Google adds TPU 8tx and TPU 8i Max to cut reliance on third-party AI chips

Google has added two new tensor processing units to its in-house silicon lineup, one aimed at confidential AI workloads and the other at memory-hungry inference. According to TechCrunch, the company unveiled the chips, TPU 8tx and TPU 8i Max, on Monday as part of a broader push to depend less on other chipmakers for its cloud computing business.

Two new members of the eighth TPU generation

Both chips extend the eighth generation of TPUs that Google introduced in April, which had focused on delivering more inference capacity and support for large training runs. The new variants push the family into more specialised territory.

TPU 8tx, developed with AMD, builds on the existing TPU 8t design and is intended for confidentially training and running AI models. In a blog post cited by TechCrunch, Google said the chip embeds security protections in the hardware itself, so that sensitive data and model weights remain shielded and walled off even while the chip is actively using them.

The chip also includes a custom inferencing network that lets it run reasoning workloads alongside training. According to Google, this makes it possible to distill and train small models on the fly, which could appeal to organisations that want to train models on proprietary data without moving that data between devices.

TPU 8i Max takes a different approach: it essentially combines two of the inference-focused TPU 8i chips launched in April into a single package. TechCrunch reports that the result doubles memory capacity, processing power and bandwidth, allowing AI models to work with more information in a single query, a growing requirement as context windows and reasoning-heavy workloads expand.

The race away from Nvidia

The launch fits a wider industry pattern. As TechCrunch notes, Microsoft, Amazon and AI labs including OpenAI and Anthropic are all working on their own AI-specific semiconductors and accompanying software, largely to loosen Nvidia's grip on the market for AI compute.

For Google, the calculus is both economic and strategic. Custom TPUs can be tuned to the exact workloads Google Cloud sells, and every job that runs on in-house silicon is one that does not hinge on a third-party supplier's pricing, allocation or release schedule.

The timing also lands awkwardly for Nvidia: TechCrunch reports that the announcement comes just after the company's chief executive, Jensen Huang, admitted over the weekend to an affair with someone alleged to be a Chinese spy.

Why it matters

Hardware-level confidential computing could become a decisive factor for regulated customers in healthcare, finance and government that have so far hesitated to run sensitive models on shared cloud infrastructure. The AMD partnership is also notable, showing a traditional chipmaker teaming up with a hyperscaler on a custom accelerator rather than competing head-on. Most importantly, a wider TPU portfolio means Google Cloud can now serve training, inference, privacy-sensitive and long-context workloads without depending on Nvidia hardware, deepening the competitive pressure that in-house silicon from every major cloud provider and AI lab is placing on the market leader.