Jensen Huang, CEO of Nvidia, said on the 7th (local time) regarding Samsung Electronics' high bandwidth memory (HBM), "I understand that Koreans are anxious about this issue, but new designs take time," and "While we are currently experiencing some difficulties, I am confident that we will recover." This statement implies that Samsung's HBM essentially needs redesign and, considering the time required for design, testing, and mass production, it will be difficult to pass quality tests for a while.
Following Huang's comments, negative forecasts have been raised regarding the supply of 5th generation HBM, which hinges on the rebound of Samsung Electronics' semiconductor business this year. However, there are observations that the partnership with Broadcom, which is rapidly emerging as a rival to Nvidia, may unfold in a different situation from Nvidia's supply.
◇ Nvidia's picky preferences… "Broadcom is different"
According to the industry on the 15th, Samsung Electronics' HBM3E (5th generation HBM) was expected to begin supplying Nvidia last year, but ultimately extended into this year. Although expectations continue that supply will occur soon this year as well, Huang's comments reaffirmed that there are still significant barriers to overcome.
The issue is that it is unclear whether the HBM design problems pointed out by Huang CEO are due to Samsung Electronics' lack of capabilities or if Nvidia is demanding excessive specifications. A semiconductor engineer from Broadcom explained, "The HBM3E products currently supplied by SK Hynix to Nvidia greatly exceed the specifications established by the Joint Electron Device Engineering Council (JEDEC) for HBM3E products," adding, "This is because the power efficiency, data processing speed, and capacity demanded by Nvidia are becoming increasingly demanding."
However, analyses suggest that there is a significant likelihood that Samsung Electronics will supply HBM to Broadcom, which is rapidly emerging as a rival to Nvidia. The fabless (semiconductor design) company Broadcom is the world's largest custom semiconductor (ASIC) design firm. Recently, big tech companies like Google, Meta, and ByteDance have been entrusting AI chip development to reduce their dependence on Nvidia. Broadcom has previously designed Google's AI-specific chip, Tensor Processing Unit (TPU), and Meta's own AI chip "MTIA."
He noted the differences in the business models and system-on-chip (SoC) design capabilities of Nvidia and Broadcom, stating, "I believe there is a high possibility of a contract for HBM between Samsung Electronics and Broadcom," adding, "Considering the chip design processes and business models of Broadcom, unlike Nvidia, which has demanded 'over-specifications' from memory semiconductor companies, Broadcom will seek a company that can supply large quantities at reasonable prices tailored to cost-effectiveness."
He continued, "Broadcom's biggest strength is its unique SoC custom design capability, and particularly its outstanding interconnect technology capabilities needed for running large language models (LLM)," adding, "This is why U.S. big tech companies are entrusting AI semiconductor requests to Broadcom instead of Nvidia." He also noted, "Traditionally, they have designed custom ASICs that are not heavily reliant on memory, so they are likely to adopt a supplier that can provide large quantities at reasonable prices instead of over-specification for HBM."
◇ Broadcom's AI chip business intensifying this year… Will it lead the脱(Nvidia) trend?
The fact that Nvidia's flagship artificial intelligence (AI) accelerator "Blackwell" has faced defect issues since its early launch is likely to be a catalyst for Broadcom's growth in the custom AI semiconductor market. Analysts suggest that Broadcom's AI chips, which are less costly and more power-efficient than Nvidia's latest AI chips that require massive investments, will spur the脱(Nvidia) movement.
Hock Tan, CEO of Broadcom, stated in an interview with the Financial Times that "the AI investment boom from big tech companies is expected to continue until the end of the 2020s" and that "over the next 3 to 5 years, we will accelerate investment in AI infrastructure." Broadcom reported that it recorded $12.2 billion (approximately 17.79 trillion won) in sales in the AI chip design sector last year, a 220% increase from the previous year. Tan also noted, "We could generate billions of dollars in additional revenue from AI chip design by 2027."
Analyses indicate that Samsung Electronics may have an opportunity, given that SK Hynix has limited additional capacity for HBM supply to meet Nvidia's orders, its main customer. Research Institute analyst Kim Rok-ho noted, "In a situation where an increase in ASIC demand led by Broadcom is anticipated, Samsung Electronics' HBM production capacity is much greater than that of SK Hynix," adding, "There remains a possibility of an increase in HBM sales revenue."
However, there are still doubts about whether Broadcom will become an AI semiconductor buyer on the scale of Nvidia. This is because it is difficult to gauge the scale of Broadcom's HBM orders. Nvidia is a major player that secures around 90% of the global HBM supply, and this year it is expected to incorporate the highest amount of high-performance HBM into its new product, Blackwell.