Google could be getting ready for one of the biggest production runs of AI chips ever attempted. According to a research note from Fubon Research, the company plans to have between 12 million and 15 million ninth generation Tensor Processing Units, or TPUs, in 2028.
Google has spent roughly a decade developing custom AI processors for its own data centres. Those chips first supported Google鈥檚 internal workloads before becoming a bigger element of its cloud business as artificial intelligence computing became more common.
The scale of the plan would mean Google wouldn鈥檛 be so far off from Nvidia in annual AI accelerator volume. Fubon estimates Nvidia supplied 8.2 million data centre AI GPUs during 2026 and expects total shipments to reach 12.4 million in 2028. If Google reaches between 12 million and 15 million TPUs during the same year, it could manufacture a similar volume of AI accelerators.
Fubon Research wrote, 鈥淏ased on our checks, Google plans to have 12 鈥� 15 million TPUs in 2028. Entering 2028, Google鈥檚 TPUs will enter the V9 generation with four compute dies, suggesting that their capacity consumption will more than double in 2028 versus 2027.鈥�
听
How Could Google Make So Many Chips?
听
Google鈥檚 production plan starts with the TPU v9 itself. According to Fubon Research, the processor will use four compute dies instead of a single piece of silicon.
This design puts multiple chiplets inside one processor, but manufacturing becomes more complicated. Every compute die must be connected using advanced packaging technology before the finished processor can be installed inside Google鈥檚 AI systems.
Producing millions of processors each year depends on manufacturing capacity as well as chip design. Making one advanced processor is difficult. Producing millions every year needs a massive manufacturing operation.
听
More from News
- TikTok Parent ByteDance Has Officially Passed The $4 Billion Mark In AI Revenue
- Russia鈥檚 Latest Move Against Pavel Durov Shows That Telegram Is No Longer Just A Messaging App
- Experts React To The UK鈥檚 Decision To Make Tech Subjects Compulsory In Schools
- Microsoft Has Confirmed Copilot鈥檚 Super App Will Launch Soon 鈥� But What Is It For?
- G2A.COM鈥檚 Autonomous AI Agent Dave Helps Sellers Resolve 14,400 Support Tickets In 63 Days
- Why One MedTech Company Chose a Computer Graphics Conference To Launch Its Next AI Platform
- 75% Of CEOs Don鈥檛 Think Marketing Drives Growth 鈥� What Are They Missing?
- OpenAI Will Soon Release Its First Tech Gadgets 鈥� Here鈥檚 What To Expect
听
Fubon believes Google鈥檚 production capacity for TPU v9 could more than double in 2028 compared with 2027 because of the four die design. Advanced packaging becomes one of the biggest engineering requirements during production.
听
Who Will Manufacture The Chips?
听
Fubon does not believe TSMC can complete Google鈥檚 full production target on its own.
Intel Foundry could become an important manufacturing partner, especially since reports published during recent months said Intel secured orders to package more than three million Google TPUs after Google evaluated Intel鈥檚 advanced packaging technology.
Fubon Research wrote, 鈥淎lthough we do not have the detailed allocation yet, we think it is difficult to reach Google鈥檚 target with TSMC alone, and Intel鈥檚 supply is a must by 2028.鈥�
There is a technical reason for this. Modern AI processors rely on specialised packaging systems that connect multiple chiplets into one processor. Intel uses EMIB and EMIB T packaging, whereas TSMC uses CoWoS L. Engineers must design processors for one packaging system or the other because the technologies are not compatible.
听
What Could This Mean For AI Computing?
听
If Google reaches between 12 million and 15 million TPUs, Fubon Research believes the company could become the world鈥檚 biggest user of AI accelerators. Google would also gain greater control over the computing hardware powering products such as Gemini and Google Cloud.
Google is not expected to stop buying Nvidia processors and producing millions of its own chips gives the company extra computing capacity for artificial intelligence services without depending entirely on one supplier.
