Microsoft AI chief Mustafa Suleyman argues that the exponential growth in compute power will keep AI models expanding far beyond current limits. Since 2010, training data has swelled 1 trillion times, from 10¹⁴ to over 10²⁶ floating‑point operations. This shift follows a cascade of hardware and software advances that let GPUs work nonstop and in concert.

Hardware gains stack up: Nvidia’s flagship chips leapt from 312 teraflops in 2020 to 2,250 teraflops today, while the new Maia 200 delivers 30 % more performance per dollar. High‑bandwidth memory, HBM3, triples data throughput, and interconnects like NVLink turn isolated GPUs into warehouse‑scale supercomputers that train a model in under four minutes.

Software breakthroughs cut compute costs dramatically. Research shows the flops needed to hit a fixed performance halve every eight months, faster than Moore’s 18‑24 month cycle. Serving costs for recent models have plummeted up to 900× annually, making large‑scale deployment economically viable and accelerating the shift from chatbots to full‑blown autonomous agents. Forecasts project 100 million H100‑equivalents by 2027.

Energy remains the bottleneck; a single AI rack consumes 120 kW, matching 100 homes. Yet solar and battery prices have dropped 100× and 97% respectively, opening a clean‑scaling path. Microsoft AI plans 100 billion‑dollar clusters drawing 10 gigawatts, positioning the company to deliver the next generation of cognitive abundance, while anticipating 200 gigawatts of additional compute online each year.