Google has officially unveiled its new generation of custom AI chips at its Cloud Next ’26 conference.
This announcement is a direct challenge to Nvidia in the new battleground of AI: inference. For a long time, the focus was on training massive AI models, which is like teaching a student. But now, the industry is shifting to inference, which is like having that student use their knowledge in the real world. Nvidia recently declared 2026 the "inference inflection" point and even licensed technology from a startup called Groq to build specialized, low-latency chips. This set the stage for Google's response.
So, how is Google tackling this? Their strategy has three key pillars. First, they are formalizing a two-track approach, offering specialized chips for both training and inference to external customers. This is a crucial shift from primarily using their TPUs (Tensor Processing Units) for their own services. They must now prove their chips offer better cost-per-token performance in the open market.
Second, Google is putting serious money behind this push. The company announced a staggering capital expenditure plan of $175–$185 billion for 2026, nearly double the previous year's. This massive investment is going directly into building out the data centers and infrastructure needed to deploy these new TPUs at a global scale and make them a core part of their Google Cloud revenue.
Third, they have secured a strong supply chain and anchor customers. Google signed a long-term deal with Broadcom to co-develop and supply these next-gen TPUs through 2031. At the same time, AI company Anthropic committed to using a massive amount of this TPU capacity, validating that there is significant external demand for Google's silicon beyond its own needs.
Google isn't alone in this fight. Other hyperscalers are also building their own chips. Microsoft recently announced its Maia 200 accelerator, specifically designed for inference. This industry-wide trend shows that major tech companies are determined to reduce their dependence on Nvidia and control their own AI hardware destiny.
Ultimately, Google's success will depend on execution. They need to demonstrate clear performance and cost advantages and, just as importantly, make their hardware easy for developers to use with popular frameworks like PyTorch. The race for the future of AI inference is officially on.
- AI Inference: The process of using a trained AI model to make predictions or generate outputs in real-time. It's the 'live' phase after the initial 'training' phase.
- TPU (Tensor Processing Unit): A custom-designed chip by Google, optimized specifically for the mathematical calculations required for artificial intelligence tasks.
- Hyperscaler: A large-scale cloud computing provider, such as Google Cloud, Amazon Web Services (AWS), or Microsoft Azure, that operates massive data centers.