Google Unveils New AI Chip “Ironwood,” Challenging NVIDIA’s Dominance

7th-Generation Inference TPU “Ironwood” Introduced

Google, the world’s largest search company, has unveiled its newly designed AI inference chip, Ironwood, challenging NVIDIA’s dominance in the AI semiconductor market.

The company announced on November 6 (local time) that its 7th-generation Tensor Processing Unit (TPU) “Ironwood” will be available to the public within the next few weeks.

According to Google, Ironwood delivers four times the performance of last year’s 6th-generation Trillium and up to ten times the performance of the 5th-generation TPU from 2023.

The system can connect up to 9,216 Ironwood chips, significantly reducing large-scale data bottlenecks and enabling more efficient processing.

Optimized for Inference, Reinforcement Learning, and Low-Latency AI

Ironwood is designed for large-scale model training, reinforcement learning, and high-volume, low-latency AI inference workloads.

As the name “Tensor Processing Unit” implies, it features an architecture optimized for matrix (tensor) computations, allowing it to outperform general-purpose GPUs and NPUs in specific AI tasks.

Google emphasized that although NVIDIA GPUs have superior versatility, Ironwood offers greater competitiveness in price, performance, and power efficiency for specialized tensor operations.

Adopted by Major AI Startups Including Anthropic

Google stated that several major customers have already shown positive responses to Ironwood-based services.

Anthropic, the developer of the AI chatbot Claude, has secured access to up to one million TPUs. Other AI startups such as Lightricks (focused on image and video generation) and Essential AI are also adopting Ironwood for their operations.

“Rising Demand for AI Infrastructure”—Google Aims to Reduce NVIDIA Dependence

Sundar Pichai, Google’s CEO, recently remarked in a post-earnings conference call that “demand for AI infrastructure based on both TPUs and GPUs is extremely high.” This statement suggests that Ironwood could become a key alternative solution in meeting that demand.

The new TPU was first previewed as a test model in April at Google Cloud Next 2025 in Las Vegas, Nevada.

With this official release, Google is expected to accelerate efforts to reduce its reliance on NVIDIA GPUs, which remain scarce and costly due to global supply constraints.

By Choi Song-ah ㅣneria97@hanmail.net