Google TPU's Rise vs NVIDIA's AI Chip Dominance: A Balanced Analysis

According to Forbes, NVIDIA held over 75% of the AI chip market share in 2025, with a market capitalization exceeding $5 trillion [3]. Google’s latest TPU 7 (Ironwood) has an inference performance of 4614 TFLOPS, which is 2x more energy-efficient than GPUs, but NVIDIA’s CUDA ecosystem has formed a strong lock-in effect [1]. Anthropic signed an order for 1 million TPUs, marking accelerated external expansion of TPUs [2]. NVIDIA’s Q3 2025 revenue was $57 billion, up 62.5% year-over-year, with its data center business contributing $51.2 billion [4][5].

Xueqiu user “Gudong Yu” pointed out that Google TPU’s closed technical stack (does not support CUDA) limits its popularity, while NVIDIA provides the computing backbone for most enterprises worldwide [6]. Reddit discussions focused on supply chain opportunities in optical modules (e.g., Lumentum) and NAND flash (e.g., SanDisk) driven by the explosion in AI computing demand.

Google TPU has obvious advantages in inference efficiency and cost control, but it is difficult to overthrow NVIDIA’s dominant position in the short term—CUDA ecosystem and versatility are key barriers. The two are not in a zero-sum competition: TPU focuses on large-scale inference, while GPU dominates general-purpose training and innovation. Investment opportunities are concentrated in computing power supply chains (optical modules, PCB), NAND flash, and AI infrastructure-related enterprises.