Big Tech Moves Away from Nvidia: Chance for Samsung, SK hynix?

Azure Maia chip of MS. (Photo courtesy of MS) As global big tech companies accelerate their movement to break away from Nvidia, attention is being drawn to whether this could become another growth axis for Samsung Electronics and SK hynix. This is due to the expectation that the supply destinations for High Bandwidth Memory (HBM) could expand accordingly, as Microsoft (MS), Google, Amazon, and others successively speed up the development of their own artificial intelligence (AI) semiconductors. In addition to memory, Samsung Electronics can also aim to expand its foundry (semiconductor contract manufacturing) orders. However, as the memory specifications and semiconductor packaging technologies required by each customer differ, customized development capabilities are expected to emerge as a new competitive variable. According to the industry on Oct. 15, MS recently announced that it would use its own AI accelerator Azure Maia chip as the main AI inference system for its products and services starting next year. According to Reuters, MS plans to distribute more than 1 million units of Maia 300 next year. Following MS’s unveiling of Maia 200 early this year, it is speeding up the unveiling of its next-generation products. It is reported that eight 6th-generation High Bandwidth Memory (HBM4) chips with a capacity of 288 gigabytes (GB) will be mounted on the Maia 300. By simple conversion, this means an increase in demand for 8 million HBM units. Compared to the initial product Maia 100 (64GB), the memory capacity has increased 4.5 times. Other big tech companies are also speeding up the development of their own AI chips. Google is continuously upgrading its Tensor Processing Unit (TPU), its own AI semiconductor. The HBM capacity of the next-generation TPU 8i is 288GB, which is expected to increase by 50% compared to the existing 7th-generation Ironwood (192GB). Amazon’s Trainium3 also has an HBM capacity of 144GB, an increase of 50% compared to its predecessor Trainium2 (96GB). Trainium4 is also under development with the goal of being supplied in 2027. Meta is also pushing ahead with the development of its own AI accelerator, MTIA. The reason big tech companies are staking their lives on developing their own AI chips is because of cost and efficiency. As the center of gravity in the AI industry expands from learning to inference, which provides actual services, power efficiency, response latency, and return on investment are becoming important. Utilizing an ASIC optimized for their own AI services can reduce unnecessary computations and improve power efficiency. The underlying intention is also to diversify the semiconductor supply chain by lowering dependence on Nvidia. Even if initial development costs are incurred, once mass production begins, the unit price per chip can be significantly reduced compared to a GPU in the long run. As the HBM demand, which had been concentrated on Nvidia GPUs, expands to the big tech companies’ own ASICs, a new market has opened for domestic memory semiconductor companies. Samsung Electronics and SK hynix show an overwhelming market share, accounting for 39% and 48% of the global HBM market last year, respectively. It is known that SK hynix is exclusively supplying the HBM3E mounted on MS’s previous work, the Maia 200. Attention is focused on whether SK hynix will also be able to win orders for MS’s volume for the Maia 300. Samsung Electronics and SK hynix’s HBMs are also being used in Google TPUs. Market research firm Counterpoint Research projected that the total bit demand for HBM mounted on ASICs for AI servers will increase 35 times from 2024 to 2028. During the same period, the average HBM capacity per ASIC is also expected to increase by about 5 times. The analysis is that the increase in memory capacity per chip, coupled with the expansion of shipments of their own AI chips, will drive up HBM demand. In particular, Samsung Electronics is also seeing new opportunities in the foundry business. This is because even if big tech companies design their own AI chips, actual production is entrusted to external semiconductor manufacturers. MS’s Maia 300 is known to utilize the advanced processes of Taiwan’s TSMC. However, as the advanced process production capacity of TSMC, the industry leader, is reaching its limit, there is a high possibility that the spillover effect of the volume that fails to keep up with demand will lead to Samsung Electronics. The growth prospect of the ASIC market is also bright. Gartner predicted in August that the global custom ASIC market size would grow approximately 3 times from $67.7 billion last year to $203.6 billion in 2030. There are also forecasts that ASIC shipments will surpass GPUs in the high-performance cloud AI accelerator market in 2027. Amid this, the importance of customized technological capabilities is expected to grow further. This is because the AI chips designed by big tech companies are all different, from their computational structures to memory capacities, data transmission speeds, and packaging methods. An industry official stated, “While how much memory is produced is important, how quickly and stably the performance desired by the customer is implemented in a customized manner is becoming important.” In this process, there is also a possibility that the burden of development costs and technical responses will increase due to differences in HBM specifications for each customer. Both Samsung Electronics and SK hynix have already begun preparing for the next-generation market. Ahn Hyun, head of development at SK hynix, recently forecasted next-generation HBM at the Semiconductor Exhibition, saying, “In the long term, the custom trend will expand,” and “SK hynix is well preparing various core technologies.” Samsung Electronics also plans to sequentially supply custom HBM samples tailored to each customer’s needs starting next year.