Google says Gemini is helping design its own AI chips faster

Google shipped four distinct Tensor Processing Unit systems in a single year. According to the company, that has never happened before.

Part of the credit goes to an unusual co-worker: Gemini, Google’s own AI model. It is reportedly cutting chip layout work from weeks down to hours. The AI is now helping build the hardware that runs the AI.

Sundar Pichai highlighted the milestone in a post on X, crediting a joint effort between Google’s hardware teams and DeepMind.

“Google’s hardware teams and DeepMind co-designed four distinct TPU systems in one year.”

Two new chips join the eighth generation

On November 9, 2026, Google expanded its eighth-generation TPU lineup with two new models: the TPU 8tx and the TPU 8i Max.

TPUs are Google’s custom processors built for machine learning.

The TPU 8tx targets secure and confidential inference. Inference is the stage where a trained model actually answers questions or performs tasks. “Confidential” here means the system is designed to keep the data being processed protected while it runs.

The TPU 8i Max goes in a different direction: raw capacity. It packs double the SRAM and HBM density into the same rack footprint as its predecessor configuration.

SRAM and HBM are two types of high-speed memory that sit close to the processor. More of it means a model can hold more information in working memory at once. Google says the payoff is larger context windows, which is the amount of text, code or data a model can consider in a single pass.

How four systems landed in one year

The two new chips follow the TPU 8t and TPU 8i, which Google unveiled at its Google Cloud Next event in April. Together, those four make up the full eighth-generation family shipped this year.

The earlier pair already set a high bar. Google said the TPU 8t and 8i delivered up to 2.7 times better training performance compared to the previous generation.

Training is the expensive, compute-heavy phase where a model learns from data. Inference is the ongoing cost of putting that model to work. The eighth-generation lineup now covers both: systems tailored for high-throughput inference and others built for scalable training.

Google frames the whole family around what it calls the agentic AI era. That refers to AI systems that carry out multi-step tasks on their own rather than simply responding to one prompt at a time.

The more interesting detail is how Google sped up the process. Chip layout, the painstaking work of arranging components on silicon, has traditionally been a slow, specialist job. Google says Gemini has compressed that step from weeks to hours.

Why it matters beyond Google’s own data centers

These chips are not just for internal use. Google says the new TPU systems support external Google Cloud customers, and the company is promoting open software stacks alongside its own workloads.

For enterprise buyers, the product split offers a clearer menu. Companies with privacy constraints get a confidentiality-focused option in the 8tx. Companies pushing the limits of long-context models get a memory-dense option in the 8i Max. Teams focused on training have the 8t and 8i.

Google has not tied the new chips to specific benchmark claims beyond the 2.7 times training figure for the earlier eighth-generation pair.