Google has unveiled what it called its highest-performing AI-only chips yet at its annual tech conference, taking direct aim at Nvidia’s dominance. Analysts say that as the AI accelerator market diversifies beyond Nvidia, Samsung Electronics and SK hynix stand to gain because they supply the key memory used in those chips.
According to the IT industry on the 23rd, Google Cloud on April 22 (local time) introduced two eighth-generation Tensor Processing Units at “Google Cloud Next 2026” in Las Vegas: the training-focused TPU 8t and the inference-focused TPU 8i. Both are scheduled for official release within the year.
TPUs are AI-specific application chips (ASICs) that Google co-designed with Broadcom. Unlike general-purpose graphics processing units, they are optimized for AI workloads and are widely viewed as more power-efficient.
The key change in the eighth-generation TPU is a design that separates training and inference for the first time. It is the first major architectural shift since Google first launched the TPU in 2015.
The TPU 8t uses a “superpod” design that can link up to 9,600 chips in a single system. Google said each pod delivers 121 exaflops of performance and trains models three times faster than the previous generation, with double the power efficiency. Through its Pathways platform, Google said more than 1 million TPUs can be pooled to operate like a single cluster, cutting development time for very large AI models from months to weeks.
The TPU 8i targets demand tied to agentic AI. Google said it delivers 9.8 times the performance of the seventh generation and includes 384MB of on-chip SRAM — more than triple the prior generation — along with 288GB of high-bandwidth memory (HBM).
Market share remains heavily tilted toward Nvidia. Google’s TPUs account for about 5% of the AI accelerator market, while Nvidia’s GPUs hold about 92%, according to the report. Nvidia’s next-generation “Vera Rubin” platform has already entered mass production and is set for commercial release in the second half of this year.
Still, conditions are shifting in Google’s favor as a prolonged shortage of Nvidia GPUs pushes customers to seek more cost-effective alternatives. Amazon Web Services is also pursuing a strategy to reduce reliance on Nvidia by promoting its own AI accelerators, including Trainium for training and Inferentia for inference. Hedge fund Citadel has built quantitative research software using Google TPUs, and 17 U.S. national laboratories under the Department of Energy are operating TPU-based “AI co-scientist” software, the report said.
The trend is drawing attention to South Korean chipmakers. Each TPU carries six to eight HBM stacks, and supply of HBM for Google’s TPUs has effectively consolidated around Samsung Electronics and SK hynix.
In Nvidia’s GPU supply chain, Samsung, SK hynix and Micron compete, but Micron has effectively fallen out of the TPU camp due to limited production capacity, the report said. The industry expects the eighth-generation TPU to use HBM4, a sixth-generation HBM standard, and forecasts Samsung’s HBM shipments to Google will more than double from this year. SK hynix, described as Google’s preferred supplier, is currently supplying HBM3E and is also reported to be the exclusive supplier of 12-high HBM3E for the power-efficiency-improved TPU 7e.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.