AI is reshaping chips and data centre architecture

Technology is increasingly defined by the physical infrastructure that enables artificial intelligence: this is the key insight put forward by McKinsey’s Technology Trends Outlook 2026, which reviews 14 trends against a backdrop that reflects the pervasiveness of AI (from robotics to scientific research and agent-based software development) and the resulting increased pressure on chips, networks, energy and data centres.

 The chapter of the report devoted to application-specific semiconductors describes the shift from large GPU clusters, focused on training state-of-the-art models, to heterogeneous architectures. Piero Montanari, a Partner at McKinsey & Company, sums up this shift with a sort of formula. “In the market for artificial intelligence chips,” the manager explains to *Il Sole 24 Ore*, “customers are also becoming suppliers: major cloud providers are now working alongside semiconductor manufacturers to design processors tailored to their AI models, which in some cases form part of the offering provided to other companies as part of their services.”

The boundaries between chip manufacturers, cloud service providers and model developers are therefore becoming increasingly blurred, giving rise to partnerships based on a greater mutual understanding of each other’s needs and areas of expertise. There is no shortage of examples of this. Amazon, Google, Meta and Microsoft are collaborating with companies such as Nvidia, Broadcom, Marvell Technologies, TSMC and others to develop components tailored to their own models; Anthropic plans to use up to one million TPUs (Tensor Processing Units) from Google Cloud; and AWS is complementing its services with its proprietary Trainium chips for the training and inference of machine learning models.

This transformation is driven above all by the growing importance of inference (that is, the work that pre-trained models carry out every time they generate a response), the use of which in businesses is becoming more widespread as costs fall and the infrastructure becomes more accessible. According to McKinsey’s estimates, by 2030 this ‘phase’ of the processing cycle could account for 30 to 40 per cent of total data centre demand and the majority of computing capacity dedicated to AI.

It is no coincidence that Montanari urges us to measure how much work is achieved per watt of power used, at what cost and with what response time: ‘This is a crucial aspect because data centre energy consumption is already putting pressure on the electricity grids, making efficiency a prerequisite for scaling up.’ As for system design, however, the manager emphasises the criteria that guide the decision-making process.