As Google and AMD boost the competitiveness of artificial intelligence (AI) accelerators, the market landscape led by Nvidia is being reshaped. As the two companies crack the "Nvidia unipolar" structure, the high-bandwidth memory (HBM) market is also shifting into a structure led by multiple "big spenders."
According to the semiconductor industry on the 5th, Fubon Research under Taiwan-based brokerage Fubon securities recently said, based on a supply chain survey, that Google's tensor processing unit (TPU) volume in 2028 could outpace Nvidia's shipments of data center graphics processing units (GPUs). It projected 12 million to 15 million units for 9th-generation TPUs and 12.4 million units for Nvidia AI GPUs. AMD, cited as a "Nvidia rival," is also increasing the HBM capacity per GPU. With major AI corporations secured as customers, accelerator shipments are also expected to rise.
HBM is a memory that stacks multiple DRAMs vertically, places them close to AI accelerators, and rapidly supplies large volumes of data. If data is not delivered in time, computational bottlenecks occur. With wide data lanes and high bandwidth, HBM has already become an essential component of high-performance AI accelerators. AMD's next-generation products include HBM4, the 6th-generation HBM, and global investment bank Morgan Stanley analyzed that Google's 9th-generation TPU could be equipped with HBM4E, the 7th-generation HBM.
◇ Nvidia 37.3% · Google 36.0%… next year's HBM demand split in two
Morgan Stanley projected total HBM demand for GPUs and application-specific integrated circuits (ASICs) next year at 48.618 billion Gb (gigabits). That is about 6.077 billion GB (gigabytes; 8 bits = 1 byte), the HBM volume that would go into roughly 31.65 million units of Nvidia's Blackwell B200, used as a mainstay accelerator for AI data centers. On a bit growth (total memory capacity growth rate) basis, demand is expected to rise about 60% from this year.
Nvidia is estimated to consume about 18.125 billion Gb, accounting for 37.3% of the total. Google is projected at about 17.507 billion Gb, or 36.0%, and AMD at about 5.9 billion Gb, or 12.1%. Nvidia will remain the largest buyer by edging out Google's demand by 1.3 percentage points, but other corporations' consumption could also rise quickly.
The HBM demand share of non-Nvidia customers such as Google, AMD, Amazon Web Services (AWS), Microsoft (MS), and Meta is expected to reach 62.7%. The HBM market, which had been swayed by Nvidia's product launch schedule and supplier selection, is effectively shifting to a multipolar system driven by multiple customers' product road maps.
Market researcher TrendForce also projected Nvidia's share of total HBM demand to drop to 58% this year from 66% last year. With cloud service providers (CSPs) such as Google TPUs increasing adoption of their own AI chips, ASIC-based products are expected to account for 27.8% of AI server shipments this year.
◇ Google expands HBM per TPU with each generation
Google, which has emerged as an HBM buyer on par with Nvidia, is increasing the HBM capacity per chip with each TPU generation. Ironwood, the 7th-generation TPU, carries 192GB of HBM per chip. The 8th-generation TPU 8t and 8i unveiled this year include 216GB and 288GB of HBM, respectively. Bandwidth is 6.5TB and 8.6TB per second, respectively, raising both capacity and throughput to rapidly exchange the computation results of large-scale AI models. The HBM capacity of the inference-oriented TPU 8i is 50% greater than Ironwood's. Morgan Stanley analyzed that HBM4E will be used in Google's upcoming 9th-generation TPU, with 576GB—double that of TPU 8i—potentially installed.
Google has also begun supplying TPUs, previously used in its own data centers and Google Cloud, directly to customers' data centers. Alphabet, Google's parent company, said in its second-quarter earnings release this year that it delivered TPU systems to customer data centers for the first time and began recognizing related revenue.
It has moved beyond the previous model of renting out TPU compute resources to establish a new revenue source. By expanding its business to directly install TPU systems in customer data centers, HBM demand is growing in tandem.
Global investment bank Barclays analyzed that Google's TPUs for external supply could increase from about 1 million units and 1.4GW (gigawatts) this year to about 4.74 million units and 11.5GW in 2028. It estimated 2028 revenue from the external TPU business at about $252.7 billion (about 361 trillion won).
◇ 31TB of HBM4 in one AMD Helios
AMD, which is securing competitiveness by expanding the HBM capacity per GPU beyond rivals, has also emerged as a new "big spender" in the memory market. The next-generation AI accelerator "Instinct MI455X" includes 432GB of HBM4. That is 50% more than the 288GB of HBM4 in Nvidia's Rubin GPU. A single rack-scale AI system "Helios" that bundles 72 MI455Xs carries a total of 31TB of HBM4. AMD has begun initial shipments of Helios and plans to ramp up supply in the fourth quarter. It has already secured major AI corporations as customers, including OpenAI, Meta, MS, and Anthropic.
As the position of the non-Nvidia camp, including Google and AMD, expands in the AI accelerator market, HBM supply competition among Samsung Electronics, SK hynix, and Micron is expected to heat up further. Previously, supply scale hinged on whether Nvidia's performance requirements were met; now, the key is the ability to deliver products tailored to each customer's accelerator characteristics on time. The industry says SK hynix still holds an edge in the Nvidia-oriented supply chain, but Samsung Electronics' influence is growing among non-Nvidia customers. Leveraging turnkey capabilities that provide DRAM, foundry, and advanced packaging together, along with a 10-nanometer-class 6th-generation (1c) DRAM process, Samsung Electronics has been raising its profile from the HBM4 transition period.
At last month's second-quarter earnings release, Samsung Electronics said, "HBM4 revenue in the third quarter of this year is expected to expand more than threefold from the previous quarter," adding, "We will secure an HBM market share on par with our overall DRAM share in the second half." According to market researcher Counterpoint Research, Samsung Electronics ranked No. 1 in the global DRAM market in the second quarter by revenue with a 39% share. Global investment bank UBS's projection of next year's HBM market shares by bit shipments—Samsung Electronics 41%, SK hynix 39%, Micron 20%—is interpreted as reflecting changes in supply to non-Nvidia customers.
A semiconductor industry official said, "As HBM buyers expand from Nvidia to Google and AMD, the importance of each customer's product road map and supply chain is growing," adding, "In the HBM4 and HBM4E transition, not only production capacity but also design, process, and packaging capabilities tailored to accelerators will determine market share."