According to the latest reports, Google is transitioning from an internal chip user to a chip retailer, directly challenging NVIDIA's market dominance. The existence of Google's latest TPUv7 "Tie Mu" processor has already impacted the price of AI computing power, causing a decline in market prices.

Previously, Google's tensor processing units (TPUs) were almost exclusively used for its own AI models, but the strategy has now changed. Research from the analysis firm SemiAnalysis shows that Google is actively selling its TPUs to third parties, aiming to compete with NVIDIA in the market. One of the new customers, Anthropic, has signed a deal for approximately one million TPUs, involving both direct hardware purchases and leasing through the Google Cloud platform.

OpenAI

Image source note: The image was generated by AI, and the image licensing service provider is Midjourney

This competitive effect is also evident in the market. According to SemiAnalysis's report, OpenAI successfully negotiated a 30% discount on its NVIDIA hardware by threatening to switch to TPUs or other alternatives. Analysts joked: "The more TPUs you buy, the more costs you save on NVIDIA GPUs."

Google's TPUs have proven they can support the training of top AI models. Recently launched powerful AI models, Google's Gemini 3 Pro and Anthropic's Claude 4.5 Opus, mainly rely on Google's TPU and Amazon's Trainium chips. TPUv7 is nearly comparable to NVIDIA's Blackwell series in terms of theoretical computing power and memory bandwidth, but it has a cost advantage.

According to SemiAnalysis's model, each TPU chip has about a 44% total cost of ownership (TCO) advantage compared to NVIDIA GB200 systems, and even external customers like Anthropic can enjoy a 30% to 50% lower cost. Google's system can connect 9,216 chips together to form a dense network, which is more convenient for distributed AI training compared to traditional NVIDIA systems.

To promote the adoption of TPUs, Google is developing native support for the popular PyTorch framework and integrating with inference libraries like vLLM, aiming to eliminate barriers for developers when migrating to TPUs.

However, facing Google's challenge, NVIDIA is preparing a technological counterattack, with its next-generation "Vera Rubin" chip expected to launch in 2026 or 2027. At the same time, Google's TPUv8 plan is facing delays, but it still aims to maintain its market competitiveness by launching a new version in collaboration with Broadcom and MediaTek.

Key Points:

💡 Google has transformed from an internal chip user to a retailer, introducing TPUv7 to challenge NVIDIA's market dominance.  

📉 OpenAI successfully secured a 30% cost discount on its hardware by switching to TPUs.  

⚙️ Google is actively promoting the development and adoption of TPUs, seeking to overcome software compatibility obstacles.