The centre of gravity in artificial intelligence infrastructure has moved decisively from building models to running them. At an event in San Francisco, AMD presented data showing that inference—the moment a user interacts with a trained model—now accounts for roughly 60 percent of AI data centre demand, up from 40 percent in 2024, while training has fallen to 40 percent. The company’s new Helios platform, which integrates processors, accelerators and networking, is designed for this reality, and AMD claims it can deliver up to 30 percent more tokens per dollar than Nvidia’s forthcoming Vera Rubin NVL72 in certain configurations. Firms including OpenAI, Meta, Anthropic, Microsoft and Oracle are adopting or evaluating the architecture.

The shift is being accelerated by the rise of agentic AI—systems that consult external tools, access databases and coordinate multiple applications before delivering a result. Each such query multiplies the processing load, making cost efficiency the decisive competitive metric. AMD’s roadmap extends to 2028 with successive GPU generations, while the broader market is expanding rapidly: in Italy, the AI market grew 58 percent in 2024 to €1.2 billion, though Istat data show that high costs, unclear regulation and a lack of skills remain the top barriers, especially for small and medium-sized enterprises. In Indonesia, Google projects that full adoption of generative AI in the creative economy could unlock Rp66 trillion annually, with workers saving an average of 390 hours per year.

Organisations are struggling to align leadership with the pace of adoption. A Microsoft survey of 20,000 users across 10 countries found that only 26 percent believe their leadership is clearly aligned on AI use, while 65 percent fear being left behind. McKinsey data indicate that 71 percent of organisations now use generative AI regularly, yet just 21 percent have significantly redesigned workflows—the factor most strongly correlated with financial returns. The Adecco Group’s chief executive, citing OECD employment figures at record highs, argues that AI is reshaping tasks rather than destroying jobs, though entry-level roles are changing and upskilling is urgent. The World Economic Forum projects a net gain of 78 million jobs globally by 2030, but estimates that 40 percent of core skills will need to change.

Adoption patterns reveal sharp divides. In Brazil, AI is present in 59 percent of federal and state public bodies, reaching 94 percent in the judiciary but only 55 percent in the executive, while among municipalities the rate drops to 22 percent in towns with fewer than 10,000 inhabitants. Italian SMEs use AI at a rate of 15.7 percent, against 53.1 percent for large firms—a gap that has widened by 17 percentage points in two years. In Colombia, an Accenture study found that 74 percent of consumers would trust an AI agent more than a friend to make a purchase on their behalf, forcing brands to optimise for both emotional connection and algorithmic evaluation. Indonesia’s deputy minister for creative economy points to new roles such as prompt engineer, while YouTube’s creative ecosystem contributed over Rp8.4 trillion to Indonesian GDP in 2025.

The next milestone is whether the infrastructure build-out and the promised efficiency gains can be distributed beyond the largest players. AMD’s next GPU generation is scheduled for 2028, and governments from Rome to Jakarta are designing incentive schemes and training programmes. The immediate watchpoint is the capacity of agentic AI to scale without widening the digital divide between large firms and SMEs, and between high- and low-income municipalities.