Google Splits Eighth-Generation TPU Into Two Chips
Google's new TPU 8t and TPU 8i are the first Tensor Processing Units from the company built as distinct designs for training and inference.
Updated on Nov. 9, 2026 in Semiconductors
Google unveiled its eighth-generation Tensor Processing Units on April 22, 2026, at Google Cloud Next, as two separate chips: the TPU 8t for large-scale pre-training and the TPU 8i for inference and high-concurrency reasoning. Both are manufactured on TSMC's N3 process and integrate Google's Arm-based Axion CPUs.
Why it matters
Google framed the split around the agentic era, in which AI systems juggle many tasks at once, and it marks the first time the company has divided a TPU generation into distinct designs. The move reflects growing demand for chips tuned separately for training and serving AI models.
The TPU 8t pairs 216 GB of High Bandwidth Memory with 128 MB of Static Random-Access Memory and delivers 12.6 FP4 PFLOPs per chip, while the TPU 8i carries 288 GB of High Bandwidth Memory, 384 MB of Static Random-Access Memory and 10.1 FP4 PFLOPs per chip. These are manufacturer-stated specifications, and their real-world performance under production workloads has not yet been independently verified.
The players
Google is the Alphabet subsidiary that designs the Tensor Processing Units and operates Google Cloud, where the chips are offered to customers.
The details
The TPU 8i uses a new Boardfly topology and includes a CAE engine, and both chips slot into Google's AI Hypercomputer architecture. Google plans to share more technical details at Hot Chips 2026 in August.
Timeline
2025: Ironwood, the seventh-generation TPU, launched.
April 22, 2026: Google unveiled the eighth-generation TPUs at Cloud Next.
August 2026: More technical details are expected at Hot Chips 2026.
Later in 2026: General availability of the new chips is expected.
The Tech Race
The TPU 8i's claimed 80% performance-per-dollar gain over Ironwood continues Google's effort to lower the cost of serving AI models against rival custom accelerators. Splitting the line into training and inference chips reflects an industry-wide push to specialize hardware as AI workloads diversify.
Developers and businesses using Google Cloud may see new options for running AI training and inference workloads once the chips reach general availability later in 2026. The claimed performance-per-dollar gain for the TPU 8i could affect what it costs to deploy AI services, though those figures have not yet been independently tested.
The takeaway
The split into two chips shows Google is treating AI training and inference as distinct engineering problems, and buyers should watch for independent benchmarks before accepting the claimed efficiency gains.
What happens next
Google plans to share more technical details at Hot Chips 2026 in August, and general availability of the new chips is expected later in 2026.
Further reading
For related chip industry coverage, see Semiconductors.
