory capacity per chip.

In particular, Samsung Electronics is eyeing new opportunities in its foundry business. This is because even when Big Tech companies design their own AI chips, they typically outsource actual production to external semiconductor manufacturers. Microsoft’s Maia 300 is reportedly set to utilize Taiwan Semiconductor Manufacturing Company’s (TSMC) cutting-edge process technology. However, as TSMC—the industry leader—nears the limits of its production capacity for advanced processes, there is a high likelihood that the “trickle-down effect” of unmet demand will benefit Samsung Electronics.

The growth outlook for the ASIC market is also bright. Last August, Gartner projected that the global custom ASIC market would nearly triple in size, growing from $67.7 billion last year to $203.6 billion by 2030. There are also forecasts that ASIC shipments will surpass those of GPUs in the high-performance cloud AI accelerator market by 2027.

Amid these trends, the importance of customer-tailored technological capabilities is expected to grow even further. This is because the AI chips designed by Big Tech companies differ in everything from computational architecture to memory capacity, data transfer speeds, and packaging methods. An industry official stated, “While the volume of memory production is important, it is becoming increasingly crucial to quickly and reliably deliver the performance customers desire through customized solutions.” In this process, differences in HBM specifications across customers could potentially increase development costs and the burden of technical adaptation.

Both Samsung Electronics and SK Hynix have already begun preparing for the next-generation market. Ahn Hyun, President of Development at SK Hynix, recently spoke about the outlook for next-generation HBM at the Semiconductor Expo, stating, “In the long term, the trend toward customization will expand,” and added, “SK Hynix is well-prepared with various core technologies.” Samsung Electronics also plans to sequentially supply custom HBM samples tailored to individual customer requirements starting next year.