Google plans to ramp up AI chip TPU production to 12-15 million per year within a couple of years, catching up with Nvidia

Google plans to ramp up production of its own AI chips, aiming to reach Nvidia's level. According to unofficial information, in 2028 Google expects to deploy from 12 million to 15 million of its newest TPU v9 accelerators. If confirmed, these numbers would match Nvidia: analysts estimate that Nvidia will ship 8.2 million data center GPUs in 2026, and that figure could rise to 12.4 million by 2028. Thus, a single cloud provider could produce AI accelerators in volumes comparable to the leading commercial chip supplier. The ninth-generation TPU, debuting in 2028, will use four compute chiplets, following the industry trend toward chiplet designs; the planned production volume is expected to more than double compared to 2027 levels. Placing multiple large chiplets on one chip requires advanced interconnect and packaging technologies, and large-scale production complicates the task. Manufacturing capacity could become a limiting factor: TSMC alone may not handle such an order, so Intel's manufacturing division might need to be involved. Packaging technologies differ between manufacturers, and Intel's EMIB or EMIB-T are not directly compatible with CoWoS-L. Google has been developing its own chips for about a decade, and their role in infrastructure is growing: initially a way to support workloads, now it's a broad cloud business strategy. If Google reaches its goal in 2028, it could become the largest user of AI accelerators. This doesn't mean it will stop buying from Nvidia, but Google would have significantly more control over its own compute infrastructure and supply. No comparative benchmarks of Google TPU v9 against Nvidia Rubin and Rubin Ultra are available yet. The scale of the search giant's plans indicates that competition in the AI hardware market increasingly depends on deployment volumes, not just chip performance.