The explosion of computing power ensures that the AI era is just beginning

While many analysts insist on the theory that artificial intelligence development will soon encounter an insurmountable limit, Mustafa Suleyman, CEO of Microsoft AI, argues that we are experiencing an unprecedented exponential progression. According to the executive, human intuition, shaped by evolution to understand the world in a linear fashion, fails when attempting to predict the current trajectory of computing. Far from reaching a ceiling, the processing capacity dedicated to frontier models is on a dizzying acceleration curve that promises to completely redefine the global economy and the execution of cognitive tasks.

The end of linear intuition

The common perception that technological progress must follow predictable rhythms, such as Moore's Law, has proven obsolete in the face of the facts. Since 2010, the volume of training data used in cutting-edge systems has jumped from approximately 10¹⁴ flops to over 10²⁶ flops—a trillion-fold increase. For Suleyman, skeptics who point to energy shortages or the slowdown in semiconductor manufacturing ignore the convergence of innovations that are keeping the gears of AI in constant and growing motion.

The engineering behind the performance leap

The secret to this advancement lies not just in larger machines, but in more efficient hardware orchestration. Historically, scaling computing was like adding more people to a room with calculators, where idle time was a constant waste. Today, the industry focuses on eliminating these bottlenecks. Three pillars support this evolution: the massive increase in raw chip performance—as exemplified by Nvidia processors and Microsoft's Maia 200 chip—the implementation of HBM (High Bandwidth Memory) technology, which accelerates data flow to processors, and the use of infrastructures like NVLink and InfiniBand, which connect hundreds of thousands of GPUs to function as a single digital brain.

Efficiency and economies of scale

Software optimization has been as important a growth vector as hardware. Data from Epoch AI indicate that the computational cost to reach a fixed level of performance is halved every eight months. As a result, the implementation cost of recent models has plummeted, in some cases by a factor of 900 on an annualized basis. What once took hours to train on a few processors is now processed in minutes in gigantic clusters, marking a transition from the AlexNet era, with two GPUs, to the current era, which mobilizes more than 100,000 GPUs in a single cluster.

The future of autonomous agents

The ultimate goal of this race is not merely to improve voice assistants or conversational chatbots. Suleyman's vision points to the creation of near-human agents, capable of carrying out complex long-term projects, managing logistics, negotiating contracts, and writing code autonomously. We are moving from the simple consultation phase into an era of collaboration between AI teams that deliberate and execute tasks. By 2028, an increase of 1,000 times in effective computing capacity is estimated, a scenario that may require annual energy inputs equivalent to the consumption of countries like the United Kingdom, France, Germany, and Italy combined.

Sustainability and cognitive abundance

Although energy consumption is a real challenge—with AI racks requiring power levels comparable to hundreds of homes—the sector is betting on convergence with another exponential trend: the drastic drop in solar energy costs and battery storage technologies. With investments in the neighborhood of $100 billion in industrial-scale supercomputers, the path to "cognitive abundance" is being paved. For the leaders at Microsoft AI, the skeptics who continue to predict diminishing returns are merely observing the margins of a historical phenomenon that, ironically, is only taking its first steps.