Google could produce 15 million TPU v9 chips

Google could produce 12 million to 15 million TPU v9 AI accelerators in 2028, according to a forecast from Fubon Research. The upper end of that estimate would put Google’s own chip output close to Nvidia’s projected data-center GPU shipments: Fubon expects Nvidia to deliver about 12.4 million devices that year.

Google has not confirmed the production plan, and it has not disclosed the specifications of TPU v9. The forecast therefore remains an analyst estimate rather than an announced product roadmap.

Four-die design and manufacturing demand

According to Tom’s Hardware, analysts expect TPU v9 to use four compute dies in a multi-die package. That design combines several chips into a single processor, rather than relying on one monolithic die.

Recommended reading

Microsoft plans one Copilot app by year-end

Fubon estimates that this configuration would more than double Google’s need for manufacturing capacity compared with 2027. Producing the accelerators at the projected scale would require advanced packaging, which connects multiple dies inside one package, along with high-speed interconnects between them.

TSMC capacity and Intel Foundry

Fubon also believes TSMC’s capacity may not be sufficient for the projected volume. As a result, Intel Foundry is being considered as a possible additional manufacturing partner.

That move would not be a simple production transfer. Intel’s EMIB packaging technology and TSMC’s CoWoS-L process differ substantially, so Google would need to adapt the chip design for each packaging approach. The company has not said whether it plans to use Intel Foundry, nor has it confirmed any TPU v9 production targets, release timing, or technical specifications.