Google unveils custom AI chips (TPU 8) to cut dependence on Nvidia and lower compute costs

Google announced two new Tensor Processing Units (TPU 8t and TPU 8i)—custom-designed chips made specifically for AI work—shipping later in 2025. Unlike most AI labs that buy expensive Nvidia chips, Google builds its own silicon in-house, avoiding what the industry calls 'the Nvidia tax' (the premium price Nvidia charges because so few suppliers exist).

The TPU 8t handles training (teaching an AI model on raw data), while the TPU 8i handles inference (the live step when an AI answers your question). By splitting these two very different tasks onto different chips, Google can make each one more efficient—meaning the company runs more AI models on the same amount of electricity and hardware investment.

For anyone using Google's AI services (like Gemini or enterprise AI tools), this matters because lower hardware costs let Google either improve service quality without raising prices, or invest more in model research. For businesses buying AI compute from cloud providers, this proves an alternative to Nvidia dominance exists—potentially breaking Nvidia's stranglehold on pricing and forcing more competition in the AI chip market.