Google to produce 12-15 million TPU v9 chips by 2028, potentially exceeding Nvidia's AI GPU output

Google's Ambitious TPU Production Targets

According to a Fubon Research note, Google intends to produce between 12 and 15 million units of its TPU v9 processors by 2028. This represents a significant increase in Google's custom AI accelerator deployment, building on its decade-long history of developing its own AI chips.

Potential to Outpace Nvidia

Fubon Research estimates that Nvidia will ship 12.4 million data center AI GPUs in 2028. If Google achieves its target of 12-15 million TPU v9 units, it could produce a comparable or even greater number of AI accelerators than Nvidia in that year. The TPU v9 will incorporate four compute dies, suggesting a focus on high performance.

Reliance on Intel Foundry for Volume

The projected production volume for Google's TPU v9 may exceed the capacity available from TSMC alone. Fubon Research suggests that Google will likely need to leverage Intel Foundry's manufacturing capabilities to meet its ambitious goals by 2028. Reports indicate that Intel has already secured orders to produce three million TPUs for Google, following extensive testing of Intel's advanced packaging technologies.

Engineering and Manufacturing Implications

Developing an AI accelerator with four large compute chiplets is a complex engineering undertaking. Google's commitment to this design for the TPU v9 highlights its investment in processor performance. The potential use of Intel Foundry also implies a strategic decision to integrate Intel's advanced packaging services, which are crucial when designing chiplets with specific packaging technologies in mind.