Google is expanding its internal chip lineup with two new specialized variants of its Tensor Processing Units (TPUs), aiming to strengthen its infrastructure for AI workloads. The latest additions—TPU 8tx and TPU 8i Max—are tailored to address secure inference and high-capacity memory needs, helping the company move further away from reliance on external chipmakers in its cloud operations.
TPU 8tx: Built-in Security for Sensitive AI Tasks
The TPU 8tx is a derivative of Google’s earlier TPU 8t, designed in collaboration with AMD. Its standout feature is hardware-level security that keeps model parameters and sensitive data protected even while in active use. This enables confidential training and inference, ensuring that proprietary data doesn’t need to be transferred between devices.
Additionally, TPU 8tx includes a custom network capable of both training and inference simultaneously. This allows it to distill and train smaller models on the fly—useful for scenarios where data privacy is paramount or regulatory requirements demand rigorous isolation.
TPU 8i Max: Doubling Up for Larger AI Models
The TPU 8i Max combines two instances of the inference-focused TPU 8i in a single package. The configuration doubles memory, processing power, and bandwidth compared to a single 8i, enabling handling of richer information during AI queries. It’s aimed at workloads that require more capacity during model inference rather than the training phases.
Both these units build on Google’s eighth generation of TPUs, which debuted in April. The newer variants are tailored to newer use cases, like training models on proprietary data without moving it externally, and supporting larger inference-based models. The goal is to offer infrastructure choices that match increasing diversity in customer AI needs.
This expansion comes amid growing efforts by hyperscale tech firms—including Microsoft, Amazon, OpenAI, and Anthropic—to design their own AI accelerators. These companies are increasingly creating custom chips and accompanying software stacks in an effort to reduce dependency on Nvidia-dominated supply chains.
The timing is notable. Google’s announcement arrives as Nvidia faces scrutiny over its leadership. While unrelated to chip design, recent non-technical controversies have underscored how leadership vulnerabilities can become focal points in a competitive market.
What this means: By introducing TPU 8tx and TPU 8i Max, Google is carving out more differentiated offerings in the AI hardware landscape. Enterprises needing secure inference or large memory capacities now have more tailored options. For industry watchers, it signals the intensifying chip wars and the shift toward custom silicon tuned for particular AI tasks. Going forward, success will depend not just on raw specs, but ecosystem support—how well these chips integrate into cloud stacks, security frameworks, and developer tools.
