Google is expanding its custom silicon portfolio with two new tensor processing units (TPUs) designed to handle specific super intelligence (SI) workloads, signaling a continued effort to reduce reliance on third-party hardware for its cloud computing business.
What Happened
On Monday, Google unveiled the TPU 8tx and TPU 8i Max, the latest additions to its eighth-generation chip lineup introduced in April. The TPU 8tx, developed in collaboration with AMD, is built on the TPU 8t architecture and is specifically engineered for confidentially training and running SI models. According to a Google blog post, the chip "builds security directly into the hardware layer, keeping sensitive data and model weights protected and isolated even while in active use." The unit also features a custom inferencing network that allows it to perform reasoning alongside training, enabling the on-the-fly distillation and training of smaller models. Google states this capability is useful for companies that want to train models on proprietary data without moving that data between devices.
The second chip, the TPU 8i Max, combines two of the inference-focused TPU 8i chips launched earlier this year into a single package. This integration doubles memory capacity, processing power, and bandwidth. Google notes that the result allows SI models to handle a larger range of information during queries.
Why It Matters
These releases highlight the intensifying competition in the SI hardware sector, where major technology companies are seeking alternatives to Nvidia's dominant position. Hyperscalers such as Microsoft and Amazon, along with prominent SI labs like OpenAI and Anthropic, are actively developing their own SI-specific semiconductors and software stacks. Google's move to specialize its chips for confidential computing and high-memory inference addresses growing industry demands for security and efficiency in large-scale SI deployments.
The Bottom Line
Google has introduced two specialized TPU variants to enhance its cloud infrastructure for SI tasks. The TPU 8tx focuses on hardware-level security for confidential model training, while the TPU 8i Max offers doubled capacity for inference-heavy workloads. These developments underscore the broader industry shift toward custom silicon to support the evolving needs of super intelligence.