![Google launched eighth‑generation TPUs at the same price as the previous generation, promising 2.8× training performance and 80% better inference. Google announced a dedicated training chip and a separate inference chip, each slated for availability later this year, and highlighted a 384 MB SRAM memory on the TPU 8i inference chip, triple the prior generation's capacity, per [CNBC] reported on 2026-04-22. ##Amin Vahdat on Specialized Chips## "With the rise of AI agents, we determined the community would benefit from chips individually specialized to the needs of training and serving," said Amin Vahdat, senior vice president and chief technologist for AI and infrastructure, underscoring the strategic split between training and inference workloads. ##Google vs. Nvidia: Memory and Throughput Compared## Google’s TPU 8i inference chip matches Nvidia’s focus on large memory for rapid responses, while the training chip delivers 2.8× the performance of its predecessor at identical pricing, though Nvidia’s exact figures remain undisclosed; Citadel Securities and all 17 U.S. Energy Department national laboratories are already leveraging the new silicon, with Anthropic committing gigawatts of TPU capacity. ##Amazon and Meta Parallel Custom Chip Strategies## Amazon previously introduced separate training and inference chips in 2018 and 2020, and Meta is collaborating with Broadcom on multiple AI processor versions, showing a broader industry move toward specialized AI silicon. Google expects both TPU chips to be available later this year, expanding the hardware portfolio for Google Cloud customers and reinforcing its AI infrastructure roadmap.](https://krolmarc.com/wp-content/uploads/2026/04/1776866436501-019db57d-8ce2-7209-a38e-c3cdc0f33db2.png)
Google launched eighth‑generation TPUs at the same price as the previous generation, promising 2.8× training performance and 80% better inference.
Google announced a dedicated training chip and a separate inference chip, each slated for availability later this year, and highlighted a 384 MB SRAM memory on the TPU 8i inference chip, triple the prior generation’s capacity, per CNBC reported on 2026-04-22.
Amin Vahdat on Specialized Chips
“With the rise of AI agents, we determined the community would benefit from chips individually specialized to the needs of training and serving,” said Amin Vahdat, senior vice president and chief technologist for AI and infrastructure, underscoring the strategic split between training and inference workloads.
Google vs. Nvidia: Memory and Throughput Compared
Google’s TPU 8i inference chip matches Nvidia’s focus on large memory for rapid responses, while the training chip delivers 2.8× the performance of its predecessor at identical pricing, though Nvidia’s exact figures remain undisclosed; Citadel Securities and all 17 U.S. Energy Department national laboratories are already leveraging the new silicon, with Anthropic committing gigawatts of TPU capacity.
Amazon and Meta Parallel Custom Chip Strategies
Amazon previously introduced separate training and inference chips in 2018 and 2020, and Meta is collaborating with Broadcom on multiple AI processor versions, showing a broader industry move toward specialized AI silicon.
Google expects both TPU chips to be available later this year, expanding the hardware portfolio for Google Cloud customers and reinforcing its AI infrastructure roadmap.