Google expands TPU sales to create new markets for Korean memory makers
Google is selling its custom AI chips to external customers, opening a massive new demand channel for high-bandwidth memory beyond Nvidia's dominance.
The short version
- Google now serves nearly 100 TPU customers, including those deploying the chips in their own private data centers.
- Counterpoint Research predicts HBM bit demand for custom AI server chips will rise 35-fold between 2024 and 2028.
- Samsung is partnering with Broadcom to secure HBM and foundry manufacturing deals potentially valued at over $200 billion through 2030.
- SK hynix is proposing to jointly design customized HBM for future TPUs to secure an earlier role in the chip development cycle.
Why it matters
This shift signals a move toward a more diverse AI hardware ecosystem where Google's custom silicon competes directly with Nvidia. For the reader, this means a broader variety of infrastructure options for AI development and a more stable supply chain for the memory components that power it.
Google's shift to a chip business
Google is moving its Tensor Processing Units (TPUs) from internal use to a broader commercial model. While these custom AI chips were originally designed to power Google's own services, the company is now actively selling them to external cloud providers and companies building AI infrastructure.
The company currently serves nearly 100 TPU customers who are deploying these processors in their own data centers. This expansion is a strategic move to turn the TPU into a standalone business that serves clients outside of Google's own ecosystem.
Alphabet expects TPU system sales to grow substantially following their contribution to Google Cloud's third-quarter growth. To support this expansion, Google recently unveiled two new systems, the TPU 8i Max and the TPU 8tx, as part of its eighth-generation lineup. The 8i Max combines two chips into one package to double its computing power and HBM capacity, while the 8tx is built for AI tasks that need higher levels of data protection.
New demand for Korean memory
The expansion of Google's TPU business creates a significant opportunity for South Korean memory manufacturers Samsung Electronics and SK hynix. These companies provide high-bandwidth memory (HBM), which is a type of memory designed to handle the massive data speeds required by AI processors.
Currently, the market for these components is heavily dominated by Nvidia's graphics processing unit (GPU) ecosystem. By selling TPUs to a wider customer base, Google provides a new avenue for Korean suppliers to secure demand for HBM beyond Nvidia's hardware.
Counterpoint Research forecasts that the demand for HBM bits used in custom AI server chips will increase by 35 times between 2024 and 2028. This growth suggests that the move toward customized AI chips will become a primary driver for the memory industry in the coming years. Industry observers note that the most important aspect of this news is not the specific technical specs of the chips, but rather Google's transition into a business that sells these chips to outside parties.
Strategic partnerships and designs
Samsung is pursuing its share of this new market through a memorandum of understanding with Broadcom, a major development partner for Google's TPUs. The agreement covers HBM, foundry manufacturing, and advanced packaging, with potential cooperation valued at more than USD 200 billion through 2030.
It is important to note that this large valuation encompasses various business areas and does not represent confirmed orders for HBM specifically. Despite this, the partnership highlights Samsung's intent to be a primary supplier for the infrastructure required to run Google's AI systems.
SK hynix is taking a different approach by proposing to jointly design customized HBM for future TPUs. By seeking to be involved earlier in the chip development process, SK hynix hopes to secure a more integral role in the hardware lifecycle rather than just acting as a component supplier. Google may also change its procurement strategy to benefit these suppliers further. By working with Taiwan's MediaTek on the TPU 8t, Google could move away from a model where a design partner handles all memory purchases, potentially allowing for more direct and cost-effective deals with memory makers.
Challenges in securing orders
Despite the massive growth potential, converting these new opportunities into actual orders remains a difficult task for memory manufacturers. Production capacity is already heavily committed to existing contracts, making it hard to pivot quickly to new demands.
SK hynix reported that agreements with 14 major customers already cover more than half of its projected production capacity for 2027 and 2028. This high level of commitment means that manufacturers must be proactive in securing long-term supply contracts early.
Because capacity is so tight, early participation in future chip designs is becoming essential. For Samsung and SK hynix, the goal is to lock in long-term agreements now to ensure they can meet the projected 35-fold increase in demand for custom AI server components. This proactive strategy is necessary to navigate a market where the most lucrative contracts are often finalized years before the hardware actually hits the shelves.
What to watch
- Whether Oracle's plan to offer Google's chips through its cloud platform next year leads to a significant spike in TPU sales.
- The final terms of the Samsung and Broadcom cooperation and how much of that $200 billion valuation is dedicated specifically to HBM.
Try it
Look up the current market share of Nvidia GPUs versus custom AI chips like Google's TPUs to see how the hardware landscape is diversifying.
Sources and links
First reported by The Economic Times. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.