Google is intensifying its push into custom silicon as the tech giant works to reduce its reliance on third-party chips for its cloud computing infrastructure. The company on Monday announced two new additions to its tensor processing unit (TPU) family - specialized chips developed in-house for artificial intelligence workloads - expanding its chip portfolio to address specific use cases including confidential inference and increased memory capacity for demanding AI tasks.

The two new chips, designated the TPU 8tx and the TPU 8i Max, are extensions of Google's eighth generation of TPUs, which were first introduced in April of this year. That original lineup was built to deliver greater inference capacity and support for large-scale model training requirements. The latest additions build on that foundation while targeting more specialized needs.

A Chip Built for Confidential AI Workloads

The TPU 8tx was developed in collaboration with AMD and is derived from the existing TPU 8t architecture. Its primary focus is on enabling confidential training and inference - in other words, running and developing AI models in a secure environment where sensitive information remains protected throughout the process. According to Google, the chip "builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use."

Beyond security, the TPU 8tx also incorporates a custom inferencing network that allows it to perform reasoning tasks simultaneously with training. This dual capability means the chip can be used to distill and train smaller models in real time. Google highlighted this as particularly valuable for organizations that need to train AI models on proprietary datasets without transferring that data across different devices or systems - a concern that is increasingly common in regulated industries such as healthcare, finance, and legal services.

Doubling Down on Inference Capacity

The TPU 8i Max takes a different approach. Rather than introducing an entirely new architecture, it effectively merges two of the inference-focused TPU 8i chips that Google launched earlier this year into a single unified package. The result is a chip that doubles the memory capacity, processing power, and memory bandwidth of its predecessor while maintaining a compact form factor.

The practical benefit of this design is that AI models running on the TPU 8i Max can process a significantly larger volume of information during a single query. This is especially relevant for large language models and other AI systems that require access to extensive context windows or large sets of data simultaneously. By consolidating the capabilities of two chips into one, Google is offering a more efficient path to scaling inference performance without requiring additional physical hardware slots or complex multi-chip configurations.

A Broader Industry Push Away from Nvidia Dependence

Google's chip announcements arrive at a time when the broader technology industry is actively working to diversify away from its heavy dependence on Nvidia, which currently dominates the market for AI-specific semiconductors. Major cloud providers including Microsoft and Amazon are each developing their own lines of custom AI chips, as are leading AI research organizations such as OpenAI and Anthropic. These efforts extend beyond hardware alone, with many companies also building out the supporting software ecosystems needed to make their chips viable alternatives to Nvidia's well-established platform.

For Google, the stakes are particularly high. As one of the largest operators of cloud infrastructure in the world, the company spends enormous sums on computing hardware. Developing its own chips allows Google to optimize performance for its specific workloads, reduce costs over time, and avoid the supply constraints and pricing pressures that come with depending on a single dominant supplier.

The announcement also comes at a moment of heightened scrutiny for Nvidia, whose chief executive Jensen Huang is facing personal controversy following recent public disclosures. The timing underscores how rapidly the competitive landscape in AI hardware is shifting, with multiple well-resourced players now racing to establish credible alternatives to Nvidia's products across both the training and inference segments of the market.

Google's TPU program has been one of the longest-running efforts in the industry to build custom AI silicon at scale. The company first introduced its tensor processing units nearly a decade ago, and each successive generation has brought meaningful improvements in performance and efficiency. With the TPU 8tx and TPU 8i Max, Google is signaling that it intends to address an increasingly diverse set of customer requirements - from enterprises handling sensitive data under strict regulatory constraints to researchers and developers pushing the boundaries of what large AI models can do in production environments.