Google’s latest Tensor processors are taking the environmentally friendly route, but at what cost?

At last year’s Google Cloud Next event, Google announced a class of Ironwood tensor processing units (TPUs) that power data centers. Designed for the AI ​​era, these TPUs focused on large-scale output, or the ability to make inferences or predictions based on what the AI ​​was trained to do (essentially what chatbots do), but without actually knowing the answer already. This year, it made further improvements to the TPU hardware and now splits the computations to serve training and output differently.

At Cloud Next 2026, Google announced the eighth generation TPU with different architectures for different purposes. The newly introduced TPUs include the TPU 8t, which will be used to train AI models, and the TPU 8i, which is specific to inference-related tasks.

Google says the split was made to meet the different power and computing requirements of both processes. This approach will help its data centers reduce energy consumption, thereby reducing operating costs and reducing AI’s negative environmental impact. This means that Your use of Gemini It may soon take less water to keep data centers cooler (hopefully!).

Training neural networks involves high-throughput memory and large groups of processing units, as it requires updating billions of parameters every second. Training involves a process calledback propagation of errors,” involves countless feedback loops that test and optimize the neural network on the training set until it begins to remember accurate information.

At the same time, the result is less intensive and can be processed on less capable hardware with less memory consumption. Using the same equipment for training and inference therefore results in a higher actual cost, which in turn increases the effective cost for the inference tasks.

Google previously introduced TPU v5e (where the “e” is supposed to stand for efficiency) for smaller-scale operations. The latest TPU 8i appears to be a large-scale adaptation based on the previous hardware. Amazon is trying to achieve the same effect with AWS Inferentia.

Although Google mentions the environmental benefits of using special grounding TPUs, we haven’t seen any promises to cut costs. It remains to be seen whether Google will pass on some of the benefits to its consumers or keep the profits for itself and its corporate allies.