Google broadens its AI hardware portfolio with two new specialized chips

Google announced on Monday two additions to its tensor processing unit lineup, the TPU 8tx and the TPU 8i Max. Both belong to the eighth generation of in-house AI chips that Google introduced in April and extend that family toward use cases the earlier parts did not target. The move fits Google's broader effort to lean less on outside chip suppliers for its cloud computing business.
TPU 8tx is aimed at confidential training and running of AI models. Google says the chip builds security into the hardware layer so that sensitive data and model weights stay protected and isolated even while actively in use. The company designed it with AMD, and it is based on the existing TPU 8t chip. It also carries a custom inferencing network that allows reasoning to run alongside training, which in practice means small models can be distilled and trained on the fly.
That combination points at a concrete scenario: a company training a model on proprietary data without shuttling that data between devices. Keeping weights and data inside a protected boundary during active computation is a meaningful distinction, because most encryption protects data at rest or in transit, not while a processor is reading it. Hardware-level isolation aims to close that gap, though the engineering cost usually shows up in performance and in how tightly software must be tied to the chip. How much throughput confidential mode gives up is not something the announcement addresses, and that is the number most buyers will eventually want.
The second chip, TPU 8i Max, is simpler in concept. It essentially packages two of the inference-focused TPU 8i chips launched in April into a single package, doubling memory capacity, processing power and bandwidth. The practical effect is that a model can hold and work through a wider range of information during queries, which matters for workloads where context length and working memory, rather than raw compute, are the bottleneck.
This suggests Google is segmenting its silicon more finely than in past generations. Rather than one general-purpose TPU per generation, the lineup now splits along trust boundaries (confidential versus standard) and along memory scale (single versus doubled inference parts). For developers, that means the choice of chip becomes a design decision rather than an infrastructure detail: teams handling regulated or proprietary data may find confidential training changes what they can do in the cloud at all, while teams running long-context inference may get more headroom without redesigning their serving stack.
The launch lands in a crowded field. Microsoft, Amazon and AI labs including OpenAI and Anthropic are all developing their own AI-specific semiconductors and the software that goes with them, largely to reduce dependence on Nvidia, which has dominated the market for AI accelerators. Google's TPUs are among the longest-running in-house efforts, having gone through eight generations, and the April parts already emphasized inference capacity and support for large training jobs. The new variants push those two threads further apart.
A likely trade-off is ecosystem lock-in. Custom silicon is most useful when the software stack is tuned for it, and Google's chips are tied to its own tooling and cloud. Companies that want portability across providers may find the confidential and high-memory capabilities harder to replicate elsewhere, which is exactly the kind of differentiation a cloud provider wants. It also raises a question for product teams about where sensitive training data can legally and practically live.
Why it matters: Teams that train on sensitive or regulated data and teams serving long-context inference get new hardware options inside Google's cloud, and the confidential-training part in particular could change which workloads are viable there at all. It also sharpens the competitive pressure on Nvidia and on other hyperscalers building their own accelerators. Whether these chips deliver on the security and memory claims in practice, and at what performance cost, is the question developers should press on before committing workloads.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Google's new tensor processing units (TPUs) broaden its chip portfolio to more use cases: confidential model training and more memory capacity.