Google has unveiled a new generation of artificial intelligence chips, marking one of its boldest moves yet to challenge Nvidia’s dominance in the fast-growing AI hardware market.
The tech giant announced that its latest tensor processing units, known as TPUs, will now be split into two specialised processors for the first time. One chip is dedicated to training AI models, while the other is built specifically for inference, the process of running those models in real-world applications such as chatbots and digital assistants.
This shift reflects a deeper change in how artificial intelligence is being developed and deployed. In the past, companies relied on a single type of chip to handle both training and inference. But as AI systems become more complex and widely used, the need for specialised hardware has grown significantly.
Google’s training chip, referred to as TPU 8t, is designed to handle massive, compute-intensive workloads required to build advanced AI models. Meanwhile, the TPU 8i is optimised for speed and efficiency when delivering AI responses to users, a critical function as demand for real-time AI applications continues to rise.
The company says this separation will improve performance, reduce costs, and increase energy efficiency. According to reports, the inference chip offers up to 80 percent better performance per dollar and significantly improved power efficiency compared to previous versions.
The announcement comes as competition in the AI infrastructure space intensifies. Nvidia currently dominates the market for AI chips, with its graphics processing units widely used by major tech companies and cloud providers. However, Google, along with rivals like Amazon and Microsoft, has been investing heavily in custom-built silicon to reduce reliance on Nvidia and gain more control over its AI ecosystem.
Despite this push, Google is not completely cutting ties with Nvidia. The company continues to offer Nvidia chips through its cloud services and plans to support upcoming Nvidia hardware, highlighting a strategy of both competition and collaboration.
The timing of the new chip launch is also significant. Industry experts say the focus is shifting from training large AI models to deploying them at scale, where inference becomes the dominant workload. This is especially true as AI agents and applications become more integrated into everyday digital services.
Google Cloud CEO Thomas Kurian emphasised that inference computing could soon rival or even surpass training in importance, as businesses increasingly demand AI systems that can operate in real time and handle complex tasks efficiently.
The chips are expected to be available to Google Cloud customers later in 2026, positioning the company to attract more enterprise clients looking for cost-effective and high-performance AI infrastructure.
Beyond performance gains, the move also highlights growing concerns about the energy demands of artificial intelligence. Training large models and running them at scale requires enormous computing power, and companies are under pressure to develop more energy-efficient solutions.
By designing chips tailored to specific AI tasks, Google aims to address these challenges while strengthening its position in the global AI race.
Ultimately, this latest development signals a broader industry trend. Major tech companies are no longer just building AI software, they are increasingly designing the hardware that powers it. As the competition intensifies, the battle for control over AI infrastructure is becoming just as important as the race to build the most advanced models.
