The shift: TPUs become a business for external customers
Google is pushing sales of its custom artificial intelligence chips beyond its own data centers, according to The Korea Herald. The American company is expanding its tensor processing unit (TPU) business to serve cloud providers and enterprises building AI infrastructure: a shift that, according to the source, could broaden the customer base for Korean high-bandwidth memory (HBM) manufacturers, which are currently heavily dependent on Nvidia’s GPU ecosystem.
The new element is not so much the technical specification of the chips as the business model. On Monday, Google presented two new TPU systems, called TPU 8i Max and TPU 8tx, bringing its eighth-generation line to four. The first combines two chips in a single package, doubling computing performance and HBM capacity; the second is intended for AI workloads that require greater data security. Google says it has nearly 100 TPU customers, including companies that install the processors in their own data centers, and Oracle plans to offer the chips through its cloud platform next year. Alphabet expects a substantial increase in TPU system sales after their contribution to Google Cloud’s third-quarter growth.
The moves by Samsung and SK hynix
Samsung is pursuing the opportunities through Broadcom, Google’s main partner in TPU development. In July, the two companies signed a memorandum of understanding on HBM, foundry and advanced packaging, with potential cooperation valued at more than $200 billion through 2030. The estimate, the source specifies, covers multiple activities and does not include confirmed HBM orders.
SK hynix, by contrast, aims to enter the chip design phase earlier. After a February meeting between group chairman Chey Tae-won and Google CEO Sundar Pichai, the company proposed jointly designing custom HBM for future TPUs. Chey also met Hock Tan, Broadcom’s CEO, to discuss next-generation AI chips. According to Counterpoint Research, Google is working with Taiwan’s MediaTek on the TPU 8t also to move away from Broadcom’s turnkey model, in which the design partner also handles memory purchasing; Counterpoint estimates that this model adds a 15-20% markup to HBM costs. Greater direct control over design and procurement could reduce costs and lead Google to negotiate more directly with memory suppliers.
Demand is growing, but the production slots are already allocated
Market prospects are expanding: Counterpoint forecasts that HBM demand in bits for custom server AI chips will grow 35-fold between 2024 and 2028, partly due to the expansion of Google’s TPUs. Custom HBM, with HBM4 and HBM4E, would allow suppliers to tailor components to specific processors and to target higher-value, longer-term contracts through early involvement in design. A Counterpoint executive calls it a major opportunity to create a high-value business.
A significant constraint remains, however: production capacity is already largely committed. SK hynix said in October that agreements with 14 major customers cover volumes equivalent to more than half of total capacity for 2027 and 2028, including products beyond HBM. As an industry source cited by the newspaper observes, winning a new customer does not necessarily mean shipping more HBM immediately: the real competition is over who enters the design of the next chip and secures production slots in time, turning the relationship into long-term orders. The actual volumes and timing of orders resulting from the TPU expansion therefore remain to be verified.
