Tech

Google to Split TPU into Training and Inference Chips, Aims at Nvidia

Google reportedly splits TPU capabilities into separate training and inference processors as part of an eighth-generation TPU lineup, with both processors expected later this year. CNBC cites anonymous sources and Google has not publicly confirmed the details, but the move signals a broader shift toward workload-optimized silicon that could reshape data-center demand.

📅 April 22, 2026
⏱️ 2 min read
Google to Split TPU into Training and Inference Chips, Aims at Nvidia

Key Takeaways

  • Dual-chip TPU design could pressure Nvidia’s pricing power if validated at scale.
  • Reported 2.8x training performance versus seventh-generation Ironwood TPU is unverified.
  • Inference performance reportedly up ~80% versus prior generation, according to sources.
  • Each TPU 8i and training chip reportedly include 384 MB SRAM, with triple Ironwood SRAM

People Involved

  • No specific individuals mentioned

Entities Involved

  • Alphabet Inc. (Google) Developer of TPUs and AI hardware for cloud services
  • NVIDIA Dominant AI accelerator supplier; market context for chip competition
  • Broadcom Inc. Expanded chip deal with Google (per reports)
  • Amazon.com, Inc. Pursuing custom silicon for AI workloads
  • Anthropic Using Google TPUs at large scale (gigawatt-scale claim)
  • Citadel Securities Uses Google TPUs for quantitative research
  • Groq, Inc. Market context for LPU hardware referenced in CNBC report
  • U.S. Department of Energy national laboratories 17 labs reportedly using AI co-scientist software on Google TPUs

MarketMoodz Analysis

Investors should view this as a potential shift in cloud compute economics. If Google’s dual-chip approach proves cost-effective at scale, it could erode Nvidia’s pricing power and alter data-center demand patterns, potentially benefiting Alphabet’s cloud traction while introducing new competitive risk for Nvidia.

The story fits into a broader industry arc: hyperscalers are pursuing custom AI silicon, and a multi-hundred-billion-dollar AI silicon market is widely cited. Google has historically built and rented TPUs since 2018 to challenge Nvidia’s dominance—a dynamic that could accelerate if the eighth-generation chips deliver on the claimed performance improvements.

What to watch next: independent benchmarking for training and inference, any official confirmation from Google, and how cloud providers price and package TPU-based services as workload-specific silicon becomes more prevalent.

Source: Original Article

Get AI-Powered Market Insights

Stay ahead of market-moving events with our real-time analysis and stock ratings.

Start Your Free Trial