For many years, the US company Nvidia shaped the foundations of modern artificial intelligence (AI). Its graphics processing units (GPUs) are a specialised type of computer chip originally designed to handle the processing demands of graphics and animation. But they’re also great for the repetitive calculations required by AI systems.
Thus, these chips have powered the rapid rise of large language models – the technology behind AI chatbots – and they have become the familiar engine behind almost every major AI breakthrough.
This hardware sat quietly in the background while most of the attention was focused on algorithms and data. Google’s decision to train Gemini on its own chips, called tensor processing units (TPUs), changes that picture. It invites the industry to look directly at the machines behind the models and to reconsider assumptions that long seemed fixed.
This moment matters because the scale of AI models has begun to expose the limits of general-purpose chips. As models grow, the demands placed on processing systems increase to levels that make hidden inefficiencies impossible to ignore.
Google’s reliance on TPUs reveals an industry that is starting to understand that hardware choices are not simply technical preferences but strategic commitments that determine who can lead the next wave of AI development.
Google’s Gemini relies on cloud systems that simplify the challenging task of coordinating devices during large-scale training (improvement) of AI models.
The design of these different chips reflects a fundamental difference in intention. Nvidia’s GPUs are general-purpose and flexible enough to run a wide range of tasks. TPUs were created for the narrow mathematical operations at the heart of AI models.
Independent comparisons highlight that TPU v5p pods can outperform high-end Nvidia systems on workloads tuned for Google’s software ecosystem. When the chip architecture, model structure and software stack align so closely, improvements in speed and efficiency become natural rather than forced.
These performance characteristics also reshape how quickly teams can experiment. When hardware works in concert with the models it is designed to train, iteration becomes faster and more scalable. This matters because the ability to test ideas quickly often determines which organisations innovate first.
These technical gains are only one part of the story. Training cutting-edge AI systems is expensive and requires enormous computing resources. Organisations that rely only on GPUs face high costs and increasing competition for supply. By developing and depending on its own hardware, Google gains more control over pricing, availability and long-term strategy.
Analysts have noted that this internal approach positions Google with lower operational costs while reducing dependence on external suppliers for chips. A particularly notable development came from Meta as it explored a multi-billion-dollar agreement to use TPU capacity.
When one of the largest consumers of GPUs evaluates a shift toward custom accelerators, it signals more than curiosity. It suggests growing recognition that relying on a single supplier may no longer be the safest or most efficient strategy in an industry where hardware availability shapes competitiveness.
These moves also raise questions about how cloud providers will position themselves. If TPUs become more widely available through Google’s cloud services, the rest of the market may gain access to hardware that was once considered proprietary. The ripple effects could reshape the economics of AI training far beyond Google’s internal research.
What This Means for Nvidia
Financial markets reacted quickly to the news. Nvidia’s stock fell as investors weighed the potential for cloud providers to split their hardware needs across more than one supplier. Even if TPUs do not replace GPUs entirely, their presence introduces competition that may influence pricing and development timelines.
The existence of credible alternatives pressures Nvidia to move faster, refine its offerings and appeal to customers who now see more than one viable path forward. Even so, Nvidia retains a strong position. Many organisations depend heavily on CUDA (a computing platform and programming model developed by Nvidia) and the large ecosystem of tools and workflows built around it.
Moving away from that environment requires significant engineering effort and may not be feasible for many teams. GPUs continue to offer unmatched flexibility for diverse workloads and will remain essential in many contexts.
However, the conversation around hardware has begun to shift. Companies building cutting-edge AI models are increasingly interested in specialised chips tuned to their exact needs. As models grow larger and more complex, organisations want greater control over the systems that support them. The idea that one chip family can meet every requirement is becoming harder to justify.
Google’s commitment to TPUs for Gemini illustrates this shift clearly. It shows that custom chips can train world-class AI models and that hardware purpose-built for AI is becoming central to future progress.
It also makes visible the growing diversification of AI infrastructure. Nvidia remains dominant, but it now shares the field with alternatives that are increasingly capable of shaping the direction of AI development.
The foundations of AI are becoming more varied and more competitive. Performance gains will come not only from new model architectures but from the hardware designed to support them.
Google’s TPU strategy marks the beginning of a new phase in which the path forward will be defined by a wider range of chips and by the organisations willing to rethink the assumptions that once held the industry together.
- is a Senior Lecturer in Artificial Intelligence and Machine Learning, University of Portsmouth
- This article first appeared in The Conversation
