Google unveiled two new Tensor Processing Units (TPUs) built for inference and training of next-gen AI agents. The chip market is now facing steeper competition as Nvidia GPUs are not the only ones suitable for emerging MoE architectures.
The tech giant unveiled its 8th-generation TPUs on April 22. TPU 8t and TPU 8i are two new processing models, each focused on a different aspect of agentic AI development. They are both targeting not just popular existing large language models (LLMs) but also massive Mixture-of-Experts (MoEs) and reasoning-heavy architectures that only start gaining traction.
TPU 8t is an engine for creating new AI models from scratch. Initial training is extremely expensive and happens on massive clusters. TPU 8t is designed to make that process cheaper and smoother at a very large scale. Google promises “2.7× better performance per dollar” for this TPU type. That means, developers would be able to train huge AI models for almost three times less money, or allocate the same budget and gain the opportunity to train a much better model with the same resources.
The second part of the agentic AI training process offered by Google is TPU 8i. This TPU is good for running AI models after they’re created and trained. Modern AI systems, especially newer designs called “Mixture-of-Experts,” don’t use their whole capacity at once. They activate only small parts depending on the task. Therefore, TPU 8i is optimized for exactly that kind of behavior and avoids wasting compute power. This model offers “80% better performance per dollar,” so that serving AI responses becomes much cheaper, and AI model owners can handle far more users without increasing costs.
Google’s TPUs are still at the announcement stage, but they are already highly expected, since the processing unit market is not so densely saturated at the moment. Nvidia GPUs are still the most flexible and easiest variants to use. Developers can plug them into almost any AI project, and they are pretty universal and efficient. But Google is now making a solid bet with its specialized next-gen hardware like TPU 8i and 8t. One of its main benefits is cost-efficiency without infringing on the product quality.
This factor may prove to be key, as we speak of billions of USD poured into AI infrastructure each year. The combined capital expenditures of Meta, Alphabet (Google’s parent), Amazon, and Microsoft have soared to more than $400 billion in 2025 and are projected to push well beyond $600 or even $650 billion this year, led by massive investments in AI data centres and cloud infrastructure.
