As the demand for artificial intelligence (AI) computing power surges, Google's homegrown tensor processing units (TPUs) are emerging as a compelling alternative to Nvidia's market-leading graphics processing units (GPUs). Google's TPUs, which serve as the engine for its Gemini chatbot and cloud-computing business, offer strong performance on AI tasks while reducing the cost of running those systems. The chips are optimized for machine learning tasks like training models and running them in real-time, known as inference. By co-designing the chips with Broadcom, Google has achieved an edge in efficiency, with TPUs consuming 20% to 40% less energy than Nvidia processors. This allows Google to charge about 20% to 30% less for excess compute capacity, attracting AI startups and enterprises to its cloud business and enterprise services. The latest eighth-generation TPUs mark a notable milestone, with two specialized variants for training and inference. These chips are up to three times faster for AI model training, offer 80% better performance per dollar, and can run more than 1 million TPUs in a single cluster. With a growing list of customers, including Anthropic and Meta Platforms, Google's TPU sales revenue is expected to generate about $3 billion in 2026 and jump to $25 billion in 2027. As the AI computing market continues to evolve, Google's custom TPUs are positioning the company as a major force in AI infrastructure, with a lucrative one-two punch of internal AI efforts and external customer wins.

Source: CNBC